<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0"
     xmlns:content="http://purl.org/rss/1.0/modules/content/"
     xmlns:dc="http://purl.org/dc/elements/1.1/"
     xmlns:dcterms="http://purl.org/dc/terms/"
     xmlns:media="http://search.yahoo.com/mrss/"
     xmlns:atom="http://www.w3.org/2005/Atom"
     xmlns:cf="https://www.futureplc.com/rss/content-flags"
>
    <channel>
                    <atom:link href="https://www.tomshardware.com/feeds/articletype/news-analysis" rel="self" type="application/rss+xml" />
                            <title><![CDATA[ Latest from Tom's Hardware in News-analysis ]]></title>
                <link>https://www.tomshardware.com/news-analysis</link>
        <description><![CDATA[ All the latest news-analysis content from the Tom's Hardware team ]]></description>
                                    <lastBuildDate>Fri, 20 Nov 2026 15:00:00 +0000</lastBuildDate>
                            <language>en</language>
                                <item>
                                                            <title><![CDATA[ VAST says petabyte SSDs will outgrow today’s parallel file system RAID ]]></title>
                                                                                                <dc:content><![CDATA[ <p>VAST Data says petabyte SSDs will outgrow RAID in today’s parallel file systems, according to a simulation of its own 900 PB system that puts the odds of data loss at 1 in 11 over 10 years. At Supercomputing 2026 (<a href="https://www.tomshardware.com/tag/premium">SC26</a>), the company explains how.</p><p>Bigger drives take longer to rebuild after a failure, leaving more time for further failures to lose data, but operators will still want those large drives because many sites are power-constrained and storage power scales with bandwidth rather than capacity. VAST’s recommended fix: system-wide erasure coding with at least three parity strips.</p><p>The company’s principal technical strategist, Glenn Lockwood, walked through the simulation, described as a “discrete event simulation” using real failure data, in an Exhibitor Forum talk during SC26’s HPC System Hardware and Architecture session. With petabyte-scale SSDs swapped in, enclosure-local declustered RAID — the approach used by DDN EXAScaler and HPE Cray ClusterStor, per Lockwood — has those exceptionally high odds of data loss. With global declustering and extra parity, operators can get this down to 1 in 54 million, a significant improvement, but VAST’s own scheme gets to about 1 in 62 billion, three orders of magnitude better.</p><p>Lockwood’s case for petabyte SSDs rests on how storage power scales: at about 28 W per SSD, regardless of capacity and including its share of the surrounding infrastructure, plus 20 W per GB/s of bandwidth. Getting 900 PB of flash from 15 TB SSDs would take more than 63K drives, each adding its own bandwidth, for 190 TB/s at 5.6 MW, or “the equivalent of 3,100 or so of Blackwell Ultra GPUs.” Lockwood says, though, that “real AI applications in our experience have never approached needing anything close to that level of performance” from the storage.</p><p>With <a href="https://www.tomshardware.com/pc-components/ssds/solidigm-reveals-122tb-ssd-the-worlds-highest-capacity-drive-for-ai-workloads-d5-p5336-offers-unlimited-write-durability">122 TB drives</a>, at almost an 8K count and much lower 8 TB/s of aggregate bandwidth, the power draw is less than a tenth of that, at about 400 kW. Within that budget, 15 TB SSDs top out at about 250 PB, even with no power left for performance. That power could be better spent on compute and, anyway, 245 TB-class drives are already arriving, with much larger drives expected.</p><p>Everything is a series of trade-offs, though, and larger drives come with their own caveats. Lockwood acknowledges the usual reaction to a 245 TB drive: that it’s “a big blast radius,” since a single failure leaves much more to rebuild. And it does take longer: at the same per-drive speed, a 122 TB drive takes twice as long to rebuild as a 60 TB drive, leaving a longer window in which a second failure in a traditional RAID 5 set loses data. Per-drive capacity is potentially outpacing bandwidth increases and, at least within a generation, “the performance of each individual drive does not change,” so this is something that must be accepted or overcome.</p><p>The simulated system involves 7,920 SSDs at 122 TB each, using 180 enclosures with 44 drives each. With a RAID 5 single parity solution, there is a 100% chance of data loss over 10 years, but “no one uses single parity anymore.” Double parity still has 1 in 150 odds of data loss over a decade and is considered “state of the art for many open source file systems,” and he explicitly names Lustre and BeeGFS within this context.</p><p>For proprietary parallel file system implementations, declustered RAID within an enclosure is the state of the art, instead. Declustered RAID spreads each drive group’s data, parity, and spare space over more drives than usual, allowing for faster rebuilds. Keeping it within an enclosure means that the spread still stays within a single box of drives. Swapping petabyte SSDs into that enclosure-level declustering is what gets to the headline 1 in 11 odds of data loss.</p><p>As for the mentioned rivals, DDN’s DCR, in one published configuration, uses RAID 6 in an 8+2 array for a 24-NVMe SSD pool. That means eight data strips plus two parity strips per stripe, allowing any two drives in the pool to fail without losing data, and each stripe is spread over 10 of the 24 drives. DDN has conceded in the past that larger systems may need more than DCR. HPE’s declustered RAID solution is known as GridRAID, with declustered parity per enclosure half, meaning two 12-drive pools per 24-drive enclosure.</p><p>Perlmutter, the supercomputer at Berkeley Lab’s National Energy Research Scientific Computing Center, where Lockwood once worked, used GridRAID with 8+2+1 RAID 6 per 12-drive object storage target, Lustre’s name for a storage unit. The +1 is one spare drive’s worth of space, spread across all 12, so a rebuild writes to every surviving drive rather than to a single replacement. Either way, protection stops at two failures in a pool of one or two dozen drives, and only those drives can assist with a rebuild. This is the scenario in VAST’s simulation with the catastrophic failure rate.</p><div ><table><caption>VAST’s simulated odds of data loss on its 900PB system over 10 years</caption><thead><tr><th class="firstcol " ><p>Protection scheme</p></th><th  ><p>Who uses it, per Lockwood</p></th><th  ><p>Drive size simulated</p></th><th  ><p>Odds of data loss over 10 years</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Single parity (RAID 5)</p></td><td  ><p>"No one anymore"</p></td><td  ><p>122TB (today’s drives)</p></td><td  ><p>100%</p></td></tr><tr><td class="firstcol " ><p>Double parity</p></td><td  ><p>Open-source Lustre and BeeGFS, typically</p></td><td  ><p>122TB (today’s drives)</p></td><td  ><p>1 in 150</p></td></tr><tr><td class="firstcol " ><p>Declustered RAID within an enclosure</p></td><td  ><p>DDN EXAScaler, HPE Cray ClusterStor</p></td><td  ><p>Petabyte-class</p></td><td  ><p>1 in 11</p></td></tr><tr><td class="firstcol " ><p>Declustered RAID across the whole system, plus extra parity</p></td><td  ><p>His stated minimum</p></td><td  ><p>Petabyte-class</p></td><td  ><p>1 in 54 million</p></td></tr><tr><td class="firstcol " ><p>VAST’s own scheme (not detailed)</p></td><td  ><p>VAST</p></td><td  ><p>Petabyte-class</p></td><td  ><p>1 in 62 billion</p></td></tr></tbody></table></div><p>Several solutions exist, with limits. Adding a parity strip for triple parity reduces the odds of data loss. Global declustering, on its own, can also reduce the odds a little bit, with faster rebuild times. However, stripes are spread over every drive in the cluster, so two random failures are more likely to hit the same stripe, meaning a larger fault domain.</p><p>If both approaches are used, which is the minimum bar for him, you get down to getting struck by “lightning on your birthday” odds. VAST already publishes quadruple parity, which is beyond this bar. Using VAST’s own “special magic,” undisclosed but beyond the minimum bar, drops the odds even further, specifically for the petabyte SSDs, which is the real target of the talk.</p><p>Then we get to the metadata, the data about data, which in this case is the index of where every piece of every file dwells. Every write necessarily updates it synchronously, so VAST keeps it on separate, fast SSDs as three full copies for replication, not parity. This means it can survive two failures, which is less than the data’s four parity strips, and the index becomes the weakest link. With 122 TB drives, he puts the data drives’ odds at 1 in 78 trillion over 10 years, against 1 in 58 million for metadata. To Lockwood, this illustrates that the surrounding infrastructure is the real risk, given the failure gap between the index and regular data.</p><p>Rebuilding in the index area starts about 12 seconds after a failure, on average, versus about two minutes for data drives, Lockwood says. VAST’s procedure more generally, for when a drive drops out, is to power-cycle it and hope, so there is always some delay before a rebuild is forced. This slight delay barely matters for data drives with four parity strips but is more impactful for the three-copy metadata index drives. For example, waiting an hour for data drives “wouldn’t make a big difference” while just 20 minutes or more on a metadata drive has “extremely high” risk.</p><p>Earlier this year, in May, Micron began shipping the <a href="https://www.tomshardware.com/pc-components/ssds/microns-industry-first-pci-6-0-ssd-promises-sequential-reads-up-to-28-000-mb-s-245-tb-ssd-also-coming-for-those-who-need-capacity-more-than-cutting-edge-speed">245 TB 6600 ION</a>, which pulls up to 30 W at maximum load. Kioxia has <a href="https://www.tomshardware.com/pc-components/ssds/kioxia-unveils-245tb-ssd-the-worlds-highest-capacity-storage-device-could-store-12-500-4k-movies">the LC9</a>, used in a Dell 2U server with 40 drives for a total of 9.8 PB. SanDisk also has its 256 TB UltraQLC, which was due in the first half of this year, and Solidigm is <a href="https://www.tomshardware.com/pc-components/ssds/solidigm-vp-talks-pcie-6-0-ssds-next-gen-floating-gate-nand-liquid-cooled-storage-and-more-avi-shetty-vp-of-ai-solutions-and-market-enablement-discusses-the-future-of-enterprise-storage-tech">targeting about 245 TB later this year</a>.</p><p>Meanwhile, multiple roadmaps reach petabyte-class drives, including <a href="https://www.tomshardware.com/tech-industry/sandisk-puts-petabyte-ssds-on-the-roadmap-has-yet-to-reveal-release-date">SanDisk’s UltraQLC</a>, <a href="https://www.tomshardware.com/pc-components/ssds/sk-hynix-unveils-ai-nand-strategy-including-gargantuan-petabyte-class-qlc-ssds-ultra-fast-hbf-and-100m-iops-ssds-also-in-the-pipeline">a target by SK hynix</a> for QLC SSDs, and Samsung at some time in the future. This strengthens VAST’s case and, while at least the power math adds up, data still needs to be compared openly and fairly to make a firm conclusion about the best pathway forward for what is clearly an unavoidable problem.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/pc-components/ssds/vast-data-says-petabyte-ssds-will-outgrow-raid-in-todays-parallel-file-systems-simulation-of-a-900-pb-system-puts-data-loss-odds-at-1-in-11-over-10-years</link>
                                                                            <description>
                            <![CDATA[ VAST Data’s Glenn Lockwood told SC26 that petabyte SSDs need global declustering plus extra parity to avoid data loss. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">v6Cbkqm4xZLhCvpVWhXzBW</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/3jeEP7YJtYDcubMn3TrA3Z-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Fri, 20 Nov 2026 15:00:00 +0000</pubDate>                                                                                                                                <updated>Fri, 20 Nov 2026 16:47:07 +0000</updated>
                                                                                                                                            <category><![CDATA[SSDs]]></category>
                                                    <category><![CDATA[PC Components]]></category>
                                                    <category><![CDATA[Storage]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>false</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/3jeEP7YJtYDcubMn3TrA3Z-1920-80.jpg">
                                                            <media:credit><![CDATA[Micron]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Micron 6600 ION 245TB SSDs in E3.L and U.2 form factors]]></media:description>                                                            <media:text><![CDATA[Micron 6600 ION 245TB SSDs in E3.L and U.2 form factors]]></media:text>
                                <media:title type="plain"><![CDATA[Micron 6600 ION 245TB SSDs in E3.L and U.2 form factors]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/3jeEP7YJtYDcubMn3TrA3Z-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>VAST Data says petabyte SSDs will outgrow RAID in today’s parallel file systems, according to a simulation of its own 900 PB system that puts the odds of data loss at 1 in 11 over 10 years. At Supercomputing 2026 (<a href="https://www.tomshardware.com/tag/premium">SC26</a>), the company explains how.</p><p>Bigger drives take longer to rebuild after a failure, leaving more time for further failures to lose data, but operators will still want those large drives because many sites are power-constrained and storage power scales with bandwidth rather than capacity. VAST’s recommended fix: system-wide erasure coding with at least three parity strips.</p><p>The company’s principal technical strategist, Glenn Lockwood, walked through the simulation, described as a “discrete event simulation” using real failure data, in an Exhibitor Forum talk during SC26’s HPC System Hardware and Architecture session. With petabyte-scale SSDs swapped in, enclosure-local declustered RAID — the approach used by DDN EXAScaler and HPE Cray ClusterStor, per Lockwood — has those exceptionally high odds of data loss. With global declustering and extra parity, operators can get this down to 1 in 54 million, a significant improvement, but VAST’s own scheme gets to about 1 in 62 billion, three orders of magnitude better.</p><p>Lockwood’s case for petabyte SSDs rests on how storage power scales: at about 28 W per SSD, regardless of capacity and including its share of the surrounding infrastructure, plus 20 W per GB/s of bandwidth. Getting 900 PB of flash from 15 TB SSDs would take more than 63K drives, each adding its own bandwidth, for 190 TB/s at 5.6 MW, or “the equivalent of 3,100 or so of Blackwell Ultra GPUs.” Lockwood says, though, that “real AI applications in our experience have never approached needing anything close to that level of performance” from the storage.</p><p>With <a href="https://www.tomshardware.com/pc-components/ssds/solidigm-reveals-122tb-ssd-the-worlds-highest-capacity-drive-for-ai-workloads-d5-p5336-offers-unlimited-write-durability">122 TB drives</a>, at almost an 8K count and much lower 8 TB/s of aggregate bandwidth, the power draw is less than a tenth of that, at about 400 kW. Within that budget, 15 TB SSDs top out at about 250 PB, even with no power left for performance. That power could be better spent on compute and, anyway, 245 TB-class drives are already arriving, with much larger drives expected.</p><p>Everything is a series of trade-offs, though, and larger drives come with their own caveats. Lockwood acknowledges the usual reaction to a 245 TB drive: that it’s “a big blast radius,” since a single failure leaves much more to rebuild. And it does take longer: at the same per-drive speed, a 122 TB drive takes twice as long to rebuild as a 60 TB drive, leaving a longer window in which a second failure in a traditional RAID 5 set loses data. Per-drive capacity is potentially outpacing bandwidth increases and, at least within a generation, “the performance of each individual drive does not change,” so this is something that must be accepted or overcome.</p><p>The simulated system involves 7,920 SSDs at 122 TB each, using 180 enclosures with 44 drives each. With a RAID 5 single parity solution, there is a 100% chance of data loss over 10 years, but “no one uses single parity anymore.” Double parity still has 1 in 150 odds of data loss over a decade and is considered “state of the art for many open source file systems,” and he explicitly names Lustre and BeeGFS within this context.</p><p>For proprietary parallel file system implementations, declustered RAID within an enclosure is the state of the art, instead. Declustered RAID spreads each drive group’s data, parity, and spare space over more drives than usual, allowing for faster rebuilds. Keeping it within an enclosure means that the spread still stays within a single box of drives. Swapping petabyte SSDs into that enclosure-level declustering is what gets to the headline 1 in 11 odds of data loss.</p><p>As for the mentioned rivals, DDN’s DCR, in one published configuration, uses RAID 6 in an 8+2 array for a 24-NVMe SSD pool. That means eight data strips plus two parity strips per stripe, allowing any two drives in the pool to fail without losing data, and each stripe is spread over 10 of the 24 drives. DDN has conceded in the past that larger systems may need more than DCR. HPE’s declustered RAID solution is known as GridRAID, with declustered parity per enclosure half, meaning two 12-drive pools per 24-drive enclosure.</p><p>Perlmutter, the supercomputer at Berkeley Lab’s National Energy Research Scientific Computing Center, where Lockwood once worked, used GridRAID with 8+2+1 RAID 6 per 12-drive object storage target, Lustre’s name for a storage unit. The +1 is one spare drive’s worth of space, spread across all 12, so a rebuild writes to every surviving drive rather than to a single replacement. Either way, protection stops at two failures in a pool of one or two dozen drives, and only those drives can assist with a rebuild. This is the scenario in VAST’s simulation with the catastrophic failure rate.</p><div ><table><caption>VAST’s simulated odds of data loss on its 900PB system over 10 years</caption><thead><tr><th class="firstcol " ><p>Protection scheme</p></th><th  ><p>Who uses it, per Lockwood</p></th><th  ><p>Drive size simulated</p></th><th  ><p>Odds of data loss over 10 years</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Single parity (RAID 5)</p></td><td  ><p>"No one anymore"</p></td><td  ><p>122TB (today’s drives)</p></td><td  ><p>100%</p></td></tr><tr><td class="firstcol " ><p>Double parity</p></td><td  ><p>Open-source Lustre and BeeGFS, typically</p></td><td  ><p>122TB (today’s drives)</p></td><td  ><p>1 in 150</p></td></tr><tr><td class="firstcol " ><p>Declustered RAID within an enclosure</p></td><td  ><p>DDN EXAScaler, HPE Cray ClusterStor</p></td><td  ><p>Petabyte-class</p></td><td  ><p>1 in 11</p></td></tr><tr><td class="firstcol " ><p>Declustered RAID across the whole system, plus extra parity</p></td><td  ><p>His stated minimum</p></td><td  ><p>Petabyte-class</p></td><td  ><p>1 in 54 million</p></td></tr><tr><td class="firstcol " ><p>VAST’s own scheme (not detailed)</p></td><td  ><p>VAST</p></td><td  ><p>Petabyte-class</p></td><td  ><p>1 in 62 billion</p></td></tr></tbody></table></div><p>Several solutions exist, with limits. Adding a parity strip for triple parity reduces the odds of data loss. Global declustering, on its own, can also reduce the odds a little bit, with faster rebuild times. However, stripes are spread over every drive in the cluster, so two random failures are more likely to hit the same stripe, meaning a larger fault domain.</p><p>If both approaches are used, which is the minimum bar for him, you get down to getting struck by “lightning on your birthday” odds. VAST already publishes quadruple parity, which is beyond this bar. Using VAST’s own “special magic,” undisclosed but beyond the minimum bar, drops the odds even further, specifically for the petabyte SSDs, which is the real target of the talk.</p><p>Then we get to the metadata, the data about data, which in this case is the index of where every piece of every file dwells. Every write necessarily updates it synchronously, so VAST keeps it on separate, fast SSDs as three full copies for replication, not parity. This means it can survive two failures, which is less than the data’s four parity strips, and the index becomes the weakest link. With 122 TB drives, he puts the data drives’ odds at 1 in 78 trillion over 10 years, against 1 in 58 million for metadata. To Lockwood, this illustrates that the surrounding infrastructure is the real risk, given the failure gap between the index and regular data.</p><p>Rebuilding in the index area starts about 12 seconds after a failure, on average, versus about two minutes for data drives, Lockwood says. VAST’s procedure more generally, for when a drive drops out, is to power-cycle it and hope, so there is always some delay before a rebuild is forced. This slight delay barely matters for data drives with four parity strips but is more impactful for the three-copy metadata index drives. For example, waiting an hour for data drives “wouldn’t make a big difference” while just 20 minutes or more on a metadata drive has “extremely high” risk.</p><p>Earlier this year, in May, Micron began shipping the <a href="https://www.tomshardware.com/pc-components/ssds/microns-industry-first-pci-6-0-ssd-promises-sequential-reads-up-to-28-000-mb-s-245-tb-ssd-also-coming-for-those-who-need-capacity-more-than-cutting-edge-speed">245 TB 6600 ION</a>, which pulls up to 30 W at maximum load. Kioxia has <a href="https://www.tomshardware.com/pc-components/ssds/kioxia-unveils-245tb-ssd-the-worlds-highest-capacity-storage-device-could-store-12-500-4k-movies">the LC9</a>, used in a Dell 2U server with 40 drives for a total of 9.8 PB. SanDisk also has its 256 TB UltraQLC, which was due in the first half of this year, and Solidigm is <a href="https://www.tomshardware.com/pc-components/ssds/solidigm-vp-talks-pcie-6-0-ssds-next-gen-floating-gate-nand-liquid-cooled-storage-and-more-avi-shetty-vp-of-ai-solutions-and-market-enablement-discusses-the-future-of-enterprise-storage-tech">targeting about 245 TB later this year</a>.</p><p>Meanwhile, multiple roadmaps reach petabyte-class drives, including <a href="https://www.tomshardware.com/tech-industry/sandisk-puts-petabyte-ssds-on-the-roadmap-has-yet-to-reveal-release-date">SanDisk’s UltraQLC</a>, <a href="https://www.tomshardware.com/pc-components/ssds/sk-hynix-unveils-ai-nand-strategy-including-gargantuan-petabyte-class-qlc-ssds-ultra-fast-hbf-and-100m-iops-ssds-also-in-the-pipeline">a target by SK hynix</a> for QLC SSDs, and Samsung at some time in the future. This strengthens VAST’s case and, while at least the power math adds up, data still needs to be compared openly and fairly to make a firm conclusion about the best pathway forward for what is clearly an unavoidable problem.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Phison details SSD with a built-in NPU and 64GB of memory that supports AI frameworks on the drive ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Phison <a href="https://www.phison.com/sc26-phison-unveils-three-next-generation-storage-innovations-for-the-new-compute-era/">detailed two new enterprise SSD designs</a> at Supercomputer 2026 (<a href="https://www.tomshardware.com/tag/sc26">SC26</a>) that aim to answer two separate problems: one with a built-in NPU and 64GB of 130 GB/s memory that supports running PyTorch, vLLM, OpenVINO, and llama.cpp directly on the drive, and another that merges fast pSLC cache with large QLC capacity behind just one controller. The company notes that “many of the products and features mentioned are still in development” and will arrive once finalized.</p><p>The computational-storage design’s premise is that a server can offload more to the SSD, which runs AI code on the data where it already resides and sends back results, instead of whole datasets. The pSLC-QLC design, meanwhile, puts the cache inside each drive, which may reduce system complexity, as every drive bay contributes usable capacity and none goes to dedicated cache drives.</p><p>Phison says that the integrated NPU is expected to deliver “36,800 DMIPS of compute performance,” or Dhrystone Million Instructions Per Second, a typical synthetic benchmark. This is admittedly an old CPU integer benchmark from the 1980s, while NPUs tend to be rated in TOPS, but that figure was not given. For reference, the Topaz NPU add-in card Phison showed at Computex carried eight NPUs rated at up to 40 TOPS each. The DMIPS figure therefore likely refers to, instead, the CPU cores that work alongside it, which we would expect to be based on an Arm Cortex design, adequate for running an SSD of this class but no measure of its AI capability.</p><p>The memory and its bandwidth are more important for the expected workload, anyway, and while the company doesn’t state the type, LPDDR5X is a popular option and may be used with the Topaz NPU, according to one secondary report. If the drive is a PCIe 6.0 part comparable to Phison’s X3 controller, it could hit up to 28 GB/s to the host, which would put the memory above 4x that speed. For reference, 130 GB/s is a bit less than half of that of the <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-dgx-spark-review/2">DGX Spark’s</a> hardware, designed for AI development.</p><p>Phison says developers can build on the drive “without specialized programming tools,” which, in this case, is a good thing. The drive does, of course, support multiple general-purpose inference servers and runtimes, but nothing was disclosed about how a host reaches the drive’s compute. The drives “scan and process data in parallel,” which indicates scalability, and return “relevant results” which Phison claims could mean 10 to 100 times less data transferred for applicable workloads. This may free up the host CPU for other tasks.</p><p>Phison’s idea to put compute on storage isn’t new. Samsung tried this with its SmartSSD — where an FPGA sat on the drive to process data without moving it — which was shown in 2018. Later, in 2022, the design was <a href="https://www.tomshardware.com/news/samsung-introduces-amd-powered-second-gen-computational-ssd">refined with AMD’s Xilinx Versal FPGA</a> in its second generation, with Samsung claiming more than 50% lower processing time, up to 70% lower energy consumption, and up to 97% lower server CPU usage for scan-heavy database queries, compared to conventional SSDs. The drive has reportedly since disappeared from Samsung’s lineup, though the first-generation model was still sold under AMD’s Xilinx name as of last year.</p><p>ScaleFlux took a narrower route, with compression handled inside the SSD controller. The server doesn’t see this, so the drive looks like any other NVMe SSD, and no application has to change its behavior. The CEO claims it is the only company with computational storage in production at hyperscalers. This suggests that not making customers write new code for the drive is the secret to success, as SmartSSD needed that for its FPGA and for each workload. Phison’s selling point here is flexibility for the client to use code written for common AI frameworks.</p><p>Company CTO Sebastien Jean, writing in <em>All About Circuits</em> in 2025, argued that generic computational storage had stalled because “each use case is highly unique and it’s simply not scalable.” He called the idea of running the Linux operating system on a drive with bigger processors a “pie-in-the-sky approach.”</p><p>Accelerators have long been widely used in data centers, and Jean’s proposal was to put targeted ones inside the SSD, each running a set operation over a range of logical block addresses (LBAs). By his reckoning, a chassis with 90 PCIe 6.0 SSDs could process data at 2.5 TB/s without affecting the CPU. The design revealed at SC26, though, appears closer to the general-purpose model, possibly due to the increasing relevance of AI and the demand for both memory and compute.</p><p>Jean also spoke at the Emerging Storage Technology panel at the conference to discuss future developments, including PCIe 7.0 drives that would operate at “50, 56 gigabytes per second,” which he called “faster than a DDR channel.” With 32 drives at this speed, a single chassis could reach “1.8 terabytes per second.” He described this specifically as computational storage for HPC.</p><p>Turning to the Mixed Media solution, that’s the “high-endurance pSLC and high-capacity QLC flash” SSD; it’s a 16-channel part, so not comparable to anything you see in retail. The pSLC portion of the drive has the 4-bit QLC flash act in a 1-bit SLC mode, which trades space for performance and endurance. Phison says the design builds on standard NVMe interfaces and fits into existing infrastructure “without requiring specialized host software.”</p><p>Jean elaborated at the same panel that the “endurance group on the drive, which is SLC,” is managed separately from the XLC — 3-bit TLC or 4-bit QLC — and each region gets a namespace, so it appears as multiple drives to the OS. Then, high-write data can be sent specifically to the SLC group, while data in the QLC is likely to be more read-heavy. This follows a general trend of enterprise shifting toward QLC for capacity, but having some of it in SLC mode can help reduce the need for specialized caching.</p><p>Right now, high-capacity SSDs usually have separate SSDs to fulfill this role, but that takes up more drive bays. The tradeoff with Phison’s approach is that a single drive failure takes out both capacity and a share of the cache. That said, Phison’s approach is more robust from a cache perspective, as “losing one drive out of 24 would result in a 4% cache loss instead of a 25% loss,” the company says, with the 25% presumably assuming four dedicated cache drives.</p><p>Dedicated cache drives need tiering handled by the host or array software, while with Mixed Media each drive holds both tiers and the host simply picks the namespace it wants to write to.</p><p>The company is not the first to have this idea, either, as Kioxia presented its Mixed Mode idea during FMS 2025, where each drive has a ratio of pSLC to QLC, set once per drive when it’s first used. Kioxia called 1:4 a “good beginning ratio,” though it said customers typically want pSLC at just 0.5% to 2% of maximum QLC capacity. That implementation also had parity protection on the pSLC space across multiple drives. Kioxia similarly marketed the benefit of needing fewer slots and field-replaceable units (FRUs), and its slides even used the term “Mixed Media” in two chart headings.</p><p>DapuStor went further with its <a href="https://www.tomshardware.com/pc-components/ssds/dapustor-splits-qlc-ssd-to-create-a-fast-pslc-region-in-dual-mode-drive-trades-6-percent-to-20-percent-of-its-qlc-capacity-for-more-than-7x-faster-random-writes">dual-mode J5060</a>, a QLC SSD with a pSLC region that was in customer validation as of August; it compared the design against “four dedicated SLC drives” in a 24-drive system, the kind of setup that would produce Phison’s 25% figure. Solidigm’s Cloud Storage Acceleration Layer works differently, with a host-side flash translation layer, where it can manage small writes on SLC or TLC drives before writing to QLC in favorable large sequential chunks.</p><p>There’s also the Enmotus FuzeDrive P200, previewed at CES 2020 as the MiDrive, which used a Phison controller and seems to be the closest to an ancestor of the current Mixed Media. That consumer drive, <a href="https://www.tomshardware.com/reviews/enmotus-fuzedrive-p200-m2-nvme-ssd-review">which we reviewed in 2021</a>, had SLC and QLC portions that acted distinctly, but they were managed by Enmotus software.</p><div ><table><caption>pSLC and QLC tiering approaches</caption><thead><tr><th class="firstcol " ><p>Design</p></th><th  ><p>Where the cache is</p></th><th  ><p>pSLC-to-QLC ratio</p></th><th  ><p>How it is managed</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Phison Mixed Media</p></td><td  ><p>In every drive, behind one 16-channel controller</p></td><td  ><p>Not stated</p></td><td  ><p>Separate NVMe endurance groups assigned to namespaces; no specialized host software, Phison says</p></td></tr><tr><td class="firstcol " ><p>Kioxia Mixed Mode</p></td><td  ><p>In every drive, with parity protection on the pSLC space across drives</p></td><td  ><p>1:4 as a “good beginning ratio”; customers typically want pSLC at 0.5% to 2% of maximum QLC capacity</p></td><td  ><p>Ratio set once, creating a pSLC namespace and endurance group</p></td></tr><tr><td class="firstcol " ><p>DapuStor dual-mode</p></td><td  ><p>In every drive, as a pSLC region beside the QLC</p></td><td  ><p>400GB, 800GB or 1.2TB of pSLC on a 30.72TB drive</p></td><td  ><p>Separate SLC and QLC block devices on one drive</p></td></tr><tr><td class="firstcol " ><p>Enmotus FuzeDrive P200</p></td><td  ><p>In the drive, as a fixed SLC portion of a consumer QLC SSD</p></td><td  ><p>24GB of SLC on the 900GB model</p></td><td  ><p>Enmotus’ own software</p></td></tr><tr><td class="firstcol " ><p>Solidigm CSAL</p></td><td  ><p>On dedicated SLC or TLC cache drives</p></td><td  ><p>Reference system: three 800GB SLC drives and seven 61.44TB QLC drives</p></td><td  ><p>Flash translation layer running on the host</p></td></tr></tbody></table></div><p>It’s also inevitable that these two approaches will be compared to Phison’s other upcoming part, the X3 enterprise SSD controller, a PCIe 6.0 design promising up to 28 GB/s of sequential throughput and 6.8 million IOPS. Those figures are <a href="https://www.tomshardware.com/pc-components/ssds/phison-shows-pcie-6-0-x3-ssd-controller-with-28-gb-s-of-bandwidth-and-6-8-million-iops-supports-2-petabytes-per-drive-also-new-power-sipping-e37t-ssds-for-pcie-5-0-systems-consume-a-mere-4-5w">unchanged since Computex in June</a>.</p><p>The X3’s IOPS sounds like a lot, but we’ve heard considerable banter about AI needing way more than that, as <a href="https://www.tomshardware.com/tech-industry/nvidia-and-kioxia-target-100-million-iops-ssd-in-2027-33-times-more-than-existing-drives-for-exclusive-use-in-ai-servers">Nvidia’s roadmap target for PCIe 7.0</a> is at 100 million IOPS, though that target counts 512-byte reads, a smaller unit than the 4KB that SSD ratings usually use, so the two figures aren’t directly comparable. Jean said at the panel that Phison is “very optimistic” it can achieve that. He added that the eightfold jump in IOPS that comes with moving from 4KB to 512-byte accesses would, on existing NAND, mean “eight times as much power,” or “roughly in the 140 to 180 watt range,” which is beyond reasonable, so perhaps a different solution will be required.</p><p>For the X3, Phison told us in June to expect customer samples by the end of the year, with volume coming in the middle of next year, though the SC26 release gives no dates. Notably, Micron and Samsung are <a href="https://www.tomshardware.com/tech-industry/the-current-state-of-pcie-6-0-ssds-and-controllers-marvell-phison-and-smi-prepare-controllers-as-drives-finally-come-to-market-following-years-of-delays">already offering PCIe 6.0 SSDs</a>, and the Bravera SC6 controller is expected from Marvell. Phison also says its PS7161, which its site lists as a customer sample, is “the first PCIe 6.0 redriver on the PCI-SIG Integrators List.”</p><p>Of the two designs, Mixed Media looks like the nearer product. It builds on standard NVMe features, and DapuStor’s similar dual-mode drive reached customer testing this summer, while the one processor figure given for the computational-storage drive says little about what it can do for AI. What remains to be seen for Mixed Media is its pSLC-to-QLC ratio and how configurable it is. CES and Computex are the notable dates on the 2027 calendar to find out more about both, as well as the X3.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/pc-components/ssds/phison-details-ssd-with-a-built-in-npu-and-64gb-of-memory-that-supports-pytorch-and-llama-cpp-running-on-the-drive-second-design-merges-fast-pslc-cache-and-qlc-capacity-behind-one-controller</link>
                                                                            <description>
                            <![CDATA[ Phison showed a computational-storage SSD with an NPU and 64GB of memory, plus Mixed Media pSLC and QLC drives, at SC26. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">3Mu3PWBXJB5RzffBfHesin</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/BzuF4iUiRQ36JLAhHEkUKF-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Fri, 20 Nov 2026 14:00:00 +0000</pubDate>                                                                                                                                                                                                                                <category><![CDATA[SSDs]]></category>
                                                    <category><![CDATA[PC Components]]></category>
                                                    <category><![CDATA[Storage]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>false</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/BzuF4iUiRQ36JLAhHEkUKF-1920-80.jpg">
                                                            <media:credit><![CDATA[Tom&amp;#39;s Hardware]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[PCIe 6.0 controller]]></media:description>                                                            <media:text><![CDATA[PCIe 6.0 controller]]></media:text>
                                <media:title type="plain"><![CDATA[PCIe 6.0 controller]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/BzuF4iUiRQ36JLAhHEkUKF-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Phison <a href="https://www.phison.com/sc26-phison-unveils-three-next-generation-storage-innovations-for-the-new-compute-era/">detailed two new enterprise SSD designs</a> at Supercomputer 2026 (<a href="https://www.tomshardware.com/tag/sc26">SC26</a>) that aim to answer two separate problems: one with a built-in NPU and 64GB of 130 GB/s memory that supports running PyTorch, vLLM, OpenVINO, and llama.cpp directly on the drive, and another that merges fast pSLC cache with large QLC capacity behind just one controller. The company notes that “many of the products and features mentioned are still in development” and will arrive once finalized.</p><p>The computational-storage design’s premise is that a server can offload more to the SSD, which runs AI code on the data where it already resides and sends back results, instead of whole datasets. The pSLC-QLC design, meanwhile, puts the cache inside each drive, which may reduce system complexity, as every drive bay contributes usable capacity and none goes to dedicated cache drives.</p><p>Phison says that the integrated NPU is expected to deliver “36,800 DMIPS of compute performance,” or Dhrystone Million Instructions Per Second, a typical synthetic benchmark. This is admittedly an old CPU integer benchmark from the 1980s, while NPUs tend to be rated in TOPS, but that figure was not given. For reference, the Topaz NPU add-in card Phison showed at Computex carried eight NPUs rated at up to 40 TOPS each. The DMIPS figure therefore likely refers to, instead, the CPU cores that work alongside it, which we would expect to be based on an Arm Cortex design, adequate for running an SSD of this class but no measure of its AI capability.</p><p>The memory and its bandwidth are more important for the expected workload, anyway, and while the company doesn’t state the type, LPDDR5X is a popular option and may be used with the Topaz NPU, according to one secondary report. If the drive is a PCIe 6.0 part comparable to Phison’s X3 controller, it could hit up to 28 GB/s to the host, which would put the memory above 4x that speed. For reference, 130 GB/s is a bit less than half of that of the <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-dgx-spark-review/2">DGX Spark’s</a> hardware, designed for AI development.</p><p>Phison says developers can build on the drive “without specialized programming tools,” which, in this case, is a good thing. The drive does, of course, support multiple general-purpose inference servers and runtimes, but nothing was disclosed about how a host reaches the drive’s compute. The drives “scan and process data in parallel,” which indicates scalability, and return “relevant results” which Phison claims could mean 10 to 100 times less data transferred for applicable workloads. This may free up the host CPU for other tasks.</p><p>Phison’s idea to put compute on storage isn’t new. Samsung tried this with its SmartSSD — where an FPGA sat on the drive to process data without moving it — which was shown in 2018. Later, in 2022, the design was <a href="https://www.tomshardware.com/news/samsung-introduces-amd-powered-second-gen-computational-ssd">refined with AMD’s Xilinx Versal FPGA</a> in its second generation, with Samsung claiming more than 50% lower processing time, up to 70% lower energy consumption, and up to 97% lower server CPU usage for scan-heavy database queries, compared to conventional SSDs. The drive has reportedly since disappeared from Samsung’s lineup, though the first-generation model was still sold under AMD’s Xilinx name as of last year.</p><p>ScaleFlux took a narrower route, with compression handled inside the SSD controller. The server doesn’t see this, so the drive looks like any other NVMe SSD, and no application has to change its behavior. The CEO claims it is the only company with computational storage in production at hyperscalers. This suggests that not making customers write new code for the drive is the secret to success, as SmartSSD needed that for its FPGA and for each workload. Phison’s selling point here is flexibility for the client to use code written for common AI frameworks.</p><p>Company CTO Sebastien Jean, writing in <em>All About Circuits</em> in 2025, argued that generic computational storage had stalled because “each use case is highly unique and it’s simply not scalable.” He called the idea of running the Linux operating system on a drive with bigger processors a “pie-in-the-sky approach.”</p><p>Accelerators have long been widely used in data centers, and Jean’s proposal was to put targeted ones inside the SSD, each running a set operation over a range of logical block addresses (LBAs). By his reckoning, a chassis with 90 PCIe 6.0 SSDs could process data at 2.5 TB/s without affecting the CPU. The design revealed at SC26, though, appears closer to the general-purpose model, possibly due to the increasing relevance of AI and the demand for both memory and compute.</p><p>Jean also spoke at the Emerging Storage Technology panel at the conference to discuss future developments, including PCIe 7.0 drives that would operate at “50, 56 gigabytes per second,” which he called “faster than a DDR channel.” With 32 drives at this speed, a single chassis could reach “1.8 terabytes per second.” He described this specifically as computational storage for HPC.</p><p>Turning to the Mixed Media solution, that’s the “high-endurance pSLC and high-capacity QLC flash” SSD; it’s a 16-channel part, so not comparable to anything you see in retail. The pSLC portion of the drive has the 4-bit QLC flash act in a 1-bit SLC mode, which trades space for performance and endurance. Phison says the design builds on standard NVMe interfaces and fits into existing infrastructure “without requiring specialized host software.”</p><p>Jean elaborated at the same panel that the “endurance group on the drive, which is SLC,” is managed separately from the XLC — 3-bit TLC or 4-bit QLC — and each region gets a namespace, so it appears as multiple drives to the OS. Then, high-write data can be sent specifically to the SLC group, while data in the QLC is likely to be more read-heavy. This follows a general trend of enterprise shifting toward QLC for capacity, but having some of it in SLC mode can help reduce the need for specialized caching.</p><p>Right now, high-capacity SSDs usually have separate SSDs to fulfill this role, but that takes up more drive bays. The tradeoff with Phison’s approach is that a single drive failure takes out both capacity and a share of the cache. That said, Phison’s approach is more robust from a cache perspective, as “losing one drive out of 24 would result in a 4% cache loss instead of a 25% loss,” the company says, with the 25% presumably assuming four dedicated cache drives.</p><p>Dedicated cache drives need tiering handled by the host or array software, while with Mixed Media each drive holds both tiers and the host simply picks the namespace it wants to write to.</p><p>The company is not the first to have this idea, either, as Kioxia presented its Mixed Mode idea during FMS 2025, where each drive has a ratio of pSLC to QLC, set once per drive when it’s first used. Kioxia called 1:4 a “good beginning ratio,” though it said customers typically want pSLC at just 0.5% to 2% of maximum QLC capacity. That implementation also had parity protection on the pSLC space across multiple drives. Kioxia similarly marketed the benefit of needing fewer slots and field-replaceable units (FRUs), and its slides even used the term “Mixed Media” in two chart headings.</p><p>DapuStor went further with its <a href="https://www.tomshardware.com/pc-components/ssds/dapustor-splits-qlc-ssd-to-create-a-fast-pslc-region-in-dual-mode-drive-trades-6-percent-to-20-percent-of-its-qlc-capacity-for-more-than-7x-faster-random-writes">dual-mode J5060</a>, a QLC SSD with a pSLC region that was in customer validation as of August; it compared the design against “four dedicated SLC drives” in a 24-drive system, the kind of setup that would produce Phison’s 25% figure. Solidigm’s Cloud Storage Acceleration Layer works differently, with a host-side flash translation layer, where it can manage small writes on SLC or TLC drives before writing to QLC in favorable large sequential chunks.</p><p>There’s also the Enmotus FuzeDrive P200, previewed at CES 2020 as the MiDrive, which used a Phison controller and seems to be the closest to an ancestor of the current Mixed Media. That consumer drive, <a href="https://www.tomshardware.com/reviews/enmotus-fuzedrive-p200-m2-nvme-ssd-review">which we reviewed in 2021</a>, had SLC and QLC portions that acted distinctly, but they were managed by Enmotus software.</p><div ><table><caption>pSLC and QLC tiering approaches</caption><thead><tr><th class="firstcol " ><p>Design</p></th><th  ><p>Where the cache is</p></th><th  ><p>pSLC-to-QLC ratio</p></th><th  ><p>How it is managed</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Phison Mixed Media</p></td><td  ><p>In every drive, behind one 16-channel controller</p></td><td  ><p>Not stated</p></td><td  ><p>Separate NVMe endurance groups assigned to namespaces; no specialized host software, Phison says</p></td></tr><tr><td class="firstcol " ><p>Kioxia Mixed Mode</p></td><td  ><p>In every drive, with parity protection on the pSLC space across drives</p></td><td  ><p>1:4 as a “good beginning ratio”; customers typically want pSLC at 0.5% to 2% of maximum QLC capacity</p></td><td  ><p>Ratio set once, creating a pSLC namespace and endurance group</p></td></tr><tr><td class="firstcol " ><p>DapuStor dual-mode</p></td><td  ><p>In every drive, as a pSLC region beside the QLC</p></td><td  ><p>400GB, 800GB or 1.2TB of pSLC on a 30.72TB drive</p></td><td  ><p>Separate SLC and QLC block devices on one drive</p></td></tr><tr><td class="firstcol " ><p>Enmotus FuzeDrive P200</p></td><td  ><p>In the drive, as a fixed SLC portion of a consumer QLC SSD</p></td><td  ><p>24GB of SLC on the 900GB model</p></td><td  ><p>Enmotus’ own software</p></td></tr><tr><td class="firstcol " ><p>Solidigm CSAL</p></td><td  ><p>On dedicated SLC or TLC cache drives</p></td><td  ><p>Reference system: three 800GB SLC drives and seven 61.44TB QLC drives</p></td><td  ><p>Flash translation layer running on the host</p></td></tr></tbody></table></div><p>It’s also inevitable that these two approaches will be compared to Phison’s other upcoming part, the X3 enterprise SSD controller, a PCIe 6.0 design promising up to 28 GB/s of sequential throughput and 6.8 million IOPS. Those figures are <a href="https://www.tomshardware.com/pc-components/ssds/phison-shows-pcie-6-0-x3-ssd-controller-with-28-gb-s-of-bandwidth-and-6-8-million-iops-supports-2-petabytes-per-drive-also-new-power-sipping-e37t-ssds-for-pcie-5-0-systems-consume-a-mere-4-5w">unchanged since Computex in June</a>.</p><p>The X3’s IOPS sounds like a lot, but we’ve heard considerable banter about AI needing way more than that, as <a href="https://www.tomshardware.com/tech-industry/nvidia-and-kioxia-target-100-million-iops-ssd-in-2027-33-times-more-than-existing-drives-for-exclusive-use-in-ai-servers">Nvidia’s roadmap target for PCIe 7.0</a> is at 100 million IOPS, though that target counts 512-byte reads, a smaller unit than the 4KB that SSD ratings usually use, so the two figures aren’t directly comparable. Jean said at the panel that Phison is “very optimistic” it can achieve that. He added that the eightfold jump in IOPS that comes with moving from 4KB to 512-byte accesses would, on existing NAND, mean “eight times as much power,” or “roughly in the 140 to 180 watt range,” which is beyond reasonable, so perhaps a different solution will be required.</p><p>For the X3, Phison told us in June to expect customer samples by the end of the year, with volume coming in the middle of next year, though the SC26 release gives no dates. Notably, Micron and Samsung are <a href="https://www.tomshardware.com/tech-industry/the-current-state-of-pcie-6-0-ssds-and-controllers-marvell-phison-and-smi-prepare-controllers-as-drives-finally-come-to-market-following-years-of-delays">already offering PCIe 6.0 SSDs</a>, and the Bravera SC6 controller is expected from Marvell. Phison also says its PS7161, which its site lists as a customer sample, is “the first PCIe 6.0 redriver on the PCI-SIG Integrators List.”</p><p>Of the two designs, Mixed Media looks like the nearer product. It builds on standard NVMe features, and DapuStor’s similar dual-mode drive reached customer testing this summer, while the one processor figure given for the computational-storage drive says little about what it can do for AI. What remains to be seen for Mixed Media is its pSLC-to-QLC ratio and how configurable it is. CES and Computex are the notable dates on the 2027 calendar to find out more about both, as well as the X3.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Power and cooling limits force AI clusters with 100,000+ accelerators to use multiple data centers  ]]></title>
                                                                                                <dc:content><![CDATA[ <p>The rapid growth of AI clusters is pushing computing infrastructure beyond the physical and power limitations of individual data centers, which creates demand for new networking architectures capable of connecting compute resources across multiple facilities and even geographical regions. At Supercomputing 2026 (<a href="https://www.tomshardware.com/tag/sc26" target="_blank">SC26</a>), researchers from The Ohio State University discussed the evolution of high-performance networking technologies. At the conference, they explained how efficient distributed computing at such a scale requires considerably more than faster network connections, highlighting the importance of scale across domains to stretch workloads across different data centers in separate geographical regions.</p><h2 id="massive-systems-call-for-new-interconnects">Massive systems call for new interconnects</h2><p>Modern AI clusters typically accommodate more than 100,000 AI accelerators and potentially millions of CPU endpoints, placing tremendous demands on network bandwidth, latency, and congestion management. Meanwhile, an accelerator's computational performance continues to increase faster than communication capabilities, which means that faster AI GPUs do not necessarily deliver proportional improvements in application performance. As a result, the efficiency of network infrastructure determines how effectively large-scale AI systems can use their computational resources. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:51.68%;"><img id="RXveXRVtEG277SibpDosrC" name="Screenshot 2026-11-19 at 01.55.21" alt="The Ohio State University" src="https://cdn.mos.cms.futurecdn.net/RXveXRVtEG277SibpDosrC-1920-80.png" mos="" align="middle" fullscreen="" width="2560" height="1323" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: The Ohio State University)</span></figcaption></figure><p>Traditionally, high-performance computing systems have relied on two major interconnect domains: scale-up and scale-out. Scale-up networks connect processors and accelerators within tightly integrated systems using technologies like AMD's Infinity Fabric, Intel's UPI, Nvidia's NVLink, or emerging alternatives such as UALink. Scale-out networks, which typically rely on Ethernet, InfiniBand, HPE Slingshot, or, more recently, Ultra Ethernet, connect individual servers or rack-scale systems into larger clusters. These interconnects have different requirements for bandwidth, latency, and scalability, which is why they serve different purposes. </p><h2 id="scale-across-interconnects-are-almost-here">Scale-across interconnects are (almost) here</h2><p>However, the massive size and power consumption of modern AI clusters create demand for a third interconnect domain, known as scale-across. Unlike scale-out networks, which typically connect systems within a single data center, scale-across architectures extend high-performance connectivity between different buildings, campuses, or even regions. </p><p>Such an approach enables operators to combine computational and other resources from multiple locations to achieve a certain level of performance rather than to bring an excessive number of accelerators to a single data center, which may be hard, or even impossible, due to power availability, cooling requirements, or physical space constraints. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:54.18%;"><img id="K7djZu3iHCSqDGqUosmJZC" name="Screenshot 2026-11-19 at 01.54.35" alt="The Ohio State University" src="https://cdn.mos.cms.futurecdn.net/K7djZu3iHCSqDGqUosmJZC-1920-80.png" mos="" align="middle" fullscreen="" width="2560" height="1387" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: The Ohio State University)</span></figcaption></figure><p>The transition toward scale-across architectures has already begun. Google introduced its Virgo network architecture earlier this year to support AI workloads that exceed the space and power capacity of individual data centers. The company intends to treat an entire campus of computing facilities as a single system, albeit with specialized networks for communication between accelerators, servers, and data centers. Cisco and Marvell are also developing technologies for scale-across deployments, including high-capacity routers, switching chips, and optical interconnects designed to connect geographically distributed clusters.</p><h2 id="not-without-caveats">Not without caveats</h2><p>Extending AI infrastructure beyond individual facilities introduces substantial technical challenges. While modern scale-up networks offer extremely high bandwidth and low latency over relatively short distances, communication between data centers must contend with the physical limitations of long-distance connections. Even optical networks cannot eliminate propagation delays, whereas congestion, packet loss, and synchronization overhead can further reduce effective performance. Consequently, connecting two large GPU clusters with high-speed links does not necessarily make them behave like one tightly integrated computing system. </p><p>These challenges are particularly important for AI training, where accelerators frequently exchange model parameters, gradients, and intermediate results. Large-scale workloads require efficient collective communications, as delays that affect individual accelerators dramatically reduce the performance of the entire cluster. Therefore, the efficiency of distributed computing depends not only on raw network bandwidth, but also on network topology, congestion management, communication protocols, and software implementations.</p><h2 id="more-sophisticated-hardware-and-software-needed">More sophisticated hardware and software needed</h2><p>To address these limitations, network adapters and switches are becoming more sophisticated than ever. Modern network interface controllers can perform Remote Direct Memory Access (RDMA) operations without continuous CPU involvement, whereas technologies such as HPE Slingshot hardware collectives and Nvidia SHARP can offload certain communication operations to the network infrastructure itself. SmartNICs and data processing units further reduce the burden on host processors as they handle selected network and infrastructure functions in dedicated hardware. Such capabilities can improve efficiency of distributed applications without necessarily requiring faster physical interconnects.</p><p>Interestingly, software plays an equally important role in the performance of large-scale AI infrastructure, as communication frameworks like libfabric, MPI, or UCX determine how applications exchange data and use the capabilities of modern network adapters. Also, GPU-aware communication and technologies like GPUDirect RDMA can eliminate unnecessary transfers through host memory, while optimized collective operations can reduce communication overhead across large numbers of accelerators, according to the researchers.  </p><p>The growing complexity of AI infrastructure also changes how operators should evaluate network performance. Although the industry continues to introduce faster interconnect technologies, including versions of GbE with nominal bandwidths of 800 Gb/s and 1.6 Tb/s, peak transfer rates alone do not fully describe the capabilities of a given system because the performance of actual applications depends on multiple factors, including how effectively network hardware and software cooperate to transfer data, prevent congestion, and synchronize computations across thousands of processors.</p><h2 id="the-evolution-continues">The evolution continues</h2><p>Ultimately, the emergence of scale-across interconnection architectures reflects a fundamental change in the design of AI infrastructure. As the computational capabilities of individual accelerators increase and AI clusters expand beyond individual facilities, efficient communication becomes just as important as processor performance. As a result, the next generation of AI systems will depend not only on faster AI accelerators or CPUs, but also on sophisticated network adapters, switches, and, perhaps more importantly, software that can unite geographically distributed computational resources without excessive communication overhead.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/supercomputers/power-and-cooling-limits-push-100-000-accelerator-ai-clusters-past-single-data-centers-google-cisco-and-marvell-target-scale-across-networking-across-regions</link>
                                                                            <description>
                            <![CDATA[ Faster GPUs are no longer enough as power constraints, network congestion, and communication overhead force the industry to rethink how massive AI systems are built. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">qYj8hmCTNbDS3XM2HSsk63</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/MQEF3FGDnEs6gC4KtkGjKG-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Fri, 20 Nov 2026 11:30:00 +0000</pubDate>                                                                                                                                <updated>Fri, 20 Nov 2026 14:09:27 +0000</updated>
                                                                                                                                            <category><![CDATA[Supercomputers]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                <author><![CDATA[ ashilov@gmail.com (Anton Shilov) ]]></author>                    <dc:creator><![CDATA[ Anton Shilov ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/uMZ5kNphxA2Ut6whdLaSQV-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Anton Shilov has been in the PC industry since 1990s playing games, building PCs, and writing stories about pretty much everything that relates to PCs, Macs, smartphones, tablets, and even fab equipment. Over his career, he has worked at a variety of high-ranking websites, including AnandTech, EE Times, TechRadar, X-bit Labs, and now Tom&#039;s Hardware. He is also a regular features contributor to Tom&#039;s Hardware Premium, writing about the latest developments in the semiconductor industry and related tech news and roadmaps. When Anton is not reading or writing about something high-tech, he is probably watching a good movie, playing a video game, or spending time with his family.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>false</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/MQEF3FGDnEs6gC4KtkGjKG-1920-80.jpg">
                                                            <media:credit><![CDATA[Google]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Google]]></media:description>                                                            <media:text><![CDATA[Google]]></media:text>
                                <media:title type="plain"><![CDATA[Google]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/MQEF3FGDnEs6gC4KtkGjKG-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>The rapid growth of AI clusters is pushing computing infrastructure beyond the physical and power limitations of individual data centers, which creates demand for new networking architectures capable of connecting compute resources across multiple facilities and even geographical regions. At Supercomputing 2026 (<a href="https://www.tomshardware.com/tag/sc26" target="_blank">SC26</a>), researchers from The Ohio State University discussed the evolution of high-performance networking technologies. At the conference, they explained how efficient distributed computing at such a scale requires considerably more than faster network connections, highlighting the importance of scale across domains to stretch workloads across different data centers in separate geographical regions.</p><h2 id="massive-systems-call-for-new-interconnects">Massive systems call for new interconnects</h2><p>Modern AI clusters typically accommodate more than 100,000 AI accelerators and potentially millions of CPU endpoints, placing tremendous demands on network bandwidth, latency, and congestion management. Meanwhile, an accelerator's computational performance continues to increase faster than communication capabilities, which means that faster AI GPUs do not necessarily deliver proportional improvements in application performance. As a result, the efficiency of network infrastructure determines how effectively large-scale AI systems can use their computational resources. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:51.68%;"><img id="RXveXRVtEG277SibpDosrC" name="Screenshot 2026-11-19 at 01.55.21" alt="The Ohio State University" src="https://cdn.mos.cms.futurecdn.net/RXveXRVtEG277SibpDosrC-1920-80.png" mos="" align="middle" fullscreen="" width="2560" height="1323" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: The Ohio State University)</span></figcaption></figure><p>Traditionally, high-performance computing systems have relied on two major interconnect domains: scale-up and scale-out. Scale-up networks connect processors and accelerators within tightly integrated systems using technologies like AMD's Infinity Fabric, Intel's UPI, Nvidia's NVLink, or emerging alternatives such as UALink. Scale-out networks, which typically rely on Ethernet, InfiniBand, HPE Slingshot, or, more recently, Ultra Ethernet, connect individual servers or rack-scale systems into larger clusters. These interconnects have different requirements for bandwidth, latency, and scalability, which is why they serve different purposes. </p><h2 id="scale-across-interconnects-are-almost-here">Scale-across interconnects are (almost) here</h2><p>However, the massive size and power consumption of modern AI clusters create demand for a third interconnect domain, known as scale-across. Unlike scale-out networks, which typically connect systems within a single data center, scale-across architectures extend high-performance connectivity between different buildings, campuses, or even regions. </p><p>Such an approach enables operators to combine computational and other resources from multiple locations to achieve a certain level of performance rather than to bring an excessive number of accelerators to a single data center, which may be hard, or even impossible, due to power availability, cooling requirements, or physical space constraints. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:54.18%;"><img id="K7djZu3iHCSqDGqUosmJZC" name="Screenshot 2026-11-19 at 01.54.35" alt="The Ohio State University" src="https://cdn.mos.cms.futurecdn.net/K7djZu3iHCSqDGqUosmJZC-1920-80.png" mos="" align="middle" fullscreen="" width="2560" height="1387" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: The Ohio State University)</span></figcaption></figure><p>The transition toward scale-across architectures has already begun. Google introduced its Virgo network architecture earlier this year to support AI workloads that exceed the space and power capacity of individual data centers. The company intends to treat an entire campus of computing facilities as a single system, albeit with specialized networks for communication between accelerators, servers, and data centers. Cisco and Marvell are also developing technologies for scale-across deployments, including high-capacity routers, switching chips, and optical interconnects designed to connect geographically distributed clusters.</p><h2 id="not-without-caveats">Not without caveats</h2><p>Extending AI infrastructure beyond individual facilities introduces substantial technical challenges. While modern scale-up networks offer extremely high bandwidth and low latency over relatively short distances, communication between data centers must contend with the physical limitations of long-distance connections. Even optical networks cannot eliminate propagation delays, whereas congestion, packet loss, and synchronization overhead can further reduce effective performance. Consequently, connecting two large GPU clusters with high-speed links does not necessarily make them behave like one tightly integrated computing system. </p><p>These challenges are particularly important for AI training, where accelerators frequently exchange model parameters, gradients, and intermediate results. Large-scale workloads require efficient collective communications, as delays that affect individual accelerators dramatically reduce the performance of the entire cluster. Therefore, the efficiency of distributed computing depends not only on raw network bandwidth, but also on network topology, congestion management, communication protocols, and software implementations.</p><h2 id="more-sophisticated-hardware-and-software-needed">More sophisticated hardware and software needed</h2><p>To address these limitations, network adapters and switches are becoming more sophisticated than ever. Modern network interface controllers can perform Remote Direct Memory Access (RDMA) operations without continuous CPU involvement, whereas technologies such as HPE Slingshot hardware collectives and Nvidia SHARP can offload certain communication operations to the network infrastructure itself. SmartNICs and data processing units further reduce the burden on host processors as they handle selected network and infrastructure functions in dedicated hardware. Such capabilities can improve efficiency of distributed applications without necessarily requiring faster physical interconnects.</p><p>Interestingly, software plays an equally important role in the performance of large-scale AI infrastructure, as communication frameworks like libfabric, MPI, or UCX determine how applications exchange data and use the capabilities of modern network adapters. Also, GPU-aware communication and technologies like GPUDirect RDMA can eliminate unnecessary transfers through host memory, while optimized collective operations can reduce communication overhead across large numbers of accelerators, according to the researchers.  </p><p>The growing complexity of AI infrastructure also changes how operators should evaluate network performance. Although the industry continues to introduce faster interconnect technologies, including versions of GbE with nominal bandwidths of 800 Gb/s and 1.6 Tb/s, peak transfer rates alone do not fully describe the capabilities of a given system because the performance of actual applications depends on multiple factors, including how effectively network hardware and software cooperate to transfer data, prevent congestion, and synchronize computations across thousands of processors.</p><h2 id="the-evolution-continues">The evolution continues</h2><p>Ultimately, the emergence of scale-across interconnection architectures reflects a fundamental change in the design of AI infrastructure. As the computational capabilities of individual accelerators increase and AI clusters expand beyond individual facilities, efficient communication becomes just as important as processor performance. As a result, the next generation of AI systems will depend not only on faster AI accelerators or CPUs, but also on sophisticated network adapters, switches, and, perhaps more importantly, software that can unite geographically distributed computational resources without excessive communication overhead.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Silicon Motion demos SM8466 PCIe 6.0 SSD controller, expected to ship by 2028 — controller boasts improved quality of service, samples imminent ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Silicon Motion is showing off its <a href="https://www.tomshardware.com/tech-industry/the-current-state-of-pcie-6-0-ssds-and-controllers-marvell-phison-and-smi-prepare-controllers-as-drives-finally-come-to-market-following-years-of-delays">PCIe 6.0</a> SSD controller for data center applications for the first time at the <a href="https://www.tomshardware.com/tag/sc26">SC26 conference in Chicago</a>. The final iteration of Silicon Motion's SM8466 controller was taped out recently, and drives based on the chip are due in late 2027 or early 2028. However, the new controller also has a secret weapon designed to ensure consistent performance of SSDs under concurrent workloads that may set it apart from competitors. </p><h2 id="the-new-controller">The new controller</h2><p>Silicon Motion introduced its high-end MonTitan SM8466 enterprise SSD controller designed for next-generation data center and hyperscale storage applications <a href="https://www.tomshardware.com/pc-components/ssds/silicon-motion-reportedly-prepping-sm8466-ssd-controller-with-a-pcie-6-0-x4-leak-claims-it-will-be-unveiled-at-fms-2025-sporting-speeds-of-up-to-28gb-s">at FMS 2025</a>. The controller features a PCIe 6.x x4 host interface and 16 NAND channels to enable sequential transfer speeds of up to 28 GB/s, random performance of up to 7 million 4K IOPS, and capacities of up to 512 TB and beyond as newer 3D NAND devices emerge. Just like other high-end data center-grade SSD controllers released in recent years, the SM8466 complies with NVMe 2.0 and OCP NVMe 2.5 SSD specifications.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="DMLwAW3aC4KBHD255d4Y8R" name="smi-sm8466" alt="Silicon Motion" src="https://cdn.mos.cms.futurecdn.net/DMLwAW3aC4KBHD255d4Y8R-1920-80.jpg" mos="" align="middle" fullscreen="" width="2560" height="1440" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Silicon Motion)</span></figcaption></figure><p>To improve performance and reduce latency, Silicon Motion has developed its Separate Command Architecture (SCA), which separates command and address transmission paths within the NAND interface to improve parallelism and increase effective throughput. Additionally, the controller supports Silicon Motion's 2<sup>nd</sup> Generation PerformaShape technology, which enables programmable performance limits and predictable quality of service (QoS) across concurrent workloads, which is arguably its most important differentiator against rivals who rely on different methods of guaranteeing performance.</p><p>In addition to its performance chops, the MonTitan SM8466 incorporates advanced error correction, virtualization, data protection, and security features. The controller's error correction engine supports low-density parity-check (LDPC) codes with 4KiB codewords and Silicon Motion's 16KiB Collaborative Codeword (CoCo) technology to improve data integrity and reliability with advanced NAND flash memory. The controller also supports Single Root I/O Virtualization (SR-IOV), MPF virtualization, SMART diagnostics, and comprehensive data protection mechanisms. Security capabilities include Secure Boot, AES-256 encryption, TCG Opal compatibility, and firmware attestation. Together, these features make the SM8466 suitable for high-capacity data center drives that are expected to deliver high performance, predictable latency, and reliable operation under demanding data center workloads.</p><h2 id="the-tactical-advantage">The tactical advantage</h2><p>Silicon Motion's 2<sup>nd</sup> Generation PerformaShape technology builds upon the original PerformaShape architecture introduced with the company's PCIe 5.0-based MonTitan SM8366 enterprise SSD controller several years ago. The 1<sup>st</sup> Generation implementation uses dedicated hardware accelerators and two-stage leaky-bucket regulators to control read and write performance in terms of IOPS and throughput, both for individual workloads and for the SSD as a whole. This approach enables data center operators to prevent one workload from monopolizing SSD resources at the expense of others.  </p><p>PerformaShape 2 retains these capabilities but doubles the number of supported quality-of-service (QoS) sets from 256 to 512 and introduces more sophisticated performance controls to accommodate a larger number of concurrent workloads with different requirements. One of the key additions to PerformaShape 2 is Multi-Dimensional Shaping, which enables the controller to coordinate performance limits across multiple I/O characteristics rather than enforce each limit independently. This gives data center operators greater flexibility when allocating SSD resources among applications with different performance requirements. </p><p>Another important enhancement of PerformaShape 2 is support for NVMe TP4176, a proposed extension to the NVMe specification that aims to standardize performance management and QoS controls. Whereas the original PerformaShape relies on Silicon Motion's vendor-specific NVMe commands and a dedicated host driver, TP4176 could eventually enable compatible host software to configure performance policies through a standardized mechanism. The technology is particularly relevant for cloud environments where SSDs serve numerous virtual machines, containers, and applications whose storage requirements can change considerably over time, according to the company.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/pc-components/ssds/silicon-motion-demos-sm8466-pcie-6-0-ssd-controller-expected-to-ship-by-2028-controller-boasts-improved-quality-of-service-samples-imminent</link>
                                                                            <description>
                            <![CDATA[ Silicon Motion's SM8466 has finally taped out and is set to enter the market in 2027-2028. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">vAgH6mcUiKD4L7SobyhVr9</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/qcTaJHoZfFpj8LgLvN26X3-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Thu, 19 Nov 2026 12:00:00 +0000</pubDate>                                                                                                                                                                                                                                <category><![CDATA[SSDs]]></category>
                                                    <category><![CDATA[PC Components]]></category>
                                                    <category><![CDATA[Storage]]></category>
                                                                                                <author><![CDATA[ ashilov@gmail.com (Anton Shilov) ]]></author>                    <dc:creator><![CDATA[ Anton Shilov ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/uMZ5kNphxA2Ut6whdLaSQV-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Anton Shilov has been in the PC industry since 1990s playing games, building PCs, and writing stories about pretty much everything that relates to PCs, Macs, smartphones, tablets, and even fab equipment. Over his career, he has worked at a variety of high-ranking websites, including AnandTech, EE Times, TechRadar, X-bit Labs, and now Tom&#039;s Hardware. He is also a regular features contributor to Tom&#039;s Hardware Premium, writing about the latest developments in the semiconductor industry and related tech news and roadmaps. When Anton is not reading or writing about something high-tech, he is probably watching a good movie, playing a video game, or spending time with his family.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>false</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/qcTaJHoZfFpj8LgLvN26X3-1920-80.jpg">
                                                            <media:credit><![CDATA[Tom&amp;#39;s Hardware]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Silicon Motion]]></media:description>                                                            <media:text><![CDATA[Silicon Motion]]></media:text>
                                <media:title type="plain"><![CDATA[Silicon Motion]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/qcTaJHoZfFpj8LgLvN26X3-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Silicon Motion is showing off its <a href="https://www.tomshardware.com/tech-industry/the-current-state-of-pcie-6-0-ssds-and-controllers-marvell-phison-and-smi-prepare-controllers-as-drives-finally-come-to-market-following-years-of-delays">PCIe 6.0</a> SSD controller for data center applications for the first time at the <a href="https://www.tomshardware.com/tag/sc26">SC26 conference in Chicago</a>. The final iteration of Silicon Motion's SM8466 controller was taped out recently, and drives based on the chip are due in late 2027 or early 2028. However, the new controller also has a secret weapon designed to ensure consistent performance of SSDs under concurrent workloads that may set it apart from competitors. </p><h2 id="the-new-controller">The new controller</h2><p>Silicon Motion introduced its high-end MonTitan SM8466 enterprise SSD controller designed for next-generation data center and hyperscale storage applications <a href="https://www.tomshardware.com/pc-components/ssds/silicon-motion-reportedly-prepping-sm8466-ssd-controller-with-a-pcie-6-0-x4-leak-claims-it-will-be-unveiled-at-fms-2025-sporting-speeds-of-up-to-28gb-s">at FMS 2025</a>. The controller features a PCIe 6.x x4 host interface and 16 NAND channels to enable sequential transfer speeds of up to 28 GB/s, random performance of up to 7 million 4K IOPS, and capacities of up to 512 TB and beyond as newer 3D NAND devices emerge. Just like other high-end data center-grade SSD controllers released in recent years, the SM8466 complies with NVMe 2.0 and OCP NVMe 2.5 SSD specifications.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="DMLwAW3aC4KBHD255d4Y8R" name="smi-sm8466" alt="Silicon Motion" src="https://cdn.mos.cms.futurecdn.net/DMLwAW3aC4KBHD255d4Y8R-1920-80.jpg" mos="" align="middle" fullscreen="" width="2560" height="1440" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Silicon Motion)</span></figcaption></figure><p>To improve performance and reduce latency, Silicon Motion has developed its Separate Command Architecture (SCA), which separates command and address transmission paths within the NAND interface to improve parallelism and increase effective throughput. Additionally, the controller supports Silicon Motion's 2<sup>nd</sup> Generation PerformaShape technology, which enables programmable performance limits and predictable quality of service (QoS) across concurrent workloads, which is arguably its most important differentiator against rivals who rely on different methods of guaranteeing performance.</p><p>In addition to its performance chops, the MonTitan SM8466 incorporates advanced error correction, virtualization, data protection, and security features. The controller's error correction engine supports low-density parity-check (LDPC) codes with 4KiB codewords and Silicon Motion's 16KiB Collaborative Codeword (CoCo) technology to improve data integrity and reliability with advanced NAND flash memory. The controller also supports Single Root I/O Virtualization (SR-IOV), MPF virtualization, SMART diagnostics, and comprehensive data protection mechanisms. Security capabilities include Secure Boot, AES-256 encryption, TCG Opal compatibility, and firmware attestation. Together, these features make the SM8466 suitable for high-capacity data center drives that are expected to deliver high performance, predictable latency, and reliable operation under demanding data center workloads.</p><h2 id="the-tactical-advantage">The tactical advantage</h2><p>Silicon Motion's 2<sup>nd</sup> Generation PerformaShape technology builds upon the original PerformaShape architecture introduced with the company's PCIe 5.0-based MonTitan SM8366 enterprise SSD controller several years ago. The 1<sup>st</sup> Generation implementation uses dedicated hardware accelerators and two-stage leaky-bucket regulators to control read and write performance in terms of IOPS and throughput, both for individual workloads and for the SSD as a whole. This approach enables data center operators to prevent one workload from monopolizing SSD resources at the expense of others.  </p><p>PerformaShape 2 retains these capabilities but doubles the number of supported quality-of-service (QoS) sets from 256 to 512 and introduces more sophisticated performance controls to accommodate a larger number of concurrent workloads with different requirements. One of the key additions to PerformaShape 2 is Multi-Dimensional Shaping, which enables the controller to coordinate performance limits across multiple I/O characteristics rather than enforce each limit independently. This gives data center operators greater flexibility when allocating SSD resources among applications with different performance requirements. </p><p>Another important enhancement of PerformaShape 2 is support for NVMe TP4176, a proposed extension to the NVMe specification that aims to standardize performance management and QoS controls. Whereas the original PerformaShape relies on Silicon Motion's vendor-specific NVMe commands and a dedicated host driver, TP4176 could eventually enable compatible host software to configure performance policies through a standardized mechanism. The technology is particularly relevant for cloud environments where SSDs serve numerous virtual machines, containers, and applications whose storage requirements can change considerably over time, according to the company.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Intel Diamond Rapids shows signs of life in SC26 demo running live server ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Intel’s next-gen Xeon Diamond Rapids CPUs appeared at Supercomputing 2026 (<a href="https://www.tomshardware.com/tag/sc26">SC26</a>) in two separate booths, but neither was from Intel directly. Most interesting was a running server at the CXL booth, showcasing a demo with a Diamond Rapids CPU. The server ran for the first day of the show, and upon returning to the booth, we were informed that it had been shut down due to noise complaints (it’s hard to mistake a running server). Elsewhere at Bull’s booth, a Diamond Rapids CPU was shown installed in a server behind a glass case. Notably, Intel didn’t have a Diamond Rapids demo, instead focusing most of its booth on Xeon 6+ and its new D270 GPU, <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/intels-crescent-island-ai-accelerator-officially-powers-new-gpu-d-series-products-chip-delivers-1-3-peak-pflops-of-fp8-compute-and-1-5-tb-s-of-memory-bandwidth"><u>otherwise known as Crescent Island</u></a>. </p><p>The demo in question is for Montage’s CXL MXC, or Memory eXpander Controller. If you’re unfamiliar, memory expansion via CXL is an increasingly common tactic to give multi-host environments access to a large pool of memory, breaking free of an individual server’s capacity via CXL. In the demo, Montage is showing off its memory expansion controller running across a Xeon 6 and a Diamond Rapids server, each sharing a pool of memory over CXL. CXL is an open standard, built atop PCIe. </p><figure role="gallery"><figure><img src="https://cdn.mos.cms.futurecdn.net/kHXpNs8MuRKsaEtRKs3JF5-1920-80.jpg" alt="Intel Diamond Rapids Demo. " /><figcaption><small role="credit">Tom's Hardware</small></figcaption></figure><figure><img src="https://cdn.mos.cms.futurecdn.net/pWfhVnLSkYUW5qijhHRPp8-1920-80.jpg" alt="Intel Diamond Rapids Demo. " /><figcaption><small role="credit">Tom's Hardware</small></figcaption></figure></figure><p>Of course, the demo doesn’t say much about Diamond Rapids performance, but it’s the first time we’ve seen a chip running in the wild. The 256 GB memory capacity shown doesn’t provide any hints, either. Intel has announced that Diamond Rapids will support 16-channel memory, with RDIMM support up to 8,000 MT/s and MRDIMM support up to 12,800 MT/s, matching <a href="https://www.tomshardware.com/pc-components/cpus/amd-zen-6-epyc-venice-deep-dive-and-benchmarking-dual-epyc-9996-with-1-024-threads-delivers-up-to-93-percent-performance-gain-gen-on-gen-9686f-outclasses-nvidias-vera"><u>AMD’s recent EPYC Venice CPUs</u></a>. However, Intel hasn’t revealed capacity, though we anticipate the maximum capacity will be far above 256 GB – for context, last-gen Granite Rapids-AP supports up to 3 TB per socket. </p><p>Bull is a French server designer focused on HPC and AI infrastructure, and it showed off two Xeon 7 Diamond Rapids chips installed in the Bullsquana XH31600 server. As expected for a 2P Diamond Rapids server, it supports 32 channels of DDR5 memory, and it’s designed for a liquid-cooled deployment.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2457px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="ZJSVRpaCtkGTUSuxavfnsC" name="IMG20261117115237" alt="Diamond Rapids server CPUs." src="https://cdn.mos.cms.futurecdn.net/ZJSVRpaCtkGTUSuxavfnsC-1920-80.jpg" mos="" align="middle" fullscreen="" width="2457" height="1382" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p>As you can see from the image, these aren’t production chips. They’re labeled with “Intel confidential,” and the code starts with “Q,” revealing that this is a sample. Originally, we expected to see Diamond Rapids in the second half of this year, perhaps around this time. However, in August, Intel revealed that Xeon 7 chips would launch in 2027. </p><p>Based on the various servers we saw and the live demo, it seems Intel is targeting a launch earlier in the year, likely in Q2; we expect a larger consumer focus in Q1. It could slip later into the year, though that seems very unlikely given that Diamond Rapids was already pushed back, and the insatiable demand for CPUs in AI data centers doesn’t appear to be drying up. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="W4WCaZBckuhGj6fxnMmRLM" name="IMG20261118135714" alt="Intel Diamond Rapids server built by Asus." src="https://cdn.mos.cms.futurecdn.net/W4WCaZBckuhGj6fxnMmRLM-1920-80.jpg" mos="" align="middle" fullscreen="" width="2560" height="1440" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p>Also pointing to a relatively quick release are the various servers we saw on the floor at SC26. Several partners showed off their next-gen servers built for Diamond Rapids CPU, sporting the LGA 9324 socket, including Dell and Asus (pictured above). However, Bull was the only vendor we could find that had an actual chip on display. </p><p>Diamond Rapids is Intel’s true next-gen CPU, following up on the E-core-only Xeon 6+ range that launched earlier this year. Diamond Rapids packs up to 256 P-cores, and it’s built on Intel’s performance-enhanced 18A-P node. The most interesting aspect of the chip is its topology, with Intel breaking from its previous layout and building something much similar to an EPYC CPU. The package features centralized memory and I/O, which were pushed out to the edges of the package with Granite Rapids. You can read more about the design of the chip in our <a href="https://www.tomshardware.com/pc-components/cpus/intel-xeon-7-diamond-rapids-comes-with-up-to-256-p-cores-1-28-gb-of-last-level-cache-next-gen-18a-p-cpu-also-brings-avx-10-2-and-uses-ucie-s-instead-of-emib"><u>Diamond Rapids recap from Hot Chips on </u><u><em>Tom’s Hardware Premium</em></u></a><em>. </em></p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/pc-components/cpus/intel-diamond-rapids-shows-signs-of-life-in-sc26-demo-running-live-server-next-gen-xeon-7-cpu-breaks-cover-at-partner-booths-no-showcase-from-intel</link>
                                                                            <description>
                            <![CDATA[ Intel’s next-gen Xeon Diamond Rapids CPUs appeared at two booths during SC26, one of which featured a running server, but neither were directly from Intel. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">QUGfiVTtNbA94fqwVQH8hQ</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/c6wWW88xn4YtKrcAiGix3h-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Thu, 19 Nov 2026 11:30:00 +0000</pubDate>                                                                                                                                                                                                                                <category><![CDATA[CPUs]]></category>
                                                    <category><![CDATA[PC Components]]></category>
                                                                                                                    <dc:creator><![CDATA[ Jake Roach ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/h6PRM8bTimCTnNfoAYfjAi-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Jake Roach has been bending pins and busting solder joints since the mid-2000s. From trying to run scratched CDs of &lt;em&gt;Delta Force &lt;/em&gt;and &lt;em&gt;Unreal Tournament &lt;/em&gt;to spitting out virtual machines on a Threadripper, Jake has been on the hunt for the latest hardware and highest performance for decades. That eventually spun up a career, with Jake serving as Lead Reporter at Digital Trends, as well as contributing to outlets like XDA, PC Invasion, Business Insider, and WIRED. At Tom’s Hardware, Jake is focused on consumer and workstation CPUs. Outside working hours, you’ll find him knee-deep in the latest roguelite taking over Steam, spending way too much money on &lt;em&gt;Magic: The Gathering, &lt;/em&gt;or forcing his lazy corgi onto walks.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>false</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/c6wWW88xn4YtKrcAiGix3h-1920-80.jpg">
                                                            <media:credit><![CDATA[Tom&amp;#39;s Hardware]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Intel Diamond Rapids CPU. ]]></media:description>                                                            <media:text><![CDATA[Intel Diamond Rapids CPU. ]]></media:text>
                                <media:title type="plain"><![CDATA[Intel Diamond Rapids CPU. ]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/c6wWW88xn4YtKrcAiGix3h-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Intel’s next-gen Xeon Diamond Rapids CPUs appeared at Supercomputing 2026 (<a href="https://www.tomshardware.com/tag/sc26">SC26</a>) in two separate booths, but neither was from Intel directly. Most interesting was a running server at the CXL booth, showcasing a demo with a Diamond Rapids CPU. The server ran for the first day of the show, and upon returning to the booth, we were informed that it had been shut down due to noise complaints (it’s hard to mistake a running server). Elsewhere at Bull’s booth, a Diamond Rapids CPU was shown installed in a server behind a glass case. Notably, Intel didn’t have a Diamond Rapids demo, instead focusing most of its booth on Xeon 6+ and its new D270 GPU, <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/intels-crescent-island-ai-accelerator-officially-powers-new-gpu-d-series-products-chip-delivers-1-3-peak-pflops-of-fp8-compute-and-1-5-tb-s-of-memory-bandwidth"><u>otherwise known as Crescent Island</u></a>. </p><p>The demo in question is for Montage’s CXL MXC, or Memory eXpander Controller. If you’re unfamiliar, memory expansion via CXL is an increasingly common tactic to give multi-host environments access to a large pool of memory, breaking free of an individual server’s capacity via CXL. In the demo, Montage is showing off its memory expansion controller running across a Xeon 6 and a Diamond Rapids server, each sharing a pool of memory over CXL. CXL is an open standard, built atop PCIe. </p><figure role="gallery"><figure><img src="https://cdn.mos.cms.futurecdn.net/kHXpNs8MuRKsaEtRKs3JF5-1920-80.jpg" alt="Intel Diamond Rapids Demo. " /><figcaption><small role="credit">Tom's Hardware</small></figcaption></figure><figure><img src="https://cdn.mos.cms.futurecdn.net/pWfhVnLSkYUW5qijhHRPp8-1920-80.jpg" alt="Intel Diamond Rapids Demo. " /><figcaption><small role="credit">Tom's Hardware</small></figcaption></figure></figure><p>Of course, the demo doesn’t say much about Diamond Rapids performance, but it’s the first time we’ve seen a chip running in the wild. The 256 GB memory capacity shown doesn’t provide any hints, either. Intel has announced that Diamond Rapids will support 16-channel memory, with RDIMM support up to 8,000 MT/s and MRDIMM support up to 12,800 MT/s, matching <a href="https://www.tomshardware.com/pc-components/cpus/amd-zen-6-epyc-venice-deep-dive-and-benchmarking-dual-epyc-9996-with-1-024-threads-delivers-up-to-93-percent-performance-gain-gen-on-gen-9686f-outclasses-nvidias-vera"><u>AMD’s recent EPYC Venice CPUs</u></a>. However, Intel hasn’t revealed capacity, though we anticipate the maximum capacity will be far above 256 GB – for context, last-gen Granite Rapids-AP supports up to 3 TB per socket. </p><p>Bull is a French server designer focused on HPC and AI infrastructure, and it showed off two Xeon 7 Diamond Rapids chips installed in the Bullsquana XH31600 server. As expected for a 2P Diamond Rapids server, it supports 32 channels of DDR5 memory, and it’s designed for a liquid-cooled deployment.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2457px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="ZJSVRpaCtkGTUSuxavfnsC" name="IMG20261117115237" alt="Diamond Rapids server CPUs." src="https://cdn.mos.cms.futurecdn.net/ZJSVRpaCtkGTUSuxavfnsC-1920-80.jpg" mos="" align="middle" fullscreen="" width="2457" height="1382" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p>As you can see from the image, these aren’t production chips. They’re labeled with “Intel confidential,” and the code starts with “Q,” revealing that this is a sample. Originally, we expected to see Diamond Rapids in the second half of this year, perhaps around this time. However, in August, Intel revealed that Xeon 7 chips would launch in 2027. </p><p>Based on the various servers we saw and the live demo, it seems Intel is targeting a launch earlier in the year, likely in Q2; we expect a larger consumer focus in Q1. It could slip later into the year, though that seems very unlikely given that Diamond Rapids was already pushed back, and the insatiable demand for CPUs in AI data centers doesn’t appear to be drying up. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="W4WCaZBckuhGj6fxnMmRLM" name="IMG20261118135714" alt="Intel Diamond Rapids server built by Asus." src="https://cdn.mos.cms.futurecdn.net/W4WCaZBckuhGj6fxnMmRLM-1920-80.jpg" mos="" align="middle" fullscreen="" width="2560" height="1440" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p>Also pointing to a relatively quick release are the various servers we saw on the floor at SC26. Several partners showed off their next-gen servers built for Diamond Rapids CPU, sporting the LGA 9324 socket, including Dell and Asus (pictured above). However, Bull was the only vendor we could find that had an actual chip on display. </p><p>Diamond Rapids is Intel’s true next-gen CPU, following up on the E-core-only Xeon 6+ range that launched earlier this year. Diamond Rapids packs up to 256 P-cores, and it’s built on Intel’s performance-enhanced 18A-P node. The most interesting aspect of the chip is its topology, with Intel breaking from its previous layout and building something much similar to an EPYC CPU. The package features centralized memory and I/O, which were pushed out to the edges of the package with Granite Rapids. You can read more about the design of the chip in our <a href="https://www.tomshardware.com/pc-components/cpus/intel-xeon-7-diamond-rapids-comes-with-up-to-256-p-cores-1-28-gb-of-last-level-cache-next-gen-18a-p-cpu-also-brings-avx-10-2-and-uses-ucie-s-instead-of-emib"><u>Diamond Rapids recap from Hot Chips on </u><u><em>Tom’s Hardware Premium</em></u></a><em>. </em></p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Nvidia says it’s ‘absurd’ to assume it predicted the agentic moment during Vera’s development ]]></title>
                                                                                                <dc:content><![CDATA[ <p>One criticism that has emerged of Nvidia’s Vera CPU, <a href="https://www.tomshardware.com/pc-components/cpus/amd-vp-and-chief-zen-architect-rebuke-claims-that-nvidia-and-arm-made-cpus-for-agentic-ai-we-didnt-magically-know-four-years-ago-that-agentic-was-going-to-be-a-thing">driven by AMD</a>, is that Nvidia – and, to a lesser degree, Arm with its <a href="https://www.tomshardware.com/pc-components/cpus/amd-vp-and-chief-zen-architect-rebuke-claims-that-nvidia-and-arm-made-cpus-for-agentic-ai-we-didnt-magically-know-four-years-ago-that-agentic-was-going-to-be-a-thing">AGI CPU</a> – couldn’t have developed an “agentic CPU,” as the chip is often characterized in Nvidia’s marketing materials. Ian Finder, the head of Nvidia’s data center CPU business, pushed back on that criticism in a recent conversation with <em>Tom’s Hardware Premium </em>at <a href="https://www.tomshardware.com/tag/sc26">SC26</a>,<em> </em>calling the notion that Nvidia somehow knew about the agentic wave four years ago when Vera development began "absurd."</p><p>“There's a conviction around the design philosophy of the part, right? Am I going to sit here and tell you that we have this magic crystal ball around the agentic moment? That would be absurd, right?” Finder said. “But what we do have as a company, and I think this has been proven out in our track record, is [that] we have design intuition.”</p><p>Instead of specifically designing Vera for agentic workloads, which would be impossible without a way to see several years into the future, Finder characterized the part by its codesign with Nvidia’s accelerators, something that the team has continued to work on since its first Grace data center CPU. Coming from the assumption that data centers and AI workloads are accelerated (and they largely are), Finder says the co-design approach on the CPU can start in a “conceptually simple space, like what kind of CPU do we need to keep the accelerator fed and happy?”</p><p>Although conceptually simple, Finder continually looped back to a particular word to describe aspects of Vera and, to a lesser extent, Grace. “Opinionated” is the word he used, which in the broader context of Vera’s design, shows up in two places. Grace, Finder says, introduced an opinionated memory system with Nvidia’s Scalable Coherency Fabric, while Vera extends that memory system while adding an opinionated microarchitecture, Nvidia’s Olympus core. </p><p>The picture Finder paints is one of Nvidia starting from scratch in this space, contrasting it with the “legacy constraints” of, presumably, Intel and AMD in particular, who have large pools of customers with specific needs. As we described in our <a href="https://www.tomshardware.com/pc-components/cpus/nvidia-spills-the-beans-on-vera-cpu-spec-benchmarks-revealed-olympus-architecture-detailed-and-more">original breakdown of the Olympus architecture</a> in Vera, Nvidia’s goal doesn’t seem to be to steal market share from AMD and Intel, but rather to grab market share in a CPU space that’s vastly accelerating on the back of AI agents. </p><p>This “clean sheet of paper” approach, and Nvidia’s codesign, led to some natural assumptions about what the CPU should do. “If we have silicon that needs really, really great single-thread performance, [and] we also need concurrency, then we need to get there with IPC instead of frequency. We need to make sure we have good loaded memory latency. We need to make sure we have a lot of bandwidth, and you start to design something on a clean sheet of paper that's not encumbered by dollar-per-core economics.”</p><p>Although Finder was clear that Vera’s design didn’t explicitly center around the agents we see today, there were early indications of the workload. “To say that four years ago, when we started putting pen to paper on Vera, [that] we absolutely saw the agentic moment, would [not] be... something I would claim, but we were starting to see the beginnings of tool calling and things like that. We certainly knew what we needed from a head node, and we knew the overall shape of this problem.”</p><p>As you can see in our <a href="https://www.tomshardware.com/pc-components/cpus/benchmarking-nvidias-vera-cpu-for-the-data-center-standalone-vera-tested-against-granite-rapids-ap-and-epyc-turin">Vera CPU benchmarking</a>, the chip doesn’t excel at every workload. And, from the data you can see in our recent <a href="https://www.tomshardware.com/pc-components/cpus/amd-zen-6-epyc-venice-deep-dive-and-benchmarking-dual-epyc-9996-with-1-024-threads-delivers-up-to-93-percent-performance-gain-gen-on-gen-9686f-outclasses-nvidias-vera">AMD EPYC Venice benchmarking</a>, the AMD camp has reclaimed some top spots in several benchmarks. Finder suggests that the superior memory bandwidth of Vera will play a much bigger role in the agentic future as these systems evolve to take better advantage of the compute available. </p><h2 id="taking-full-advantage-of-the-cpu">Taking full advantage of the CPU</h2><p>We need to rewind for just a moment. When the agentic wave hit earlier this year, some characterized the demand as a shift away from GPUs and toward CPUs. That is incorrect. Rather, as is becoming increasingly clear as these agentic workloads progress, these loops can’t take full advantage of the GPUs in a GPU-centric environment, with GPUs often oversubscribed on host nodes. The solution isn’t to get rid of GPUs, of course, nor is it to stop buying them. Rather, it’s <a href="https://www.tomshardware.com/pc-components/cpus/shifting-need-for-cpus-in-ai-workloads-drives-intensifying-shortages-price-hikes">to <em>also </em>buy enough CPUs</a> to properly leverage that compute. </p><p>A portion of the AI industry is focused on tackling the problem of fully utilizing the GPU, Nvidia included. GPUs shouldn’t sit idle for very long (if at all) with the demand today, and they don’t. The challenge today is spinning up these agent sandboxes quickly, within milliseconds, so that the CPU isn’t sitting idle, either. </p><p>Finder provided an example of using a coding agent. If you’ve used a coding agent before, you’re probably familiar with this process. After prompting, the model will make several small tool calls to read and locate files, passing back to the model in between. These are tiny calls, almost negligible, and the vast majority of the time is spent in the model on the GPU. Then, it may start compiling and testing the project, which is much more bound by the CPU. The GPU, Finder says, isn't sitting idle during that time. It’s being leveraged for another purpose. Finder suggests “that very same thing is going to happen on the CPU.”</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:56.28%;"><img id="eXLFBd3VeLbbGiDKVLt9D8" name="image3" alt="Nvidia Vera CPU" src="https://cdn.mos.cms.futurecdn.net/eXLFBd3VeLbbGiDKVLt9D8-1920-80.jpg" mos="" align="middle" fullscreen="" width="1999" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Nvidia)</span></figcaption></figure><p>“The idea that we're just going to say, ‘OK, the CPU can just chill out here, and we're just going to wait,’ that's not going to happen,” Finder said. “What's going to happen is we're going to pack multiple environments together, and we're going to subscribe the socket.”</p><p>This “very interesting computer science problem,” as Finder described it, is being investigated by AI startups such as Daytona and E2B to spin up and close down agentic sandboxes in milliseconds. And serving more agents will naturally put pressure on the memory system, which is an area where Vera has an edge. “We built a product that does really, really well without such optimizations, but as a design point that really will lean into them when the industry starts to catch up to this,” Finder said. </p><p>Vera is a fascinating chip, regardless of what you think of it and how it’s presented in marketing materials (from Nvidia itself or elsewhere). Although Grace was the starting point internally, Vera has ridden the wave of agentic to become a much bigger player in the world of CPUs, and Nvidia has already announced that it will <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-updates-data-center-roadmap-with-rosa-cpu-and-stacked-feynman-gpus-optical-nvlink-groq-lpus-with-nvfp4-and-nvlink-also-on-deck">follow up with its Rosa CPU</a> featuring the new Rigel microarchitecture. Vera is an opinionated CPU indeed, and time will tell if that opinion is correct. </p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/pc-components/cpus/nvidia-says-its-absurd-to-assume-it-predicted-the-agentic-moment-during-veras-development-product-intuition-helped-drive-cpu-design-superior-memory-bandwidth-will-play-critical-role-in-future-ai-agents-cpu-boss-claims</link>
                                                                            <description>
                            <![CDATA[ AMD has criticized Nvidia, and to a lesser extent Arm, for claims about creating a CPU designed for agentic AI. Nvidia is pushing back on that characterization, instead saying Vera was designed with “base principles, product intuition, and oftentimes simple allegorical laws of computer science” in a recent interview with Tom’s Hardware. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">oHGBQrBrddqJQiZsazTPg9</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/MX9ZgfJcjpjFTu4aNfmAhP-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Wed, 18 Nov 2026 15:00:00 +0000</pubDate>                                                                                                                                <updated>Wed, 18 Nov 2026 15:49:42 +0000</updated>
                                                                                                                                            <category><![CDATA[CPUs]]></category>
                                                    <category><![CDATA[PC Components]]></category>
                                                                                                                    <dc:creator><![CDATA[ Jake Roach ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/h6PRM8bTimCTnNfoAYfjAi-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Jake Roach has been bending pins and busting solder joints since the mid-2000s. From trying to run scratched CDs of &lt;em&gt;Delta Force &lt;/em&gt;and &lt;em&gt;Unreal Tournament &lt;/em&gt;to spitting out virtual machines on a Threadripper, Jake has been on the hunt for the latest hardware and highest performance for decades. That eventually spun up a career, with Jake serving as Lead Reporter at Digital Trends, as well as contributing to outlets like XDA, PC Invasion, Business Insider, and WIRED. At Tom’s Hardware, Jake is focused on consumer and workstation CPUs. Outside working hours, you’ll find him knee-deep in the latest roguelite taking over Steam, spending way too much money on &lt;em&gt;Magic: The Gathering, &lt;/em&gt;or forcing his lazy corgi onto walks.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>false</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/MX9ZgfJcjpjFTu4aNfmAhP-1920-80.jpg">
                                                            <media:credit><![CDATA[Tom&amp;#39;s Hardware]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Nvidia&#039;s Vera CPU]]></media:description>                                                            <media:text><![CDATA[Nvidia&#039;s Vera CPU]]></media:text>
                                <media:title type="plain"><![CDATA[Nvidia&#039;s Vera CPU]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/MX9ZgfJcjpjFTu4aNfmAhP-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>One criticism that has emerged of Nvidia’s Vera CPU, <a href="https://www.tomshardware.com/pc-components/cpus/amd-vp-and-chief-zen-architect-rebuke-claims-that-nvidia-and-arm-made-cpus-for-agentic-ai-we-didnt-magically-know-four-years-ago-that-agentic-was-going-to-be-a-thing">driven by AMD</a>, is that Nvidia – and, to a lesser degree, Arm with its <a href="https://www.tomshardware.com/pc-components/cpus/amd-vp-and-chief-zen-architect-rebuke-claims-that-nvidia-and-arm-made-cpus-for-agentic-ai-we-didnt-magically-know-four-years-ago-that-agentic-was-going-to-be-a-thing">AGI CPU</a> – couldn’t have developed an “agentic CPU,” as the chip is often characterized in Nvidia’s marketing materials. Ian Finder, the head of Nvidia’s data center CPU business, pushed back on that criticism in a recent conversation with <em>Tom’s Hardware Premium </em>at <a href="https://www.tomshardware.com/tag/sc26">SC26</a>,<em> </em>calling the notion that Nvidia somehow knew about the agentic wave four years ago when Vera development began "absurd."</p><p>“There's a conviction around the design philosophy of the part, right? Am I going to sit here and tell you that we have this magic crystal ball around the agentic moment? That would be absurd, right?” Finder said. “But what we do have as a company, and I think this has been proven out in our track record, is [that] we have design intuition.”</p><p>Instead of specifically designing Vera for agentic workloads, which would be impossible without a way to see several years into the future, Finder characterized the part by its codesign with Nvidia’s accelerators, something that the team has continued to work on since its first Grace data center CPU. Coming from the assumption that data centers and AI workloads are accelerated (and they largely are), Finder says the co-design approach on the CPU can start in a “conceptually simple space, like what kind of CPU do we need to keep the accelerator fed and happy?”</p><p>Although conceptually simple, Finder continually looped back to a particular word to describe aspects of Vera and, to a lesser extent, Grace. “Opinionated” is the word he used, which in the broader context of Vera’s design, shows up in two places. Grace, Finder says, introduced an opinionated memory system with Nvidia’s Scalable Coherency Fabric, while Vera extends that memory system while adding an opinionated microarchitecture, Nvidia’s Olympus core. </p><p>The picture Finder paints is one of Nvidia starting from scratch in this space, contrasting it with the “legacy constraints” of, presumably, Intel and AMD in particular, who have large pools of customers with specific needs. As we described in our <a href="https://www.tomshardware.com/pc-components/cpus/nvidia-spills-the-beans-on-vera-cpu-spec-benchmarks-revealed-olympus-architecture-detailed-and-more">original breakdown of the Olympus architecture</a> in Vera, Nvidia’s goal doesn’t seem to be to steal market share from AMD and Intel, but rather to grab market share in a CPU space that’s vastly accelerating on the back of AI agents. </p><p>This “clean sheet of paper” approach, and Nvidia’s codesign, led to some natural assumptions about what the CPU should do. “If we have silicon that needs really, really great single-thread performance, [and] we also need concurrency, then we need to get there with IPC instead of frequency. We need to make sure we have good loaded memory latency. We need to make sure we have a lot of bandwidth, and you start to design something on a clean sheet of paper that's not encumbered by dollar-per-core economics.”</p><p>Although Finder was clear that Vera’s design didn’t explicitly center around the agents we see today, there were early indications of the workload. “To say that four years ago, when we started putting pen to paper on Vera, [that] we absolutely saw the agentic moment, would [not] be... something I would claim, but we were starting to see the beginnings of tool calling and things like that. We certainly knew what we needed from a head node, and we knew the overall shape of this problem.”</p><p>As you can see in our <a href="https://www.tomshardware.com/pc-components/cpus/benchmarking-nvidias-vera-cpu-for-the-data-center-standalone-vera-tested-against-granite-rapids-ap-and-epyc-turin">Vera CPU benchmarking</a>, the chip doesn’t excel at every workload. And, from the data you can see in our recent <a href="https://www.tomshardware.com/pc-components/cpus/amd-zen-6-epyc-venice-deep-dive-and-benchmarking-dual-epyc-9996-with-1-024-threads-delivers-up-to-93-percent-performance-gain-gen-on-gen-9686f-outclasses-nvidias-vera">AMD EPYC Venice benchmarking</a>, the AMD camp has reclaimed some top spots in several benchmarks. Finder suggests that the superior memory bandwidth of Vera will play a much bigger role in the agentic future as these systems evolve to take better advantage of the compute available. </p><h2 id="taking-full-advantage-of-the-cpu">Taking full advantage of the CPU</h2><p>We need to rewind for just a moment. When the agentic wave hit earlier this year, some characterized the demand as a shift away from GPUs and toward CPUs. That is incorrect. Rather, as is becoming increasingly clear as these agentic workloads progress, these loops can’t take full advantage of the GPUs in a GPU-centric environment, with GPUs often oversubscribed on host nodes. The solution isn’t to get rid of GPUs, of course, nor is it to stop buying them. Rather, it’s <a href="https://www.tomshardware.com/pc-components/cpus/shifting-need-for-cpus-in-ai-workloads-drives-intensifying-shortages-price-hikes">to <em>also </em>buy enough CPUs</a> to properly leverage that compute. </p><p>A portion of the AI industry is focused on tackling the problem of fully utilizing the GPU, Nvidia included. GPUs shouldn’t sit idle for very long (if at all) with the demand today, and they don’t. The challenge today is spinning up these agent sandboxes quickly, within milliseconds, so that the CPU isn’t sitting idle, either. </p><p>Finder provided an example of using a coding agent. If you’ve used a coding agent before, you’re probably familiar with this process. After prompting, the model will make several small tool calls to read and locate files, passing back to the model in between. These are tiny calls, almost negligible, and the vast majority of the time is spent in the model on the GPU. Then, it may start compiling and testing the project, which is much more bound by the CPU. The GPU, Finder says, isn't sitting idle during that time. It’s being leveraged for another purpose. Finder suggests “that very same thing is going to happen on the CPU.”</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:56.28%;"><img id="eXLFBd3VeLbbGiDKVLt9D8" name="image3" alt="Nvidia Vera CPU" src="https://cdn.mos.cms.futurecdn.net/eXLFBd3VeLbbGiDKVLt9D8-1920-80.jpg" mos="" align="middle" fullscreen="" width="1999" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Nvidia)</span></figcaption></figure><p>“The idea that we're just going to say, ‘OK, the CPU can just chill out here, and we're just going to wait,’ that's not going to happen,” Finder said. “What's going to happen is we're going to pack multiple environments together, and we're going to subscribe the socket.”</p><p>This “very interesting computer science problem,” as Finder described it, is being investigated by AI startups such as Daytona and E2B to spin up and close down agentic sandboxes in milliseconds. And serving more agents will naturally put pressure on the memory system, which is an area where Vera has an edge. “We built a product that does really, really well without such optimizations, but as a design point that really will lean into them when the industry starts to catch up to this,” Finder said. </p><p>Vera is a fascinating chip, regardless of what you think of it and how it’s presented in marketing materials (from Nvidia itself or elsewhere). Although Grace was the starting point internally, Vera has ridden the wave of agentic to become a much bigger player in the world of CPUs, and Nvidia has already announced that it will <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-updates-data-center-roadmap-with-rosa-cpu-and-stacked-feynman-gpus-optical-nvlink-groq-lpus-with-nvfp4-and-nvlink-also-on-deck">follow up with its Rosa CPU</a> featuring the new Rigel microarchitecture. Vera is an opinionated CPU indeed, and time will tell if that opinion is correct. </p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Nvidia claims its Vera CPU can gain up to 39% performance boost via software optimization at SC26 — architectural fine-tuning is needed to fully unleash custom Olympus cores ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Nvidia's Vera CPU, designed as both a standalone data-center part and a host for Nvidia's Vera Rubin NVL72 servers, offers <a href="https://www.tomshardware.com/pc-components/cpus/benchmarking-nvidias-vera-cpu-for-the-data-center-standalone-vera-tested-against-granite-rapids-ap-and-epyc-turin">impressive performance</a> thanks to substantially higher execution resources per core than its predecessor, Grace. However, Nvidia admits at <a href="https://www.tomshardware.com/tag/sc26">Supercomputing 2026 (SC26) </a>that existing software cannot always take full advantage of Vera's resources without adaptation for the architecture.</p><p>The company shared that architecture-specific software optimizations can boost Vera's performance by nearly 39% in certain computationally intensive workloads. But despite offering unbeatable single-thread performance, the CPU is not positioned for general-purpose workloads.</p><div  class="fancy-box"><div class="fancy_box-title">Get the Tom's Hardware App</div><div class="fancy_box_body"><figure class="van-image-figure "  ><div class='image-full-width-wrapper'><div class='image-widthsetter' ><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="wy2jRCvahmJNbu4bh6m86M" name="image (9)" caption="" alt="Tom's Hardware App" src="https://cdn.mos.cms.futurecdn.net/wy2jRCvahmJNbu4bh6m86M-1920-80.png" mos="" link="" align="" fullscreen="" width="" height="" attribution="" endorsement="" class="pinterest-pin-exclude" /></p></div></div><figcaption itemprop="caption description" class=""><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p class="fancy-box__body-text">Get the full <em>Tom's Hardware</em> experience on your device with our brand-new app. Bookmark articles, use our smart search feature, personalize your feed, and get push notifications so you never miss a beat.</p><p class="fancy-box__body-text"><ul><li><a data-analytics-id="inline-link" href="https://apps.apple.com/ua/app/toms-hardware-pc-tech-news/id6785006008" target="_blank"><strong>Download now on Apple App Store</strong></a></li><li><a data-analytics-id="inline-link" href="https://play.google.com/store/apps/details?id=com.futureplc.peregrine.tomshardware" target="_blank"><strong>Download now on the Google Play Store</strong></a></li></ul></p></div></div><p>Unlike Grace, which relies on off-the-shelf Neoverse V2 CPU cores developed by Arm, Vera uses Nvidia's proprietary Olympus cores specifically designed to maximize per-core performance as well as to ensure high aggregate throughput. To that end, Nvidia equipped Olympus with a wide front end that features a 10-wide decoder, a neural branch predictor, a deep out-of-order execution engine, six vector execution units, four load pipelines, two store pipelines, all backed with large private caches and advanced prefetch mechanisms. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:65.82%;"><img id="tNinGuat4fBcaghWKj5mcP" name="Screenshot 2026-11-17 at 18.21.12" alt="Nvidia" src="https://cdn.mos.cms.futurecdn.net/tNinGuat4fBcaghWKj5mcP-1920-80.png" mos="" align="middle" fullscreen="" width="2560" height="1685" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Nvidia)</span></figcaption></figure><p>In fact, Nvidia deliberately opted for 88 high-performance general-purpose cores with support for its proprietary spatial multithreading technology (176 threads in total) rather than maximizing the number of cores per socket, as the maximum possible single-thread performance was one of the key design goals for Vera.</p><p>"CPU design is really just a game of trade-offs," said Hannah Coutand, who leads Vera CPU marketing at Nvidia, in an interview with <em>Tom's Hardware Premium</em>. "You could maximize for a lot of cores, but literally when you maximize the die for lots of cores, you do not actually have space to maximize for IPC, like instructions per clock."</p><p>Although Nvidia primarily positions Vera as a CPU for agentic AI and other AI-related workloads, the company originally conceived its Olympus microarchitecture with data processing applications in mind. As it turned out, the key attributes of Vera — high per-core performance, high memory bandwidth, and low memory access latency — are precisely what agentic AI workloads need as well, according to Coutand.</p><p>"We are very clear in our strategy with Vera and our positioning with Vera that it is not a general-purpose CPU," Coutand said. " If you want a web server for your cloud data center, there are probably other choices that are better suited for that than Vera. [Vera is made for] workloads […] that require that high per-core speed, high memory bandwidth, low latency. […] It has great fungibility for lots of workloads as long as those workloads value the things that Vera is good for."</p><h2 id="software-optimizations-make-good-things-better">Software optimizations make good things better</h2><p>To assess the computational capabilities of Nvidia's custom Olympus microarchitecture and its advantages over Neoverse V2, Nvidia's researchers used a double-precision general matrix multiplication (DGEMM) workload based on an assembly kernel from OpenBLAS, an open-source linear algebra library. While OpenBLAS features a kernel tailored for the Arm Neoverse V2 cores that power Nvidia's Grace processor, it lacks one specifically designed for Olympus. </p><p>To examine the efficiency of the two microarchitectures, the researchers used a standalone benchmark called miniKernelGEMM with a 128 × 128 matrix multiplication problem that fits into the L1 cache of both processors.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:54.88%;"><img id="WzHRg3hA8ru96QppKsfFfP" name="Screenshot 2026-11-17 at 18.16.07" alt="Nvidia" src="https://cdn.mos.cms.futurecdn.net/WzHRg3hA8ru96QppKsfFfP-1920-80.png" mos="" align="middle" fullscreen="" width="2560" height="1405" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Nvidia)</span></figcaption></figure><p>The Grace CPU achieved 4.51 instructions per clock (IPC) and 13.03 FP64 operations per cycle with the Neoverse V2 kernel, whereas Vera executed 5.43 instructions per clock and 15.69 FP64 operations per cycle with the very same code, an improvement of approximately 20%. </p><p>However, when the researchers used a kernel specifically optimized for Olympus, Vera’s performance increased to 7.53 instructions per clock and 21.75 FP64 operations per cycle. As a result, Vera outperformed its predecessor by nearly 67% on a per-core, per-clock basis in this particular workload.</p><p>Grace's Neoverse V2 microarchitecture has four 128-bit SIMD pipelines that can perform a total of 16 FP64 operations per clock, whereas Nvidia's Olympus has six such pipelines that can execute up to 24 FP64 operations per cycle, a 50% improvement. </p><p>But where the Neoverse V2-optimized kernel uses 81.4% of Grace's peak FP64 throughput, it uses only about 66.5% of Vera's computational capabilities. With changes to matrix tile organization, register allocation, and instruction scheduling that better suit the Olympus microarchitecture, the researchers managed to utilize 90.6% of its peak FP64 throughput, according to Nvidia's tutorial published at SC26. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:41.68%;"><img id="YJxJdu5bzLRhUHQreJZEAP" name="Screenshot 2026-11-17 at 18.16.38" alt="Nvidia" src="https://cdn.mos.cms.futurecdn.net/YJxJdu5bzLRhUHQreJZEAP-1920-80.png" mos="" align="middle" fullscreen="" width="2560" height="1067" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Nvidia)</span></figcaption></figure><p>While software developed and optimized for existing Arm-based data center processors can still benefit from the capabilities of Nvidia's Olympus cores without any specific optimizations, the company's SC26 benchmark results demonstrate that at least some applications will need to be fine-tuned for the new microarchitecture to take full advantage of its execution resources. Of course, it remains to be seen how much performance can be gained through Olympus-specific optimizations <a href="https://www.tomshardware.com/pc-components/cpus/benchmarking-nvidias-vera-cpu-for-the-data-center-standalone-vera-tested-against-granite-rapids-ap-and-epyc-turin" target="_blank">in workloads other than matrix multiplication</a>.</p><p>As Nvidia expands beyond CPUs that primarily serve as companions to its data center GPUs and chases the standalone server CPU market, the company will have to ensure that performance-critical software takes advantage of the architectural peculiarities of its processors, much like Intel and AMD have done for decades. </p><p>But the company is already deeply experienced with partner and ecosystem development for platforms like CUDA, so as Vera is more widely adopted, that experience could prove useful in spurring optimizations for these CPUs going forward. </p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/pc-components/cpus/nvidia-claims-its-vera-cpu-can-gain-up-to-39-percent-performance-boost-via-software-optimization-at-sc26-architectural-fine-tuning-is-needed-to-fully-unleash-custom-olympus-cores</link>
                                                                            <description>
                            <![CDATA[ Nvidia's Vera CPU, powered by its custom Olympus Arm cores, already offers strong performance with unoptimized code, but at SC26, the company demonstrated that further software optimizations can improve performance by nearly 39% with certain workloads. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">Nkrnqxh8tCpVAYsSsC8A68</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/HLdVAsCHvfXFb5ZjVFtz2b-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Tue, 17 Nov 2026 17:30:00 +0000</pubDate>                                                                                                                                                                                                                                <category><![CDATA[CPUs]]></category>
                                                    <category><![CDATA[PC Components]]></category>
                                                                                                <author><![CDATA[ ashilov@gmail.com (Anton Shilov) ]]></author>                    <dc:creator><![CDATA[ Anton Shilov ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/uMZ5kNphxA2Ut6whdLaSQV-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Anton Shilov has been in the PC industry since 1990s playing games, building PCs, and writing stories about pretty much everything that relates to PCs, Macs, smartphones, tablets, and even fab equipment. Over his career, he has worked at a variety of high-ranking websites, including AnandTech, EE Times, TechRadar, X-bit Labs, and now Tom&#039;s Hardware. He is also a regular features contributor to Tom&#039;s Hardware Premium, writing about the latest developments in the semiconductor industry and related tech news and roadmaps. When Anton is not reading or writing about something high-tech, he is probably watching a good movie, playing a video game, or spending time with his family.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>false</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/HLdVAsCHvfXFb5ZjVFtz2b-1920-80.jpg">
                                                            <media:credit><![CDATA[Tom&#039;s Hardware]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[SOCAMM 2 modules spotted on the Vera CPU tray ]]></media:description>                                                            <media:text><![CDATA[SOCAMM 2 modules spotted on the Vera CPU tray ]]></media:text>
                                <media:title type="plain"><![CDATA[SOCAMM 2 modules spotted on the Vera CPU tray ]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/HLdVAsCHvfXFb5ZjVFtz2b-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Nvidia's Vera CPU, designed as both a standalone data-center part and a host for Nvidia's Vera Rubin NVL72 servers, offers <a href="https://www.tomshardware.com/pc-components/cpus/benchmarking-nvidias-vera-cpu-for-the-data-center-standalone-vera-tested-against-granite-rapids-ap-and-epyc-turin">impressive performance</a> thanks to substantially higher execution resources per core than its predecessor, Grace. However, Nvidia admits at <a href="https://www.tomshardware.com/tag/sc26">Supercomputing 2026 (SC26) </a>that existing software cannot always take full advantage of Vera's resources without adaptation for the architecture.</p><p>The company shared that architecture-specific software optimizations can boost Vera's performance by nearly 39% in certain computationally intensive workloads. But despite offering unbeatable single-thread performance, the CPU is not positioned for general-purpose workloads.</p><div  class="fancy-box"><div class="fancy_box-title">Get the Tom's Hardware App</div><div class="fancy_box_body"><figure class="van-image-figure "  ><div class='image-full-width-wrapper'><div class='image-widthsetter' ><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="wy2jRCvahmJNbu4bh6m86M" name="image (9)" caption="" alt="Tom's Hardware App" src="https://cdn.mos.cms.futurecdn.net/wy2jRCvahmJNbu4bh6m86M-1920-80.png" mos="" link="" align="" fullscreen="" width="" height="" attribution="" endorsement="" class="pinterest-pin-exclude" /></p></div></div><figcaption itemprop="caption description" class=""><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p class="fancy-box__body-text">Get the full <em>Tom's Hardware</em> experience on your device with our brand-new app. Bookmark articles, use our smart search feature, personalize your feed, and get push notifications so you never miss a beat.</p><p class="fancy-box__body-text"><ul><li><a data-analytics-id="inline-link" href="https://apps.apple.com/ua/app/toms-hardware-pc-tech-news/id6785006008" target="_blank"><strong>Download now on Apple App Store</strong></a></li><li><a data-analytics-id="inline-link" href="https://play.google.com/store/apps/details?id=com.futureplc.peregrine.tomshardware" target="_blank"><strong>Download now on the Google Play Store</strong></a></li></ul></p></div></div><p>Unlike Grace, which relies on off-the-shelf Neoverse V2 CPU cores developed by Arm, Vera uses Nvidia's proprietary Olympus cores specifically designed to maximize per-core performance as well as to ensure high aggregate throughput. To that end, Nvidia equipped Olympus with a wide front end that features a 10-wide decoder, a neural branch predictor, a deep out-of-order execution engine, six vector execution units, four load pipelines, two store pipelines, all backed with large private caches and advanced prefetch mechanisms. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:65.82%;"><img id="tNinGuat4fBcaghWKj5mcP" name="Screenshot 2026-11-17 at 18.21.12" alt="Nvidia" src="https://cdn.mos.cms.futurecdn.net/tNinGuat4fBcaghWKj5mcP-1920-80.png" mos="" align="middle" fullscreen="" width="2560" height="1685" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Nvidia)</span></figcaption></figure><p>In fact, Nvidia deliberately opted for 88 high-performance general-purpose cores with support for its proprietary spatial multithreading technology (176 threads in total) rather than maximizing the number of cores per socket, as the maximum possible single-thread performance was one of the key design goals for Vera.</p><p>"CPU design is really just a game of trade-offs," said Hannah Coutand, who leads Vera CPU marketing at Nvidia, in an interview with <em>Tom's Hardware Premium</em>. "You could maximize for a lot of cores, but literally when you maximize the die for lots of cores, you do not actually have space to maximize for IPC, like instructions per clock."</p><p>Although Nvidia primarily positions Vera as a CPU for agentic AI and other AI-related workloads, the company originally conceived its Olympus microarchitecture with data processing applications in mind. As it turned out, the key attributes of Vera — high per-core performance, high memory bandwidth, and low memory access latency — are precisely what agentic AI workloads need as well, according to Coutand.</p><p>"We are very clear in our strategy with Vera and our positioning with Vera that it is not a general-purpose CPU," Coutand said. " If you want a web server for your cloud data center, there are probably other choices that are better suited for that than Vera. [Vera is made for] workloads […] that require that high per-core speed, high memory bandwidth, low latency. […] It has great fungibility for lots of workloads as long as those workloads value the things that Vera is good for."</p><h2 id="software-optimizations-make-good-things-better">Software optimizations make good things better</h2><p>To assess the computational capabilities of Nvidia's custom Olympus microarchitecture and its advantages over Neoverse V2, Nvidia's researchers used a double-precision general matrix multiplication (DGEMM) workload based on an assembly kernel from OpenBLAS, an open-source linear algebra library. While OpenBLAS features a kernel tailored for the Arm Neoverse V2 cores that power Nvidia's Grace processor, it lacks one specifically designed for Olympus. </p><p>To examine the efficiency of the two microarchitectures, the researchers used a standalone benchmark called miniKernelGEMM with a 128 × 128 matrix multiplication problem that fits into the L1 cache of both processors.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:54.88%;"><img id="WzHRg3hA8ru96QppKsfFfP" name="Screenshot 2026-11-17 at 18.16.07" alt="Nvidia" src="https://cdn.mos.cms.futurecdn.net/WzHRg3hA8ru96QppKsfFfP-1920-80.png" mos="" align="middle" fullscreen="" width="2560" height="1405" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Nvidia)</span></figcaption></figure><p>The Grace CPU achieved 4.51 instructions per clock (IPC) and 13.03 FP64 operations per cycle with the Neoverse V2 kernel, whereas Vera executed 5.43 instructions per clock and 15.69 FP64 operations per cycle with the very same code, an improvement of approximately 20%. </p><p>However, when the researchers used a kernel specifically optimized for Olympus, Vera’s performance increased to 7.53 instructions per clock and 21.75 FP64 operations per cycle. As a result, Vera outperformed its predecessor by nearly 67% on a per-core, per-clock basis in this particular workload.</p><p>Grace's Neoverse V2 microarchitecture has four 128-bit SIMD pipelines that can perform a total of 16 FP64 operations per clock, whereas Nvidia's Olympus has six such pipelines that can execute up to 24 FP64 operations per cycle, a 50% improvement. </p><p>But where the Neoverse V2-optimized kernel uses 81.4% of Grace's peak FP64 throughput, it uses only about 66.5% of Vera's computational capabilities. With changes to matrix tile organization, register allocation, and instruction scheduling that better suit the Olympus microarchitecture, the researchers managed to utilize 90.6% of its peak FP64 throughput, according to Nvidia's tutorial published at SC26. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:41.68%;"><img id="YJxJdu5bzLRhUHQreJZEAP" name="Screenshot 2026-11-17 at 18.16.38" alt="Nvidia" src="https://cdn.mos.cms.futurecdn.net/YJxJdu5bzLRhUHQreJZEAP-1920-80.png" mos="" align="middle" fullscreen="" width="2560" height="1067" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Nvidia)</span></figcaption></figure><p>While software developed and optimized for existing Arm-based data center processors can still benefit from the capabilities of Nvidia's Olympus cores without any specific optimizations, the company's SC26 benchmark results demonstrate that at least some applications will need to be fine-tuned for the new microarchitecture to take full advantage of its execution resources. Of course, it remains to be seen how much performance can be gained through Olympus-specific optimizations <a href="https://www.tomshardware.com/pc-components/cpus/benchmarking-nvidias-vera-cpu-for-the-data-center-standalone-vera-tested-against-granite-rapids-ap-and-epyc-turin" target="_blank">in workloads other than matrix multiplication</a>.</p><p>As Nvidia expands beyond CPUs that primarily serve as companions to its data center GPUs and chases the standalone server CPU market, the company will have to ensure that performance-critical software takes advantage of the architectural peculiarities of its processors, much like Intel and AMD have done for decades. </p><p>But the company is already deeply experienced with partner and ecosystem development for platforms like CUDA, so as Vera is more widely adopted, that experience could prove useful in spurring optimizations for these CPUs going forward. </p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ AMD attacks Nvidia's FP64 emulation at SC26, touts native double-precision on Instinct MI430X ]]></title>
                                                                                                <dc:content><![CDATA[ <p>As Supercomputing 2026 (SC26) opens in Chicago this week, AMD is indirectly needling Nvidia about the lack of native FP64 data type support in its data center GPU products. FP64, or double-precision floating-point math, is widely used in the high-performance scientific computing and simulation fields that are the focus of SC26. For those customers, <a href="https://www.amd.com/en/blogs/2026/exascale-was-the-beginning-amd-instinct--mi430x-gpus.html" target="_blank">AMD is trumpeting the value of its latest Instinct MI430X accelerator</a> for customers that still need native FP64, like national laboratories and supercomputing centers. </p><p>Over the past few years, Nvidia has reallocated die area on its data center GPUs to favor the lower-precision matrix math that boosts the operations needed for much of modern AI inference. That reallocation only makes business sense, as the total addressable market for AI compute has exploded while the demand for double-precision math in large supercomputing installations has remained flat by comparison. </p><p>To make up for the lack of native FP64 support, Nvidia has turned to techniques that allow it to emulate FP64 operations using methods like the Ozaki schemes. But AMD claims this kind of emulation is nowhere close to a replacement for native double-precision processing for a few reasons—hence continued development of specialized products like the MI430X. </p><p>For one, AMD notes that Ozaki emulation assumes that the workload in question resembles a dense general matrix multiply (GEMM). In its own application profiling, the company says "less than 10%" of HPC workloads perform dense GEMM operations, so this emulation is a good replacement for native compute. Dense linear algebra is the one common workload for which AMD says Ozaki emulation is a natural fit.</p><p>For workloads like sparse iterative linear algebra, stencils and hydrodynamics, high-order finite element analysis, Monte Carlo and particle transport, and classical molecular dynamics, AMD says developers still need strong FP64 vector performance for the best results.</p><p>Furthermore, even for workloads that map well onto GEMM operations, AMD says matrix sizes must be large in order to achieve the best performance, and that might also not be a given for every operation. The operands of those matrices also have to be sized well to the hardware, as incorrectly scaled operands can reduce performance. The FP64 vector math capabilities of the MI430X are insensitive to those operand sizing issues.</p><p>Beyond those stumbling blocks, AMD also notes that using Ozaki emulation methods can cause the memory footprint of an application to balloon, requiring up to three times more capacity than native FP64 would. </p><p>Since HBM capacity is directly tied to GPUs in modern supercomputing installations, fitting a workload to a machine that can only emulate FP64 operations could have cascading effects on the number of GPUs and compute nodes required, causing knock-on challenges for the floor area, power, and cooling required to support the system. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1920px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="UahiUVXoQLojHq4AZy3u7U" name="discovery-sc" alt="A render of the Discovery supercomputer" src="https://cdn.mos.cms.futurecdn.net/UahiUVXoQLojHq4AZy3u7U-1920-80.png" mos="" align="middle" fullscreen="" width="1920" height="1080" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="caption-text">The Oak Ridge National Laboratory Discovery supercomputer, powered by MI430X GPUs </span><span class="credit" itemprop="copyrightHolder">(Image credit: Oak Ridge National Laboratory)</span></figcaption></figure><p>AMD says the high 432GB of HBM4 memory capacity per MI430X node allows its customers to provision fewer GPUs and compute nodes in a system, saving on networking, power, and floor space. </p><p>In addition to its FP64 HPC performance, AMD positions an MI430X-powered supercomputer as an AI inference platform for developing workloads that utilize those models as part of the research process. But there are some key differences between those systems and the Helios rack-scale accelerator powered by <a href="https://www.tomshardware.com/pc-components/gpus/amd-takes-the-wraps-off-its-instinct-mi455x-ai-accelerator-cdna-5-and-helios-rack-scale-architecture-combine-to-take-the-fight-to-nvidia-in-the-data-center">MI455X GPUs</a> from the same CDNA 5 architectural family. </p><p>Each MI430X GPU offers 9.2 PFLOPS of theoretical FP4 compute performance for inference, but that should be understood in the context of the whopping 40.3 PFLOPS of FP4 compute from each AI-focused MI455X accelerator. </p><p>Beyond raw performance, the native scale-up domain size of four MI430Xs over AMD’s Infinity Fabric interconnect when coupled with a sixth-generation EPYC host CPU pales in comparison to the domain size of 72 MI455X accelerators in <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/amds-helios-mi455x-ai-platform-breaks-cover-initial-systems-use-ualink-over-ethernet-interconnects-amds-vera-rubin-rival-surfaces-but-the-downsides-of-ethernet-could-hamstring-performance">a Helios rack</a>. AMD says this interconnect is optimized more for latency-sensitive operations rather than a massive scale-up domain.</p><p>Of course, every customer and user of such a system will have different requirements, and institutional supercomputing installations usually aren’t trying to do double duty as cloud inference providers. </p><p>For scientists and engineers whose AI models of choice work well within the 1.72 TB of aggregate HBM4 capacity and 93.2 TB/s of memory bandwidth offered by four MI430X's working together across the 12.3 Tbps of scale-up bandwidth available within such a node, the inference performance of each node is likely more than up to the task of the kinds of AI applications that need to be run alongside double-precision compute workloads. </p><p>In any event, the MI430X can already count several supercomputing design wins in its corner. The upcoming Discovery supercomputer at Oak Ridge National Laboratory, the Alice Recoque installation in France, the LUMI-AI system at Finland’s CSC, and the Herder system in Germany will all take advantage of the native FP64 support offered by the MI430X. </p><p>AMD’s continued support for native FP64, even as Nvidia retreats from that capability in its products, will doubtless be a point of pride for its architects and engineers, even as those institutional supercomputers will be dwarfed by both AMD and Nvidia-powered AI inference installations the world over. The company's semi-custom silicon design chops and chiplet experience give it the flexibility needed to tailor products like the MI430X to these specialized markets instead of pursuing a one-size-fits-all approach. </p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/pc-components/gpus/amd-attacks-nvidias-fp64-emulation-at-sc26-touts-native-double-precision-on-instinct-mi430x-says-ozaki-schemes-fit-less-than-10-percent-of-hpc-workloads-and-triple-memory-use-and-emulated-fp64-is-a-poor-fit-for-many-hpc-workloads</link>
                                                                            <description>
                            <![CDATA[ AMD is touting the benefits of the native FP64 support provided by its MI430X accelerator at Supercomputing 2026. The company says that Ozaki FP64 emulation only covers the needs of a small portion of HPC workloads, making the MI430X a unique foundation for next-gen supercomputing installations. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">7uCV8qHDFePnZ8Jpdou9qA</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/aYqjYmHikYgdUmCHUJ3Wbk-1920-80.png" type="image/png" length="0"></enclosure>
                                                                        <pubDate>Tue, 17 Nov 2026 14:30:00 +0000</pubDate>                                                                                                                                <updated>Tue, 17 Nov 2026 16:47:47 +0000</updated>
                                                                                                                                            <category><![CDATA[GPUs]]></category>
                                                    <category><![CDATA[PC Components]]></category>
                                                                                                                    <dc:creator><![CDATA[ Jeffrey Kampman ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/8JCjGs5yVZds2YdKmzjUDE-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Jeff Kampman has been playing PC games ever since he learned how to fire up freeware CDs from the DOS command line. He started building his own PCs in the mid-aughts and later turned that passion into a career, working as a news and guides writer, reviewer, and ultimately Editor-in-Chief at The Tech Report, where he dove deep on CPUs and GPUs (and more) in pursuit of the smoothest gaming experiences around. Jeff later took on roles at Asus and Intel as a technical marketer before joining Tom&#039;s Hardware. As Senior Analyst, Graphics, Jeff covers everything from integrated graphics processors to discrete graphics cards to the massive data center GPU installations powering our AI future. Jeff is also a hobbyist photographer, Twitch streamer, espresso enthusiast, and runner.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>false</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/png" url="https://cdn.mos.cms.futurecdn.net/aYqjYmHikYgdUmCHUJ3Wbk-1920-80.png">
                                                            <media:credit><![CDATA[AMD]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Two views of the AMD MI430X accelerator]]></media:description>                                                            <media:text><![CDATA[Two views of the AMD MI430X accelerator]]></media:text>
                                <media:title type="plain"><![CDATA[Two views of the AMD MI430X accelerator]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/aYqjYmHikYgdUmCHUJ3Wbk-1920-80.png" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>As Supercomputing 2026 (SC26) opens in Chicago this week, AMD is indirectly needling Nvidia about the lack of native FP64 data type support in its data center GPU products. FP64, or double-precision floating-point math, is widely used in the high-performance scientific computing and simulation fields that are the focus of SC26. For those customers, <a href="https://www.amd.com/en/blogs/2026/exascale-was-the-beginning-amd-instinct--mi430x-gpus.html" target="_blank">AMD is trumpeting the value of its latest Instinct MI430X accelerator</a> for customers that still need native FP64, like national laboratories and supercomputing centers. </p><p>Over the past few years, Nvidia has reallocated die area on its data center GPUs to favor the lower-precision matrix math that boosts the operations needed for much of modern AI inference. That reallocation only makes business sense, as the total addressable market for AI compute has exploded while the demand for double-precision math in large supercomputing installations has remained flat by comparison. </p><p>To make up for the lack of native FP64 support, Nvidia has turned to techniques that allow it to emulate FP64 operations using methods like the Ozaki schemes. But AMD claims this kind of emulation is nowhere close to a replacement for native double-precision processing for a few reasons—hence continued development of specialized products like the MI430X. </p><p>For one, AMD notes that Ozaki emulation assumes that the workload in question resembles a dense general matrix multiply (GEMM). In its own application profiling, the company says "less than 10%" of HPC workloads perform dense GEMM operations, so this emulation is a good replacement for native compute. Dense linear algebra is the one common workload for which AMD says Ozaki emulation is a natural fit.</p><p>For workloads like sparse iterative linear algebra, stencils and hydrodynamics, high-order finite element analysis, Monte Carlo and particle transport, and classical molecular dynamics, AMD says developers still need strong FP64 vector performance for the best results.</p><p>Furthermore, even for workloads that map well onto GEMM operations, AMD says matrix sizes must be large in order to achieve the best performance, and that might also not be a given for every operation. The operands of those matrices also have to be sized well to the hardware, as incorrectly scaled operands can reduce performance. The FP64 vector math capabilities of the MI430X are insensitive to those operand sizing issues.</p><p>Beyond those stumbling blocks, AMD also notes that using Ozaki emulation methods can cause the memory footprint of an application to balloon, requiring up to three times more capacity than native FP64 would. </p><p>Since HBM capacity is directly tied to GPUs in modern supercomputing installations, fitting a workload to a machine that can only emulate FP64 operations could have cascading effects on the number of GPUs and compute nodes required, causing knock-on challenges for the floor area, power, and cooling required to support the system. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1920px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="UahiUVXoQLojHq4AZy3u7U" name="discovery-sc" alt="A render of the Discovery supercomputer" src="https://cdn.mos.cms.futurecdn.net/UahiUVXoQLojHq4AZy3u7U-1920-80.png" mos="" align="middle" fullscreen="" width="1920" height="1080" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="caption-text">The Oak Ridge National Laboratory Discovery supercomputer, powered by MI430X GPUs </span><span class="credit" itemprop="copyrightHolder">(Image credit: Oak Ridge National Laboratory)</span></figcaption></figure><p>AMD says the high 432GB of HBM4 memory capacity per MI430X node allows its customers to provision fewer GPUs and compute nodes in a system, saving on networking, power, and floor space. </p><p>In addition to its FP64 HPC performance, AMD positions an MI430X-powered supercomputer as an AI inference platform for developing workloads that utilize those models as part of the research process. But there are some key differences between those systems and the Helios rack-scale accelerator powered by <a href="https://www.tomshardware.com/pc-components/gpus/amd-takes-the-wraps-off-its-instinct-mi455x-ai-accelerator-cdna-5-and-helios-rack-scale-architecture-combine-to-take-the-fight-to-nvidia-in-the-data-center">MI455X GPUs</a> from the same CDNA 5 architectural family. </p><p>Each MI430X GPU offers 9.2 PFLOPS of theoretical FP4 compute performance for inference, but that should be understood in the context of the whopping 40.3 PFLOPS of FP4 compute from each AI-focused MI455X accelerator. </p><p>Beyond raw performance, the native scale-up domain size of four MI430Xs over AMD’s Infinity Fabric interconnect when coupled with a sixth-generation EPYC host CPU pales in comparison to the domain size of 72 MI455X accelerators in <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/amds-helios-mi455x-ai-platform-breaks-cover-initial-systems-use-ualink-over-ethernet-interconnects-amds-vera-rubin-rival-surfaces-but-the-downsides-of-ethernet-could-hamstring-performance">a Helios rack</a>. AMD says this interconnect is optimized more for latency-sensitive operations rather than a massive scale-up domain.</p><p>Of course, every customer and user of such a system will have different requirements, and institutional supercomputing installations usually aren’t trying to do double duty as cloud inference providers. </p><p>For scientists and engineers whose AI models of choice work well within the 1.72 TB of aggregate HBM4 capacity and 93.2 TB/s of memory bandwidth offered by four MI430X's working together across the 12.3 Tbps of scale-up bandwidth available within such a node, the inference performance of each node is likely more than up to the task of the kinds of AI applications that need to be run alongside double-precision compute workloads. </p><p>In any event, the MI430X can already count several supercomputing design wins in its corner. The upcoming Discovery supercomputer at Oak Ridge National Laboratory, the Alice Recoque installation in France, the LUMI-AI system at Finland’s CSC, and the Herder system in Germany will all take advantage of the native FP64 support offered by the MI430X. </p><p>AMD’s continued support for native FP64, even as Nvidia retreats from that capability in its products, will doubtless be a point of pride for its architects and engineers, even as those institutional supercomputers will be dwarfed by both AMD and Nvidia-powered AI inference installations the world over. The company's semi-custom silicon design chops and chiplet experience give it the flexibility needed to tailor products like the MI430X to these specialized markets instead of pursuing a one-size-fits-all approach. </p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Samsung, SK hynix, and Micron form unlikely alliance to standardize robot charging as chipmakers push toward unmanned fabs ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Samsung Electronics, SK hynix, and Micron are reportedly working together to develop common charging standards for mobile robots used in semiconductor fabrication plants, according to a new report. Representatives from the three companies have reportedly formed an early-stage consultative body to lead standardization efforts, alongside major industry stakeholders such as the Semiconductor Equipment and Materials International (SEMI) industry association.</p><p>According to a <a href="https://en.sedaily.com/business/2026/11/08/memorys-big-three-to-set-robot-charging-standard"><u>report</u></a> by Seoul Economic Daily, the initiative aims to establish common technical specifications that would allow robots from different manufacturers to share charging equipment, reducing the cost and complexity of deploying large robotic fleets across semiconductor manufacturing facilities. The broader goal is to accelerate factory automation, particularly in older fabs, while laying the groundwork for fully autonomous semiconductor manufacturing.</p><p>So why are three of the world's biggest memory manufacturers, who are also competitors, coming together to establish standards for the robotics industry? What exactly does the collaboration involve? And what are the implications of something as seemingly minor as robot charging for the broader semiconductor and robotics industries?</p><h2 id="why-robot-charging-is-becoming-a-problem-for-semiconductor-fabs">Why robot charging is becoming a problem for semiconductor fabs</h2><p><a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/samsung-and-sk-hynix-warn-ai-driven-memory-shortages-could-last-until-2027-and-beyond-as-hbm-demand-explodes-customers-already-reserving-supply-years-ahead-while-the-wider-dram-market-begins-to-tighten" target="_blank">Soaring demand for semiconductor memory chips</a> used in AI infrastructure has driven manufacturers to automate as many production processes as possible to optimize manufacturing speed and capacity in existing facilities to increase output, while they <a href="https://www.tomshardware.com/tech-industry/semiconductors/south-korea-unveils-usd520-billion-investment-plan-with-samsung-and-sk-hynix-to-expand-memory-chip-dominance-plan-includes-four-new-fabs-and-hbm-facilities-amid-strong-government-support" target="_blank">build additional fabs</a>. Modern semiconductor fabs already have extensive automation systems. However, even these facilities still require human workers for certain operations, such as moving materials between floors and unpacking products, that existing automation systems cannot perform. The issue is even more pronounced in older fabs, which generally have fewer automated material-handling systems and still rely on workers to move wafer containers using manual carts.</p><p>One potential solution is to <a href="https://www.tomshardware.com/tech-industry/semiconductors/stmicro-to-deploy-humanoid-robots-to-its-legacy-fabs-in-europe-over-100-humanoid-robots-to-be-used-for-routine-and-physically-demanding-tasks-in-fight-for-efficiency" target="_blank">hand these remaining manual operations over to mobile robots</a>, allowing manufacturers to automate tasks without undertaking extensive renovations or installing complex fixed transport systems. These could include autonomous mobile robots (AMRs) that transport materials around the factory, mobile manipulators that combine robotic arms with moving platforms, and <a href="https://www.tomshardware.com/tech-industry/semiconductors/nvidia-products-could-be-made-using-humanoid-robots-for-the-first-time-ever-company-in-talks-with-foxconn-to-deploy-them-in-houston-factory-building-gb300-ai-servers" target="_blank">humanoid robots capable of performing more complicated handling tasks</a>.</p><p>The problem, however, is that these robots come in different builds and configurations, often with their own batteries and accompanying charging infrastructure. A semiconductor fab deploying multiple robot models from different manufacturers could consequently end up requiring separate charging stations for each model. In addition to increasing equipment and maintenance costs, these installations consume valuable factory floor space.</p><p>The proposed initiative seeks to address this problem by establishing common charging specifications that different robot manufacturers can adopt. Standardized charging infrastructure would allow multiple types of robots to use compatible charging stations, reducing the need for separate installations and making it easier for semiconductor manufacturers to expand their robotic fleets without adding operational inefficiencies.</p><h2 id="what-the-proposed-standard-covers-and-how-it-would-work">What the proposed standard covers and how it would work</h2><p>The alliance is considering standards covering battery specifications for different mobile robots, charging-station equipment and communication protocols, charging connectors, and battery management systems (BMS). Together, these specifications would establish common requirements for the physical and electronic interfaces involved in charging robots inside semiconductor fabs.</p><p>The initial objective is to achieve charging compatibility across different robot models. In addition to letting a fab deploy robots from several manufacturers without having to install a separate charging system for every model, this would also give manufacturers greater flexibility when purchasing new robots, since their choice of supplier would no longer be as heavily influenced by the charging infrastructure already installed in the facility.</p><p>A subsequent step would involve integrating robot charging data with the fab's manufacturing execution system (MES), the software that coordinates and monitors manufacturing operations. With access to information such as battery charge levels, robot availability, and production schedules, the system could coordinate when individual robots operate and when they return for charging. For example, robots could be scheduled to recharge during periods of lower demand or while other robots handle critical material transfers, minimizing interruptions to production. This integration would become increasingly important as manufacturers deploy larger fleets of robots across the hundreds of individual processes involved in semiconductor fabrication.</p><p>Interestingly, the three chipmakers leading the standardization efforts are primarily customers of industrial robotics manufacturers, not manufacturers of the robots themselves, except for Samsung, which has an established robotics business. The objective is to develop specifications that robotics manufacturers and charging-equipment suppliers can incorporate into their products.</p><p>On the other hand, the three companies are not independently creating specifications and dictating them to suppliers. The initiative would likely follow an industry standardization process in which robotics manufacturers, equipment suppliers, and other stakeholders contribute to developing and validating the technical requirements. For now, the project remains in its early stages, with participants still determining which components and interfaces should be standardized and consulting robotics and smart-factory specialists.</p><h2 id="why-the-chipmaking-rivals-are-joining-forces-now-and-the-broader-implications">Why the chipmaking rivals are joining forces now, and the broader implications</h2><p>According to DigiTimes, semiconductor industry participants first explored establishing common robotics standards roughly a decade ago, but discussions made little progress until recently. The renewed momentum is largely attributed to the ongoing semiconductor boom, which has coincided with rapid advancements in industrial robotics.</p><p>Unprecedented demand for AI-related memory products, along with <a href="https://www.tomshardware.com/tech-industry/samsungs-chip-division-expects-to-out-earn-its-entire-40-year-history-in-2026" target="_blank">record financial performance</a>, has given semiconductor manufacturers both the incentive and the financial resources to invest more aggressively in factory automation. While manufacturers are expanding production capacity to meet demand, new fabs take years to construct and bring online. As a result, they are simultaneously looking to squeeze out as much capacity as possible from existing facilities by increasing automation and improving productivity.</p><p>As for why Samsung, SK hynix, and Micron are collaborating despite their intense rivalry, the answer likely comes down to shared commercial interests. Incompatible robot charging infrastructure is an industry-wide problem, and all three companies stand to benefit from a common solution. The AI boom has already encouraged numerous partnerships across the semiconductor supply chain, including <a href="https://www.tomshardware.com/tech-industry/sandisk-and-sk-hynix-join-forces-to-standardize-high-bandwidth-flash-memory-a-nand-based-alternative-to-hbm-for-ai-gpus-move-could-enable-8-16x-higher-capacity-compared-to-dram" target="_blank">between companies that are considered fierce rivals</a>, as manufacturers look for ways to overcome common production and infrastructure constraints.</p><p>The involvement of these particular companies also gives the initiative considerable weight. Together,<a href="https://www.tomshardware.com/tech-industry/samsung-sk-hynix-and-micron-sued-over-alleged-dram-price-fixing-amid-record-memory-costs" target="_blank"> Samsung, SK hynix, and Micron control almost the entire global DRAM market</a>. They also operate extensive semiconductor manufacturing networks and represent a substantial potential market for fab-specific robotic equipment. A common set of specifications backed by all three would therefore provide robotics manufacturers with a strong commercial incentive to develop compatible products, potentially shortening development cycles and accelerating adoption.</p><p>Although the current collaboration focuses on a relatively small aspect of factory automation, it is a significant step for broader standardization across the semiconductor manufacturing and industrial robotics industries. Common charging infrastructure would make it easier to deploy and manage mixed robotic fleets, while future integration with production management systems could bring fabs closer to operating with minimal human intervention. This longer-term ambition of completely autonomous facilities is already taking shape. Both SK hynix and Samsung have <a href="https://www.tomshardware.com/tech-industry/samsung-plans-to-eliminate-humans-from-its-chip-fabs-by-2030-push-for-full-automation-continues-at-full-steam" target="_blank">outlined plans to develop an autonomous semiconductor fab by 2030</a>.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/semiconductors/samsung-sk-hynix-and-micron-form-unlikely-alliance-to-standardize-robot-charging-as-chipmakers-push-toward-unmanned-chip-fabs-standard-specs-would-allow-manufacturer-agnostic-charging-equipment</link>
                                                                            <description>
                            <![CDATA[ Samsung, SK hynix, and Micron are working to standardize charging infrastructure for mobile fab robot batteries to simplify automation and help fabs deploy larger mixed robotic fleets. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">BTeV2PsAyeBHbSJaShTbVC</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/nDTtYSUtcMdaq6MduKFbR5-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Tue, 17 Nov 2026 11:30:00 +0000</pubDate>                                                                                                                                                                                                                                <category><![CDATA[Semiconductors]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                    <category><![CDATA[Manufacturing]]></category>
                                                                                                                    <dc:creator><![CDATA[ Etiido Uko ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/BBrMt7jWtSo2Dc3iKoroyD-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Etiido Uko is a mechanical engineer and senior technical writer with over nine years of experience in documentation and reporting. He is deeply passionate about all things engineering and technology, and is an expert in gadgets, manufacturing, robotics, automotive, and aerospace. His work spans content creation for industry leaders across multiple sectors, including Autodesk, Siemens, Xometry, Telus, and Coca-Cola. When he is not writing or keeping up with the latest innovations, you can find him exploring lands unknown. Check out more of his work at etiidowrites.com.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/nDTtYSUtcMdaq6MduKFbR5-1920-80.jpg">
                                                            <media:credit><![CDATA[Micron]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Micron DRAM fab, Taichung]]></media:description>                                                            <media:text><![CDATA[Micron DRAM fab, Taichung]]></media:text>
                                <media:title type="plain"><![CDATA[Micron DRAM fab, Taichung]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/nDTtYSUtcMdaq6MduKFbR5-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Samsung Electronics, SK hynix, and Micron are reportedly working together to develop common charging standards for mobile robots used in semiconductor fabrication plants, according to a new report. Representatives from the three companies have reportedly formed an early-stage consultative body to lead standardization efforts, alongside major industry stakeholders such as the Semiconductor Equipment and Materials International (SEMI) industry association.</p><p>According to a <a href="https://en.sedaily.com/business/2026/11/08/memorys-big-three-to-set-robot-charging-standard"><u>report</u></a> by Seoul Economic Daily, the initiative aims to establish common technical specifications that would allow robots from different manufacturers to share charging equipment, reducing the cost and complexity of deploying large robotic fleets across semiconductor manufacturing facilities. The broader goal is to accelerate factory automation, particularly in older fabs, while laying the groundwork for fully autonomous semiconductor manufacturing.</p><p>So why are three of the world's biggest memory manufacturers, who are also competitors, coming together to establish standards for the robotics industry? What exactly does the collaboration involve? And what are the implications of something as seemingly minor as robot charging for the broader semiconductor and robotics industries?</p><h2 id="why-robot-charging-is-becoming-a-problem-for-semiconductor-fabs">Why robot charging is becoming a problem for semiconductor fabs</h2><p><a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/samsung-and-sk-hynix-warn-ai-driven-memory-shortages-could-last-until-2027-and-beyond-as-hbm-demand-explodes-customers-already-reserving-supply-years-ahead-while-the-wider-dram-market-begins-to-tighten" target="_blank">Soaring demand for semiconductor memory chips</a> used in AI infrastructure has driven manufacturers to automate as many production processes as possible to optimize manufacturing speed and capacity in existing facilities to increase output, while they <a href="https://www.tomshardware.com/tech-industry/semiconductors/south-korea-unveils-usd520-billion-investment-plan-with-samsung-and-sk-hynix-to-expand-memory-chip-dominance-plan-includes-four-new-fabs-and-hbm-facilities-amid-strong-government-support" target="_blank">build additional fabs</a>. Modern semiconductor fabs already have extensive automation systems. However, even these facilities still require human workers for certain operations, such as moving materials between floors and unpacking products, that existing automation systems cannot perform. The issue is even more pronounced in older fabs, which generally have fewer automated material-handling systems and still rely on workers to move wafer containers using manual carts.</p><p>One potential solution is to <a href="https://www.tomshardware.com/tech-industry/semiconductors/stmicro-to-deploy-humanoid-robots-to-its-legacy-fabs-in-europe-over-100-humanoid-robots-to-be-used-for-routine-and-physically-demanding-tasks-in-fight-for-efficiency" target="_blank">hand these remaining manual operations over to mobile robots</a>, allowing manufacturers to automate tasks without undertaking extensive renovations or installing complex fixed transport systems. These could include autonomous mobile robots (AMRs) that transport materials around the factory, mobile manipulators that combine robotic arms with moving platforms, and <a href="https://www.tomshardware.com/tech-industry/semiconductors/nvidia-products-could-be-made-using-humanoid-robots-for-the-first-time-ever-company-in-talks-with-foxconn-to-deploy-them-in-houston-factory-building-gb300-ai-servers" target="_blank">humanoid robots capable of performing more complicated handling tasks</a>.</p><p>The problem, however, is that these robots come in different builds and configurations, often with their own batteries and accompanying charging infrastructure. A semiconductor fab deploying multiple robot models from different manufacturers could consequently end up requiring separate charging stations for each model. In addition to increasing equipment and maintenance costs, these installations consume valuable factory floor space.</p><p>The proposed initiative seeks to address this problem by establishing common charging specifications that different robot manufacturers can adopt. Standardized charging infrastructure would allow multiple types of robots to use compatible charging stations, reducing the need for separate installations and making it easier for semiconductor manufacturers to expand their robotic fleets without adding operational inefficiencies.</p><h2 id="what-the-proposed-standard-covers-and-how-it-would-work">What the proposed standard covers and how it would work</h2><p>The alliance is considering standards covering battery specifications for different mobile robots, charging-station equipment and communication protocols, charging connectors, and battery management systems (BMS). Together, these specifications would establish common requirements for the physical and electronic interfaces involved in charging robots inside semiconductor fabs.</p><p>The initial objective is to achieve charging compatibility across different robot models. In addition to letting a fab deploy robots from several manufacturers without having to install a separate charging system for every model, this would also give manufacturers greater flexibility when purchasing new robots, since their choice of supplier would no longer be as heavily influenced by the charging infrastructure already installed in the facility.</p><p>A subsequent step would involve integrating robot charging data with the fab's manufacturing execution system (MES), the software that coordinates and monitors manufacturing operations. With access to information such as battery charge levels, robot availability, and production schedules, the system could coordinate when individual robots operate and when they return for charging. For example, robots could be scheduled to recharge during periods of lower demand or while other robots handle critical material transfers, minimizing interruptions to production. This integration would become increasingly important as manufacturers deploy larger fleets of robots across the hundreds of individual processes involved in semiconductor fabrication.</p><p>Interestingly, the three chipmakers leading the standardization efforts are primarily customers of industrial robotics manufacturers, not manufacturers of the robots themselves, except for Samsung, which has an established robotics business. The objective is to develop specifications that robotics manufacturers and charging-equipment suppliers can incorporate into their products.</p><p>On the other hand, the three companies are not independently creating specifications and dictating them to suppliers. The initiative would likely follow an industry standardization process in which robotics manufacturers, equipment suppliers, and other stakeholders contribute to developing and validating the technical requirements. For now, the project remains in its early stages, with participants still determining which components and interfaces should be standardized and consulting robotics and smart-factory specialists.</p><h2 id="why-the-chipmaking-rivals-are-joining-forces-now-and-the-broader-implications">Why the chipmaking rivals are joining forces now, and the broader implications</h2><p>According to DigiTimes, semiconductor industry participants first explored establishing common robotics standards roughly a decade ago, but discussions made little progress until recently. The renewed momentum is largely attributed to the ongoing semiconductor boom, which has coincided with rapid advancements in industrial robotics.</p><p>Unprecedented demand for AI-related memory products, along with <a href="https://www.tomshardware.com/tech-industry/samsungs-chip-division-expects-to-out-earn-its-entire-40-year-history-in-2026" target="_blank">record financial performance</a>, has given semiconductor manufacturers both the incentive and the financial resources to invest more aggressively in factory automation. While manufacturers are expanding production capacity to meet demand, new fabs take years to construct and bring online. As a result, they are simultaneously looking to squeeze out as much capacity as possible from existing facilities by increasing automation and improving productivity.</p><p>As for why Samsung, SK hynix, and Micron are collaborating despite their intense rivalry, the answer likely comes down to shared commercial interests. Incompatible robot charging infrastructure is an industry-wide problem, and all three companies stand to benefit from a common solution. The AI boom has already encouraged numerous partnerships across the semiconductor supply chain, including <a href="https://www.tomshardware.com/tech-industry/sandisk-and-sk-hynix-join-forces-to-standardize-high-bandwidth-flash-memory-a-nand-based-alternative-to-hbm-for-ai-gpus-move-could-enable-8-16x-higher-capacity-compared-to-dram" target="_blank">between companies that are considered fierce rivals</a>, as manufacturers look for ways to overcome common production and infrastructure constraints.</p><p>The involvement of these particular companies also gives the initiative considerable weight. Together,<a href="https://www.tomshardware.com/tech-industry/samsung-sk-hynix-and-micron-sued-over-alleged-dram-price-fixing-amid-record-memory-costs" target="_blank"> Samsung, SK hynix, and Micron control almost the entire global DRAM market</a>. They also operate extensive semiconductor manufacturing networks and represent a substantial potential market for fab-specific robotic equipment. A common set of specifications backed by all three would therefore provide robotics manufacturers with a strong commercial incentive to develop compatible products, potentially shortening development cycles and accelerating adoption.</p><p>Although the current collaboration focuses on a relatively small aspect of factory automation, it is a significant step for broader standardization across the semiconductor manufacturing and industrial robotics industries. Common charging infrastructure would make it easier to deploy and manage mixed robotic fleets, while future integration with production management systems could bring fabs closer to operating with minimal human intervention. This longer-term ambition of completely autonomous facilities is already taking shape. Both SK hynix and Samsung have <a href="https://www.tomshardware.com/tech-industry/samsung-plans-to-eliminate-humans-from-its-chip-fabs-by-2030-push-for-full-automation-continues-at-full-steam" target="_blank">outlined plans to develop an autonomous semiconductor fab by 2030</a>.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ China retains fastest supercomputer in the world crown as AI data centers soak up new hardware ]]></title>
                                                                                                <dc:content><![CDATA[ <p>China's LineShine remains the world's highest-performing supercomputer, according to the 68<sup>th</sup> edition of the Top500 list published at SC26 in Chicago on Monday. For the first time in decades, the supercomputer ranking saw exceptionally few additions despite unprecedented demand for high-performance computing hardware, as the vast amounts of compute resources deployed these days are destined for artificial intelligence (AI) data centers rather than traditional supercomputers.</p><h2 id="top-10-supercomputers-unchanged">Top 10 supercomputers unchanged</h2><p>China's <a href="https://www.tomshardware.com/tech-industry/supercomputers/chinas-lineshine-supercomputer-dethrones-us-el-capitan-secures-first-place-in-top-500-list-first-machine-in-the-rankings-to-sustain-more-than-2-exaflops-of-double-precision-performance-using-only-cpus">LineShine</a>, installed at the National Supercomputer Center in Shenzhen, offers sustained FP64 compute performance of 2.198 FP64 ExaFLOPS in the High Performance Linpack (HPL) benchmark at 42.22 MW. The machine relies on the homegrown LingKun platform with 304-core Armx9-based LX2 processors clocked at 1.55 GHz (13,789,440 cores in total) and the proprietary LingQi interconnect. Unlike the other four exascale-class supercomputers, LineShine does not use any compute accelerators and exclusively relies on CPUs, which is largely why it only gains 3.6X performance when moving from FP64 to mixed-precision data, which is lower than systems that use low-precision accelerators. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1094px;"><p class="vanilla-image-block" style="padding-top:35.47%;"><img id="PchizavCQC3mEngxrrs6bg" name="Screenshot 2026-11-17 at 01.57.42" alt="Top 500" src="https://cdn.mos.cms.futurecdn.net/PchizavCQC3mEngxrrs6bg-1920-80.png" mos="" align="middle" fullscreen="" width="1094" height="388" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Top 500)</span></figcaption></figure><p>America's most powerful El Capitan supercomputer at the Lawrence Livermore National Laboratory relies on AMD's Instinct MI300A accelerated processing units (APUs), and the HPE Cray EX255a platform delivers 1.809 FP64 ExaFLOPS at 29.69 MW, which makes it significantly more power-efficient than its rival from China. The Frontier supercomputer at the Oak Ridge National Laboratory, based on AMD's 3<sup>rd</sup> Generation EPYC processors and Instinct MI250X compute GPUs, achieves 1.353 FP64 ExaFLOPS at 24.61 MW. Meanwhile, Aurora at the Argonne National Laboratory, which uses Intel's Xeon CPU Max 9470 processors and Data Center GPU Max accelerators, offers 1.012 FP64 ExaFLOPS at 38.70 MW.  </p><p>All three American exascale supercomputers use HPE’s Cray EX platforms and Slingshot-11 interconnect. </p><p>Germany’s Jupiter Booster at the Jülich Supercomputing Center, which employs Nvidia’s GH200 Grace Hopper Superchips, the BullSequana XH3000 platform, and Nvidia’s quad-rail InfiniBand NDR200 interconnect, delivers 1.0 FP64 ExaFLOPS at 15.79 MW and remains the most energy-efficient exascale supercomputer on the list.</p><p>Remarkably, neither the 10 highest-performing supercomputers nor the 10 most energy-efficient machines have changed their positions since June. </p><h2 id="only-27-new-supercomputers-in-top-500">Only 27 new supercomputers in Top 500 </h2><p>Only 27 new supercomputers (5.4% of all machines on the list) made it to the November edition of the Top 500, the lowest number since the project was founded in 1993. The most powerful newcomer is South Korea's Hanggang at the KISTI Supercomputing Center, which occupies the 12<sup>th</sup> position with 342.6 FP64 PFLOPS. The machine uses Nvidia's GH200 Grace Hopper Superchips and HPE's Cray EX254n platform. Hanggang also ranks eighth in the HPCG benchmark with 4.29 HPCG PFLOPS.</p><p>The aggregate HPL performance of the world's 500 highest-performing supercomputers increased by a modest 4.4% over the past six months to 19.57 FP64 ExaFLOPS. Nonetheless, it now takes a machine with 2.857 PFLOPS of sustained FP64 performance to qualify for the TOP500, up 7.3% from 2.662 PFLOPS in June. The performance of the systems ranked 50<sup>th</sup>  and 100<sup>th</sup> rose from 49.85 PFLOPS to 55.81 PFLOPS (12%) and from 21.85 PFLOPS to 22.74 PFLOPS, respectively. Also, there are now 27 machines that deliver at least 100 PFLOPS of FP64 performance, up from 25 in June, whereas the number of supercomputers equipped with compute accelerators increased from 275 to 291 (58.2%).</p><h2 id="everything-goes-to-ai-data-centers">Everything goes to AI data centers </h2><p>The historically low number of new entrants to the Top 500 is particularly noteworthy given the unprecedented scale of investments in advanced computing infrastructure these days. According to estimates by Epoch AI cited by the Top 500 authors, the amount of AI compute capacity added globally every quarter has increased by 2.5X per year. Essentially, the industry's priorities have shifted from construction of machines capable of solving exceptionally complex scientific problems to deployment of vast amounts of compute capacity for commercial AI workloads. </p><p>In fact, the world's installed AI compute capacity already exceeds the equivalent of 15 million Nvidia H100 GPUs, according to Epoch AI. Furthermore, as of the fourth quarter of 2025, five major American hyperscalers accounted for an estimated 71% of the world's AI compute resources, much of which was deployed in massive clusters. While demand for advanced computing hardware is at an all-time high, it appears that the vast majority of new hardware never makes it to the world's most known supercomputer ranking.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/supercomputers/china-retains-fastest-supercomputer-in-the-world-crown-as-ai-data-centers-soak-up-new-hardware-top500-records-fewest-new-systems-in-33-years</link>
                                                                            <description>
                            <![CDATA[ The Top 500 list records the fewest new entrants in its history amid a massive shift in compute infrastructure investments towards AI factories. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">rE8V2bK6rGuc8WNfx6iMGZ</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/kiQWCdGYbYcrsLVKdCSDi3-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Mon, 16 Nov 2026 22:17:02 +0000</pubDate>                                                                                                                                                                                                                                <category><![CDATA[Supercomputers]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                <author><![CDATA[ ashilov@gmail.com (Anton Shilov) ]]></author>                    <dc:creator><![CDATA[ Anton Shilov ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/uMZ5kNphxA2Ut6whdLaSQV-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Anton Shilov has been in the PC industry since 1990s playing games, building PCs, and writing stories about pretty much everything that relates to PCs, Macs, smartphones, tablets, and even fab equipment. Over his career, he has worked at a variety of high-ranking websites, including AnandTech, EE Times, TechRadar, X-bit Labs, and now Tom&#039;s Hardware. He is also a regular features contributor to Tom&#039;s Hardware Premium, writing about the latest developments in the semiconductor industry and related tech news and roadmaps. When Anton is not reading or writing about something high-tech, he is probably watching a good movie, playing a video game, or spending time with his family.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>false</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/kiQWCdGYbYcrsLVKdCSDi3-1920-80.jpg">
                                                            <media:credit><![CDATA[Oak Ridge National Laboratory]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Lux supercomputer.]]></media:description>                                                            <media:text><![CDATA[Lux supercomputer.]]></media:text>
                                <media:title type="plain"><![CDATA[Lux supercomputer.]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/kiQWCdGYbYcrsLVKdCSDi3-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>China's LineShine remains the world's highest-performing supercomputer, according to the 68<sup>th</sup> edition of the Top500 list published at SC26 in Chicago on Monday. For the first time in decades, the supercomputer ranking saw exceptionally few additions despite unprecedented demand for high-performance computing hardware, as the vast amounts of compute resources deployed these days are destined for artificial intelligence (AI) data centers rather than traditional supercomputers.</p><h2 id="top-10-supercomputers-unchanged">Top 10 supercomputers unchanged</h2><p>China's <a href="https://www.tomshardware.com/tech-industry/supercomputers/chinas-lineshine-supercomputer-dethrones-us-el-capitan-secures-first-place-in-top-500-list-first-machine-in-the-rankings-to-sustain-more-than-2-exaflops-of-double-precision-performance-using-only-cpus">LineShine</a>, installed at the National Supercomputer Center in Shenzhen, offers sustained FP64 compute performance of 2.198 FP64 ExaFLOPS in the High Performance Linpack (HPL) benchmark at 42.22 MW. The machine relies on the homegrown LingKun platform with 304-core Armx9-based LX2 processors clocked at 1.55 GHz (13,789,440 cores in total) and the proprietary LingQi interconnect. Unlike the other four exascale-class supercomputers, LineShine does not use any compute accelerators and exclusively relies on CPUs, which is largely why it only gains 3.6X performance when moving from FP64 to mixed-precision data, which is lower than systems that use low-precision accelerators. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1094px;"><p class="vanilla-image-block" style="padding-top:35.47%;"><img id="PchizavCQC3mEngxrrs6bg" name="Screenshot 2026-11-17 at 01.57.42" alt="Top 500" src="https://cdn.mos.cms.futurecdn.net/PchizavCQC3mEngxrrs6bg-1920-80.png" mos="" align="middle" fullscreen="" width="1094" height="388" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Top 500)</span></figcaption></figure><p>America's most powerful El Capitan supercomputer at the Lawrence Livermore National Laboratory relies on AMD's Instinct MI300A accelerated processing units (APUs), and the HPE Cray EX255a platform delivers 1.809 FP64 ExaFLOPS at 29.69 MW, which makes it significantly more power-efficient than its rival from China. The Frontier supercomputer at the Oak Ridge National Laboratory, based on AMD's 3<sup>rd</sup> Generation EPYC processors and Instinct MI250X compute GPUs, achieves 1.353 FP64 ExaFLOPS at 24.61 MW. Meanwhile, Aurora at the Argonne National Laboratory, which uses Intel's Xeon CPU Max 9470 processors and Data Center GPU Max accelerators, offers 1.012 FP64 ExaFLOPS at 38.70 MW.  </p><p>All three American exascale supercomputers use HPE’s Cray EX platforms and Slingshot-11 interconnect. </p><p>Germany’s Jupiter Booster at the Jülich Supercomputing Center, which employs Nvidia’s GH200 Grace Hopper Superchips, the BullSequana XH3000 platform, and Nvidia’s quad-rail InfiniBand NDR200 interconnect, delivers 1.0 FP64 ExaFLOPS at 15.79 MW and remains the most energy-efficient exascale supercomputer on the list.</p><p>Remarkably, neither the 10 highest-performing supercomputers nor the 10 most energy-efficient machines have changed their positions since June. </p><h2 id="only-27-new-supercomputers-in-top-500">Only 27 new supercomputers in Top 500 </h2><p>Only 27 new supercomputers (5.4% of all machines on the list) made it to the November edition of the Top 500, the lowest number since the project was founded in 1993. The most powerful newcomer is South Korea's Hanggang at the KISTI Supercomputing Center, which occupies the 12<sup>th</sup> position with 342.6 FP64 PFLOPS. The machine uses Nvidia's GH200 Grace Hopper Superchips and HPE's Cray EX254n platform. Hanggang also ranks eighth in the HPCG benchmark with 4.29 HPCG PFLOPS.</p><p>The aggregate HPL performance of the world's 500 highest-performing supercomputers increased by a modest 4.4% over the past six months to 19.57 FP64 ExaFLOPS. Nonetheless, it now takes a machine with 2.857 PFLOPS of sustained FP64 performance to qualify for the TOP500, up 7.3% from 2.662 PFLOPS in June. The performance of the systems ranked 50<sup>th</sup>  and 100<sup>th</sup> rose from 49.85 PFLOPS to 55.81 PFLOPS (12%) and from 21.85 PFLOPS to 22.74 PFLOPS, respectively. Also, there are now 27 machines that deliver at least 100 PFLOPS of FP64 performance, up from 25 in June, whereas the number of supercomputers equipped with compute accelerators increased from 275 to 291 (58.2%).</p><h2 id="everything-goes-to-ai-data-centers">Everything goes to AI data centers </h2><p>The historically low number of new entrants to the Top 500 is particularly noteworthy given the unprecedented scale of investments in advanced computing infrastructure these days. According to estimates by Epoch AI cited by the Top 500 authors, the amount of AI compute capacity added globally every quarter has increased by 2.5X per year. Essentially, the industry's priorities have shifted from construction of machines capable of solving exceptionally complex scientific problems to deployment of vast amounts of compute capacity for commercial AI workloads. </p><p>In fact, the world's installed AI compute capacity already exceeds the equivalent of 15 million Nvidia H100 GPUs, according to Epoch AI. Furthermore, as of the fourth quarter of 2025, five major American hyperscalers accounted for an estimated 71% of the world's AI compute resources, much of which was deployed in massive clusters. While demand for advanced computing hardware is at an all-time high, it appears that the vast majority of new hardware never makes it to the world's most known supercomputer ranking.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Researchers ‘recognize that AI will never replace HPC,’ says AMD executive — high-precision workloads will ‘exist for a very long time’ even as AI usage expands ]]></title>
                                                                                                <dc:content><![CDATA[ <p>While AI continues to reshape industries across science and mathematics, traditional supercomputing has seemingly taken a back seat. However, AMD's Scott Tease told <em>Tom's Hardware Premium</em> that most HPC sites “recognize that AI will never replace HPC,” and suggests that building a heterogeneous compute stack is the best path forward to tackle bleeding-edge research, despite Nvidia pushing performance on lower-precision data types higher for AI workloads at the cost of native performance for the important double-precision FP64 format that’s critical for HPC workloads.</p><p>A heterogeneous compute stack brings together quantum, HPC, and AI under one roof, a fleet that AMD is helping pilot at Oak Ridge National Laboratory alongside IBM. Tease is AMD’s corporate vice president of HPC and sovereign AI, who joined the company earlier this year after more than a decade at Lenovo (and, before that, another 13 years at IBM). We spoke with Tease ahead of <a href="https://www.tomshardware.com/tag/sc26">SC26 </a>to get a read on the state of HPC, how it intersects with AI, and what’s actually driving the research community in a time when computer chips are driving headlines about international espionage plots. </p><h2 id="an-awful-lot-of-numbers">An awful lot of numbers</h2><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1920px;"><p class="vanilla-image-block" style="padding-top:60.00%;"><img id="3BFHbcRr5aki3FZCTBPAG5" name="instinct-hero" alt="AMD Instinct MI455X Accelerator" src="https://cdn.mos.cms.futurecdn.net/3BFHbcRr5aki3FZCTBPAG5-1920-80.png" mos="" align="middle" fullscreen="" width="1920" height="1152" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: AMD)</span></figcaption></figure><p>FP64 is important for HPC workloads because of its precision. FP64 is a hot-button issue, largely due to how Nvidia’s GPUs have progressed over the past two generations. Or rather, how they <em>haven't </em>progressed. <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-details-rubin-architectural-optimizations-for-inference-improvements-target-better-performance-and-efficiency-from-the-gpu-to-the-rack">Nvidia’s latest Rubin GPU</a> has 33 TFLOPs of FP64 compute, which is lower than the four-year-old H100’s 34 TFLOPs of FP64. For context, AMD’s year-old Instinct <a href="https://www.tomshardware.com/pc-components/gpus/amd-announces-mi350x-and-mi355x-ai-gpus-claims-up-to-4x-generational-gain-up-to-35x-faster-inference-performance">MI355X</a> peaks at 78.6 TFLOPs of FP64 compute, while its upcoming HPC-angled <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/amd-touts-instinct-mi430x-mi440x-and-mi455x-ai-accelerators-and-helios-rack-scale-ai-architecture-at-ces-full-mi400-series-family-fulfills-a-broad-range-of-infrastructure-and-customer-requirements">MI430X </a>peaks at 288 TFLOPs. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:5712px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="pk3xH2YfEacihEjQauD3nJ" name="IMG_1755.jpg" alt="Nvidia RTX Pro 6000 Blackwell GPUs" src="https://cdn.mos.cms.futurecdn.net/pk3xH2YfEacihEjQauD3nJ-1920-80.jpg" mos="" align="middle" fullscreen="" width="5712" height="3213" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p>Nvidia has leaned on emulation to drive FP64 performance higher, a <a href="https://www.amd.com/en/blogs/2026/exascale-was-the-beginning-amd-instinct--mi430x-gpus.html"><u>tactic that AMD is now pushing back on</u></a>, but regardless, it’s easy to see why the HPC world could feel left behind in this new AI era. AI is defined by fuzzy, low-precision data formats, which is diametrically opposed to the accuracy of high-precision FP64. And, over the last generations of hardware from the big green giant that is Nvidia, FP64 performance has seemingly stalled in favor of driving low-precision data formats for AI workloads. </p><p>“Maybe a few years ago, people in HPC might have felt a little threatened by AI because it was taking some of the limelight,” Tease told <em>Tom’s Hardware Premium. </em>“But I think now they’ve come to realize that AI is an essential part of scientific discovery today. It’s just, how are you going to deploy it?”</p><p>That central question recontextualizes the apparent conflict between AI compute and HPC compute. Tease pointed to researchers spawning dozens of agents to explore viable paths for a given intent, while still leaving the complex simulations for a traditional FP64-driven HPC fleet. “...to try to do that with old approaches where you have islands of CPUs or islands of GPUs or something like that, and they're very different. They're not interconnected. It's just not going to be able to keep up,” Tease said. </p><h2 id="enter-quantum">Enter quantum</h2><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2000px;"><p class="vanilla-image-block" style="padding-top:75.00%;"><img id="fHioj7u5XgSaFQhFyg8ASd" name="IBM-Quantum-Nighthawk-r2-processor-angle-in-hand" alt="IBM Quantum Nighthawk r2 processor." src="https://cdn.mos.cms.futurecdn.net/fHioj7u5XgSaFQhFyg8ASd-1920-80.jpg" mos="" align="middle" fullscreen="" width="2000" height="1500" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: IBM)</span></figcaption></figure><p>One of the challenges right now is actually connecting those systems, which is an area AMD is <a href="https://newsroom.amd.com/news/amd-builds-classical-foundation-quantum-future/"><u>exploring with its recent collaboration</u></a> with IBM and Oak Ridge National Laboratory to deploy AMD’s “classical” computing stack in the Umbra HPC system alongside IBM’s 129-qubit <a href="https://www.tomshardware.com/tech-industry/semiconductors/ibm-unveils-new-120-qubit-processor-and-software-stack">Quantum Nighthawk</a> quantum computer. The systems will connect to the Oak Ridge Leadership Computing Facility, allowing researchers to “explore how integrated quantum, AI, and scientific computing can accelerate scientific discoveries and national security.” </p><p>If the HPC crowd felt slighted by AI, you can imagine how quantum researchers probably feel. </p><p>Regardless, Tease says the future of scientific computing looks like a unified fleet that brings together AI, HPC, and quantum. “There [are] going to be a lot of problems where quantum will not be the be-all, end-all. There are some problems like human health, chemistry, where quantum’s going to allow us to do things we really can’t do with classical computing today. So having those things coexist, having them in a single workload, is where we want to get to.” </p><p>Scheduling workloads for HPC is a solved game, with open-source schedulers like Slurm driving much of the TOP500 to properly schedule workloads on supercomputers. Tease described something that goes deeper to connect these systems. “What we want to be able to do is not only do those kinds of ‘where should it go, here or there’ for HPC, we want to figure out when AI could be applied to it, when [GPU acceleration is] appropriate… and also when is it best to go in a quantum environment?”</p><p>The Oak Ridge Leadership Computing Facility is a test bed for figuring out how to hook these systems together properly. The press release explicitly states that it’s a “shared environment to evaluate the architectures, software, and workflows that future hybrid quantum systems will require.”  The system is a co-design initiative with IBM, bringing together quantum and classical computing into one system. </p><p>With the system, researchers are exploring how to accelerate quantum error correction with classical systems, a critical step in quantum computing to weed out environmental influences. The goal is ultimately to accelerate quantum systems in the Quantum Genesis Q competition, an initiative launched by the U.S. Department of Energy that will provide up to $215 million in funding for quantum research. </p><p>By late 2028, the <a href="https://www.energy.gov/science/articles/doe-launches-competition-accelerate-development-worlds-first-fault-tolerant"><u>competition calls for quantum computers</u></a> “with at least 100 logical qubits that are capable of performing hundreds of millions [of] fault-tolerant operations.” IBM describes the error correction research that will be performed at Oak Ridge as a “key component” in driving toward that goal. Separately, <a href="https://www.businesswire.com/news/home/20261112481657/en/PsiQuantum-and-AMD-Collaborate-to-Build-the-Computing-Foundation-for-Utility-Scale-Quantum-Computing"><u>AMD announced last week</u></a> an investment in quantum outfit PsiQuantum to provide CPUs, GPUs, and FPGAs to drive the company’s first fault-tolerant quantum system. </p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/supercomputers/researchers-recognize-that-ai-will-never-replace-hpc-says-amd-executive-high-precision-workloads-will-exist-for-a-very-long-time-even-as-ai-usage-expands</link>
                                                                            <description>
                            <![CDATA[ The emergence of AI has seemingly posed a threat to high-precision FP64 computation, but AMD doesn't think the HPC world is moving away from those complex calculations in the wake of AI. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">8T85LdfCECTAyYnNX5ymK4</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/odrYnFH3kdkxQe7pL2cCGb-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Mon, 16 Nov 2026 21:30:00 +0000</pubDate>                                                                                                                                                                                                                                <category><![CDATA[Supercomputers]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Jake Roach ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/h6PRM8bTimCTnNfoAYfjAi-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Jake Roach has been bending pins and busting solder joints since the mid-2000s. From trying to run scratched CDs of &lt;em&gt;Delta Force &lt;/em&gt;and &lt;em&gt;Unreal Tournament &lt;/em&gt;to spitting out virtual machines on a Threadripper, Jake has been on the hunt for the latest hardware and highest performance for decades. That eventually spun up a career, with Jake serving as Lead Reporter at Digital Trends, as well as contributing to outlets like XDA, PC Invasion, Business Insider, and WIRED. At Tom’s Hardware, Jake is focused on consumer and workstation CPUs. Outside working hours, you’ll find him knee-deep in the latest roguelite taking over Steam, spending way too much money on &lt;em&gt;Magic: The Gathering, &lt;/em&gt;or forcing his lazy corgi onto walks.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>false</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/odrYnFH3kdkxQe7pL2cCGb-1920-80.jpg">
                                                            <media:credit><![CDATA[IBM]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[IBM quantum system alongside Umbra supercomputer.]]></media:description>                                                            <media:text><![CDATA[IBM quantum system alongside Umbra supercomputer.]]></media:text>
                                <media:title type="plain"><![CDATA[IBM quantum system alongside Umbra supercomputer.]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/odrYnFH3kdkxQe7pL2cCGb-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>While AI continues to reshape industries across science and mathematics, traditional supercomputing has seemingly taken a back seat. However, AMD's Scott Tease told <em>Tom's Hardware Premium</em> that most HPC sites “recognize that AI will never replace HPC,” and suggests that building a heterogeneous compute stack is the best path forward to tackle bleeding-edge research, despite Nvidia pushing performance on lower-precision data types higher for AI workloads at the cost of native performance for the important double-precision FP64 format that’s critical for HPC workloads.</p><p>A heterogeneous compute stack brings together quantum, HPC, and AI under one roof, a fleet that AMD is helping pilot at Oak Ridge National Laboratory alongside IBM. Tease is AMD’s corporate vice president of HPC and sovereign AI, who joined the company earlier this year after more than a decade at Lenovo (and, before that, another 13 years at IBM). We spoke with Tease ahead of <a href="https://www.tomshardware.com/tag/sc26">SC26 </a>to get a read on the state of HPC, how it intersects with AI, and what’s actually driving the research community in a time when computer chips are driving headlines about international espionage plots. </p><h2 id="an-awful-lot-of-numbers">An awful lot of numbers</h2><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1920px;"><p class="vanilla-image-block" style="padding-top:60.00%;"><img id="3BFHbcRr5aki3FZCTBPAG5" name="instinct-hero" alt="AMD Instinct MI455X Accelerator" src="https://cdn.mos.cms.futurecdn.net/3BFHbcRr5aki3FZCTBPAG5-1920-80.png" mos="" align="middle" fullscreen="" width="1920" height="1152" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: AMD)</span></figcaption></figure><p>FP64 is important for HPC workloads because of its precision. FP64 is a hot-button issue, largely due to how Nvidia’s GPUs have progressed over the past two generations. Or rather, how they <em>haven't </em>progressed. <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-details-rubin-architectural-optimizations-for-inference-improvements-target-better-performance-and-efficiency-from-the-gpu-to-the-rack">Nvidia’s latest Rubin GPU</a> has 33 TFLOPs of FP64 compute, which is lower than the four-year-old H100’s 34 TFLOPs of FP64. For context, AMD’s year-old Instinct <a href="https://www.tomshardware.com/pc-components/gpus/amd-announces-mi350x-and-mi355x-ai-gpus-claims-up-to-4x-generational-gain-up-to-35x-faster-inference-performance">MI355X</a> peaks at 78.6 TFLOPs of FP64 compute, while its upcoming HPC-angled <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/amd-touts-instinct-mi430x-mi440x-and-mi455x-ai-accelerators-and-helios-rack-scale-ai-architecture-at-ces-full-mi400-series-family-fulfills-a-broad-range-of-infrastructure-and-customer-requirements">MI430X </a>peaks at 288 TFLOPs. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:5712px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="pk3xH2YfEacihEjQauD3nJ" name="IMG_1755.jpg" alt="Nvidia RTX Pro 6000 Blackwell GPUs" src="https://cdn.mos.cms.futurecdn.net/pk3xH2YfEacihEjQauD3nJ-1920-80.jpg" mos="" align="middle" fullscreen="" width="5712" height="3213" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p>Nvidia has leaned on emulation to drive FP64 performance higher, a <a href="https://www.amd.com/en/blogs/2026/exascale-was-the-beginning-amd-instinct--mi430x-gpus.html"><u>tactic that AMD is now pushing back on</u></a>, but regardless, it’s easy to see why the HPC world could feel left behind in this new AI era. AI is defined by fuzzy, low-precision data formats, which is diametrically opposed to the accuracy of high-precision FP64. And, over the last generations of hardware from the big green giant that is Nvidia, FP64 performance has seemingly stalled in favor of driving low-precision data formats for AI workloads. </p><p>“Maybe a few years ago, people in HPC might have felt a little threatened by AI because it was taking some of the limelight,” Tease told <em>Tom’s Hardware Premium. </em>“But I think now they’ve come to realize that AI is an essential part of scientific discovery today. It’s just, how are you going to deploy it?”</p><p>That central question recontextualizes the apparent conflict between AI compute and HPC compute. Tease pointed to researchers spawning dozens of agents to explore viable paths for a given intent, while still leaving the complex simulations for a traditional FP64-driven HPC fleet. “...to try to do that with old approaches where you have islands of CPUs or islands of GPUs or something like that, and they're very different. They're not interconnected. It's just not going to be able to keep up,” Tease said. </p><h2 id="enter-quantum">Enter quantum</h2><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2000px;"><p class="vanilla-image-block" style="padding-top:75.00%;"><img id="fHioj7u5XgSaFQhFyg8ASd" name="IBM-Quantum-Nighthawk-r2-processor-angle-in-hand" alt="IBM Quantum Nighthawk r2 processor." src="https://cdn.mos.cms.futurecdn.net/fHioj7u5XgSaFQhFyg8ASd-1920-80.jpg" mos="" align="middle" fullscreen="" width="2000" height="1500" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: IBM)</span></figcaption></figure><p>One of the challenges right now is actually connecting those systems, which is an area AMD is <a href="https://newsroom.amd.com/news/amd-builds-classical-foundation-quantum-future/"><u>exploring with its recent collaboration</u></a> with IBM and Oak Ridge National Laboratory to deploy AMD’s “classical” computing stack in the Umbra HPC system alongside IBM’s 129-qubit <a href="https://www.tomshardware.com/tech-industry/semiconductors/ibm-unveils-new-120-qubit-processor-and-software-stack">Quantum Nighthawk</a> quantum computer. The systems will connect to the Oak Ridge Leadership Computing Facility, allowing researchers to “explore how integrated quantum, AI, and scientific computing can accelerate scientific discoveries and national security.” </p><p>If the HPC crowd felt slighted by AI, you can imagine how quantum researchers probably feel. </p><p>Regardless, Tease says the future of scientific computing looks like a unified fleet that brings together AI, HPC, and quantum. “There [are] going to be a lot of problems where quantum will not be the be-all, end-all. There are some problems like human health, chemistry, where quantum’s going to allow us to do things we really can’t do with classical computing today. So having those things coexist, having them in a single workload, is where we want to get to.” </p><p>Scheduling workloads for HPC is a solved game, with open-source schedulers like Slurm driving much of the TOP500 to properly schedule workloads on supercomputers. Tease described something that goes deeper to connect these systems. “What we want to be able to do is not only do those kinds of ‘where should it go, here or there’ for HPC, we want to figure out when AI could be applied to it, when [GPU acceleration is] appropriate… and also when is it best to go in a quantum environment?”</p><p>The Oak Ridge Leadership Computing Facility is a test bed for figuring out how to hook these systems together properly. The press release explicitly states that it’s a “shared environment to evaluate the architectures, software, and workflows that future hybrid quantum systems will require.”  The system is a co-design initiative with IBM, bringing together quantum and classical computing into one system. </p><p>With the system, researchers are exploring how to accelerate quantum error correction with classical systems, a critical step in quantum computing to weed out environmental influences. The goal is ultimately to accelerate quantum systems in the Quantum Genesis Q competition, an initiative launched by the U.S. Department of Energy that will provide up to $215 million in funding for quantum research. </p><p>By late 2028, the <a href="https://www.energy.gov/science/articles/doe-launches-competition-accelerate-development-worlds-first-fault-tolerant"><u>competition calls for quantum computers</u></a> “with at least 100 logical qubits that are capable of performing hundreds of millions [of] fault-tolerant operations.” IBM describes the error correction research that will be performed at Oak Ridge as a “key component” in driving toward that goal. Separately, <a href="https://www.businesswire.com/news/home/20261112481657/en/PsiQuantum-and-AMD-Collaborate-to-Build-the-Computing-Foundation-for-Utility-Scale-Quantum-Computing"><u>AMD announced last week</u></a> an investment in quantum outfit PsiQuantum to provide CPUs, GPUs, and FPGAs to drive the company’s first fault-tolerant quantum system. </p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ AMD grabs over 40% of new TOP500 list for supercomputers, up 14% since last year  ]]></title>
                                                                                                <dc:content><![CDATA[ <p>With the coming of SC26, also comes a fresh edition of the TOP500 list, which ranks the world's most powerful supercomputers. The newly released list illustrates that AMD powers 40% of the list, up 14% compared to 2025. AMD-powered systems now tally up to 202 on the list, up from 177 last year. New to the list is Lux, a “SI + HPC supercomputer” developed by AMD and the U.S. Department of Energy (DOE), which debuted at position 14. </p><p>AMD’s representation in the top 10 is familiar. After debuting in June, <a href="https://www.tomshardware.com/tech-industry/supercomputers/chinas-lineshine-supercomputer-dethrones-us-el-capitan-secures-first-place-in-top-500-list-first-machine-in-the-rankings-to-sustain-more-than-2-exaflops-of-double-precision-performance-using-only-cpus"><u>China’s LineShine remains</u></a> in the top slot. El Capitan, powered by AMD chips, is in the second slot, followed by Frontier in third, HPC7 in sixth, and HPC6 in eighth, the same as the most recent TOP500 list. </p><p>Lux was announced a little over a year ago in <a href="https://www.tomshardware.com/tech-industry/supercomputers/u-s-department-of-energy-and-amd-cut-a-usd1-billion-deal-for-two-ai-supercomputers-pairing-has-already-birthed-the-two-fastest-machines-on-the-planet"><u>a $1 billion deal between the DOE and AMD</u></a> to develop two supercomputers. This is the first of the two, and the first time it’s been on the TOP500. The system was built by HPE and developed by AMD and DOE as part of the U.S. government’s Genesis Mission. The machine is housed in Oak Ridge National Laboratory’s Oak Ridge Leadership Computing Facility (OLCF), alongside machines like Frontier. </p><p>Lux <a href="https://www.olcf.ornl.gov/olcf-resources/compute-systems/lux/"><u>relies on over 4,000 Instinct MI355X GPUs</u></a> separated into over 500 nodes. For HPC workloads using FP64, the machine is capable of 317 PetaFLOPS. For AI training, the machine peaks at 20 ExaFLOPS of FP8, and for inference, 40 ExaFLOPS of FP4. With 288 GB of HBM3E per GPU, the entire system has access to 1.1 PB. </p><p>The machine uses AMD EPYC host CPUs — we weren’t able to find details on the specific generation— and AMD’s Pensando networking, with nodes built inside HPE’s ProLiant Compute XD685.</p><div ><table><caption>TOP500 Top 10 (Nov. 2026)</caption><tbody><tr><td class="firstcol " ><p><strong>Rank</strong></p></td><td  ><p><strong>Prev. Rank</strong></p></td><td  ><p><strong>Name</strong></p></td><td  ><p><strong>System</strong></p></td></tr><tr><td class="firstcol " ><p>1</p></td><td  ><p>1</p></td><td  ><p>LineShine</p></td><td  ><p>LingKun, LX2 304C 1.55GHz, LingQi, Kylin OS</p></td></tr><tr><td class="firstcol " ><p>2</p></td><td  ><p>2</p></td><td  ><p>El Capitan</p></td><td  ><p>HPE Cray EX255a, AMD 4th Gen EPYC 24C 1.8GHz, AMD Instinct MI300A, Slingshot-11, TOSS</p></td></tr><tr><td class="firstcol " ><p>3</p></td><td  ><p>3</p></td><td  ><p>Frontier</p></td><td  ><p>HPE Cray EX235a, AMD Optimized 3rd Generation EPYC 64C 2GHz, AMD Instinct MI250X, Slingshot-11, HPE Cray OS</p></td></tr><tr><td class="firstcol " ><p>4</p></td><td  ><p>4</p></td><td  ><p>Aurora</p></td><td  ><p>HPE Cray EX - Intel Exascale Compute Blade, Xeon CPU Max 9470 52C 2.4GHz, Intel Data Center GPU Max, Slingshot-11</p></td></tr><tr><td class="firstcol " ><p>5</p></td><td  ><p>5</p></td><td  ><p>JUPITER Booster</p></td><td  ><p>BullSequana XH3000, GH Superchip 72C 3GHz, NVIDIA GH200 Superchip, Quad-Rail NVIDIA InfiniBand NDR200, RedHat Enterprise Linux</p></td></tr><tr><td class="firstcol " ><p>6</p></td><td  ><p>6</p></td><td  ><p>HPC7</p></td><td  ><p>HPE Cray EX255a, AMD 4th Gen EPYC 24C 1.8GHz, AMD Instinct MI300A, Slingshot-11, RHEL 9</p></td></tr><tr><td class="firstcol " ><p>7</p></td><td  ><p>7</p></td><td  ><p>Eagle</p></td><td  ><p>Microsoft NDv5, Xeon Platinum 8480C 48C 2GHz, NVIDIA H100, NVIDIA Infiniband NDR</p></td></tr><tr><td class="firstcol " ><p>8</p></td><td  ><p>8</p></td><td  ><p>HPC6</p></td><td  ><p>HPE Cray EX235a, AMD Optimized 3rd Generation EPYC 64C 2GHz, AMD Instinct MI250X, Slingshot-11, RHEL 8.9</p></td></tr><tr><td class="firstcol " ><p>9</p></td><td  ><p>9</p></td><td  ><p>Supercomputer Fugaku</p></td><td  ><p>Supercomputer Fugaku, A64FX 48C 2.2GHz, Tofu interconnect D</p></td></tr><tr><td class="firstcol " ><p>10</p></td><td  ><p>10</p></td><td  ><p>Alps</p></td><td  ><p>HPE Cray EX254n, NVIDIA Grace 72C 3.1GHz, NVIDIA GH200 Superchip, Slingshot-11, HPE Cray OS</p></td></tr></tbody></table></div><p>The machine has 3.5 million node-hours (one node running for one hour) per year, half of which are reserved for Genesis Mission research. <a href="https://www.whitehouse.gov/presidential-actions/2025/11/launching-the-genesis-mission/"><u>Genesis Mission</u></a> was established through an executive order in November 2025 with the stated goal of creating a “dedicated, coordinated national effort to unleash a new age of AI‑accelerated innovation and discovery that can solve the most challenging problems of this century.” </p><p>More specifically, the order lays out training scientific foundation models on Federal datasets, and creating “AI agents to test new hypotheses, automate research workflows, and accelerate scientific breakthroughs.” Lux is the first dedicated supercomputer for the Genesis Mission, but it joins existing machines that have been at least partially dedicated to research, including the Intel-powered Aurora at Argonne National Laboratory, which arrived fourth in the most recent TOP500. </p><p>Lux is built primarily for AI research, not as a successor to Frontier. Discovery is on track to debut as a next-gen flagship exascale supercomputer in 2028, also developed jointly by the DOE, AMD, and HPE. </p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/supercomputers/amd-grabs-over-40-percent-of-new-top500-list-for-supercomputers-up-14-percent-since-last-year-new-lux-system-developed-between-amd-and-department-of-energy-debuts-at-no-14-with-epyc-cpus-and-instinct-mi355x-gpus</link>
                                                                            <description>
                            <![CDATA[ AMD's Lux supercomputer debuted at position 14 on the new TOP500 list published during SC26. It's one of two machines AMD is developing as part of a $1 billion deal with the U.S. Department of Energy. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">HpUGmzokNexc4c5zeXbixc</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/kiQWCdGYbYcrsLVKdCSDi3-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Mon, 16 Nov 2026 21:00:00 +0000</pubDate>                                                                                                                                <updated>Mon, 16 Nov 2026 21:05:40 +0000</updated>
                                                                                                                                            <category><![CDATA[Supercomputers]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Jake Roach ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/h6PRM8bTimCTnNfoAYfjAi-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Jake Roach has been bending pins and busting solder joints since the mid-2000s. From trying to run scratched CDs of &lt;em&gt;Delta Force &lt;/em&gt;and &lt;em&gt;Unreal Tournament &lt;/em&gt;to spitting out virtual machines on a Threadripper, Jake has been on the hunt for the latest hardware and highest performance for decades. That eventually spun up a career, with Jake serving as Lead Reporter at Digital Trends, as well as contributing to outlets like XDA, PC Invasion, Business Insider, and WIRED. At Tom’s Hardware, Jake is focused on consumer and workstation CPUs. Outside working hours, you’ll find him knee-deep in the latest roguelite taking over Steam, spending way too much money on &lt;em&gt;Magic: The Gathering, &lt;/em&gt;or forcing his lazy corgi onto walks.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>false</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/kiQWCdGYbYcrsLVKdCSDi3-1920-80.jpg">
                                                            <media:credit><![CDATA[Oak Ridge National Laboratory]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Lux supercomputer.]]></media:description>                                                            <media:text><![CDATA[Lux supercomputer.]]></media:text>
                                <media:title type="plain"><![CDATA[Lux supercomputer.]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/kiQWCdGYbYcrsLVKdCSDi3-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>With the coming of SC26, also comes a fresh edition of the TOP500 list, which ranks the world's most powerful supercomputers. The newly released list illustrates that AMD powers 40% of the list, up 14% compared to 2025. AMD-powered systems now tally up to 202 on the list, up from 177 last year. New to the list is Lux, a “SI + HPC supercomputer” developed by AMD and the U.S. Department of Energy (DOE), which debuted at position 14. </p><p>AMD’s representation in the top 10 is familiar. After debuting in June, <a href="https://www.tomshardware.com/tech-industry/supercomputers/chinas-lineshine-supercomputer-dethrones-us-el-capitan-secures-first-place-in-top-500-list-first-machine-in-the-rankings-to-sustain-more-than-2-exaflops-of-double-precision-performance-using-only-cpus"><u>China’s LineShine remains</u></a> in the top slot. El Capitan, powered by AMD chips, is in the second slot, followed by Frontier in third, HPC7 in sixth, and HPC6 in eighth, the same as the most recent TOP500 list. </p><p>Lux was announced a little over a year ago in <a href="https://www.tomshardware.com/tech-industry/supercomputers/u-s-department-of-energy-and-amd-cut-a-usd1-billion-deal-for-two-ai-supercomputers-pairing-has-already-birthed-the-two-fastest-machines-on-the-planet"><u>a $1 billion deal between the DOE and AMD</u></a> to develop two supercomputers. This is the first of the two, and the first time it’s been on the TOP500. The system was built by HPE and developed by AMD and DOE as part of the U.S. government’s Genesis Mission. The machine is housed in Oak Ridge National Laboratory’s Oak Ridge Leadership Computing Facility (OLCF), alongside machines like Frontier. </p><p>Lux <a href="https://www.olcf.ornl.gov/olcf-resources/compute-systems/lux/"><u>relies on over 4,000 Instinct MI355X GPUs</u></a> separated into over 500 nodes. For HPC workloads using FP64, the machine is capable of 317 PetaFLOPS. For AI training, the machine peaks at 20 ExaFLOPS of FP8, and for inference, 40 ExaFLOPS of FP4. With 288 GB of HBM3E per GPU, the entire system has access to 1.1 PB. </p><p>The machine uses AMD EPYC host CPUs — we weren’t able to find details on the specific generation— and AMD’s Pensando networking, with nodes built inside HPE’s ProLiant Compute XD685.</p><div ><table><caption>TOP500 Top 10 (Nov. 2026)</caption><tbody><tr><td class="firstcol " ><p><strong>Rank</strong></p></td><td  ><p><strong>Prev. Rank</strong></p></td><td  ><p><strong>Name</strong></p></td><td  ><p><strong>System</strong></p></td></tr><tr><td class="firstcol " ><p>1</p></td><td  ><p>1</p></td><td  ><p>LineShine</p></td><td  ><p>LingKun, LX2 304C 1.55GHz, LingQi, Kylin OS</p></td></tr><tr><td class="firstcol " ><p>2</p></td><td  ><p>2</p></td><td  ><p>El Capitan</p></td><td  ><p>HPE Cray EX255a, AMD 4th Gen EPYC 24C 1.8GHz, AMD Instinct MI300A, Slingshot-11, TOSS</p></td></tr><tr><td class="firstcol " ><p>3</p></td><td  ><p>3</p></td><td  ><p>Frontier</p></td><td  ><p>HPE Cray EX235a, AMD Optimized 3rd Generation EPYC 64C 2GHz, AMD Instinct MI250X, Slingshot-11, HPE Cray OS</p></td></tr><tr><td class="firstcol " ><p>4</p></td><td  ><p>4</p></td><td  ><p>Aurora</p></td><td  ><p>HPE Cray EX - Intel Exascale Compute Blade, Xeon CPU Max 9470 52C 2.4GHz, Intel Data Center GPU Max, Slingshot-11</p></td></tr><tr><td class="firstcol " ><p>5</p></td><td  ><p>5</p></td><td  ><p>JUPITER Booster</p></td><td  ><p>BullSequana XH3000, GH Superchip 72C 3GHz, NVIDIA GH200 Superchip, Quad-Rail NVIDIA InfiniBand NDR200, RedHat Enterprise Linux</p></td></tr><tr><td class="firstcol " ><p>6</p></td><td  ><p>6</p></td><td  ><p>HPC7</p></td><td  ><p>HPE Cray EX255a, AMD 4th Gen EPYC 24C 1.8GHz, AMD Instinct MI300A, Slingshot-11, RHEL 9</p></td></tr><tr><td class="firstcol " ><p>7</p></td><td  ><p>7</p></td><td  ><p>Eagle</p></td><td  ><p>Microsoft NDv5, Xeon Platinum 8480C 48C 2GHz, NVIDIA H100, NVIDIA Infiniband NDR</p></td></tr><tr><td class="firstcol " ><p>8</p></td><td  ><p>8</p></td><td  ><p>HPC6</p></td><td  ><p>HPE Cray EX235a, AMD Optimized 3rd Generation EPYC 64C 2GHz, AMD Instinct MI250X, Slingshot-11, RHEL 8.9</p></td></tr><tr><td class="firstcol " ><p>9</p></td><td  ><p>9</p></td><td  ><p>Supercomputer Fugaku</p></td><td  ><p>Supercomputer Fugaku, A64FX 48C 2.2GHz, Tofu interconnect D</p></td></tr><tr><td class="firstcol " ><p>10</p></td><td  ><p>10</p></td><td  ><p>Alps</p></td><td  ><p>HPE Cray EX254n, NVIDIA Grace 72C 3.1GHz, NVIDIA GH200 Superchip, Slingshot-11, HPE Cray OS</p></td></tr></tbody></table></div><p>The machine has 3.5 million node-hours (one node running for one hour) per year, half of which are reserved for Genesis Mission research. <a href="https://www.whitehouse.gov/presidential-actions/2025/11/launching-the-genesis-mission/"><u>Genesis Mission</u></a> was established through an executive order in November 2025 with the stated goal of creating a “dedicated, coordinated national effort to unleash a new age of AI‑accelerated innovation and discovery that can solve the most challenging problems of this century.” </p><p>More specifically, the order lays out training scientific foundation models on Federal datasets, and creating “AI agents to test new hypotheses, automate research workflows, and accelerate scientific breakthroughs.” Lux is the first dedicated supercomputer for the Genesis Mission, but it joins existing machines that have been at least partially dedicated to research, including the Intel-powered Aurora at Argonne National Laboratory, which arrived fourth in the most recent TOP500. </p><p>Lux is built primarily for AI research, not as a successor to Frontier. Discovery is on track to debut as a next-gen flagship exascale supercomputer in 2028, also developed jointly by the DOE, AMD, and HPE. </p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Nvidia announces Rubin HPC GPU with 'record number of CUDA cores' at Supercomputing 2026  ]]></title>
                                                                                                <dc:content><![CDATA[ <p>The Supercomputing 2026 conference is under way, and despite continuing turmoil elsewhere in the company, Nvidia has lots to reveal at the show. Foremost among its announcements today, Nvidia took the wraps off a new member of the Rubin GPU family <a href="https://blogs.nvidia.com/blog/rubin-hpc/" target="_blank">called Rubin HPC</a>, a chip that the company says is meant to accelerate agentic workflows and simulation for scientists, engineers, and analysts working with large data sets.</p><p>To accelerate those most demanding workloads, Nvidia says Rubin HPC has the highest number of CUDA cores on any of its GPUs - "more than 50,000" - which might imply that this chip uses a configuration similar to the four-die Rubin Ultra GPU, which links those dies together using NVLink-C2C. </p><p>Today's Rubin GPU has 224 SMs and 28,672 CUDA cores, so scaling compute resources to the degree Rubin HPC does would necessarily require more dies on a package. Nvidia didn't reveal more details of the chip's physical layout alongside the announcement. (The representative image shared by the company wouldn't appear to take this into account.)</p><p>To keep data moving between Rubin HPC GPUs, Nvidia will use its NVLink 7 interconnect. The company didn't share any new details of the bandwidth or physical layer of this interconnect within the rack, <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-demonstrates-rubin-ultra-tray-worlds-1st-ai-gpu-with-1tb-of-hbm4e" target="_blank">but past disclosures about the tech</a> have indicated that it retains the 3.6 TB/s of per-GPU bandwidth that NVLink 6 delivers while enlarging the size of its scale-up domain.</p><p>Nvidia didn't share any details of the Rubin HPC GPU's HBM capacity or bandwidth, and it also didn't share whether this part includes a rebalancing of compute resources towards native FP64, as AMD and Intel are touting at the conference with their own GPUs this week.</p><p>Nvidia instead touts the suitability of this chip for scientists and engineers using agentic workflows, for which the company expects faster simulation performance and therefore faster opportunities for discovery in those fields thanks to better agent performance in those domains.</p><p>In fields like climate research, genomic analysis, and financial simulation, Nvidia says the Rubin HPC GPU's "balanced design," with both higher compute resources and (presumably) faster data movement between nodes, will help feed the chip with the massive data sets typical of those workloads, minimizing idle compute resources and achieving better utilization.</p><p>Researchers, engineers, and analysts marshaling agent swarms to explore the problem spaces in their domains will also be able to take advantage of Nvidia's CUDA-X domain-specific libraries to boost performance on Rubin HPC, tightening the reasoning loop from hypothesis to simulation to refinement and hastening the pursuit of the most promising results.</p><p>Nvidia says Rubin HPC will arrive sometime in 2028.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/pc-components/gpus/nvidia-announces-rubin-hpc-gpu-with-record-number-of-cuda-cores-at-supercomputing-2026-new-chip-accelerates-agentic-hpc-workloads-arrives-in-2028</link>
                                                                            <description>
                            <![CDATA[ Nvidia took the wraps off its Rubin HPC GPU at Supercomputing 2026. This new GPU will give scientists, engineers, and analysts over 50,000 CUDA cores to speed up agentic HPC workloads when it arrives in 2028. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">VrcYsqV72WUvRXsBKEao7U</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/wJ2qnaAhPDGJQVsKBgLAri-1920-80.png" type="image/png" length="0"></enclosure>
                                                                        <pubDate>Mon, 16 Nov 2026 19:05:24 +0000</pubDate>                                                                                                                                                                                                                                <category><![CDATA[GPUs]]></category>
                                                    <category><![CDATA[PC Components]]></category>
                                                                                                                    <dc:creator><![CDATA[ Jeffrey Kampman ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/8JCjGs5yVZds2YdKmzjUDE-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Jeff Kampman has been playing PC games ever since he learned how to fire up freeware CDs from the DOS command line. He started building his own PCs in the mid-aughts and later turned that passion into a career, working as a news and guides writer, reviewer, and ultimately Editor-in-Chief at The Tech Report, where he dove deep on CPUs and GPUs (and more) in pursuit of the smoothest gaming experiences around. Jeff later took on roles at Asus and Intel as a technical marketer before joining Tom&#039;s Hardware. As Senior Analyst, Graphics, Jeff covers everything from integrated graphics processors to discrete graphics cards to the massive data center GPU installations powering our AI future. Jeff is also a hobbyist photographer, Twitch streamer, espresso enthusiast, and runner.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>false</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/png" url="https://cdn.mos.cms.futurecdn.net/wJ2qnaAhPDGJQVsKBgLAri-1920-80.png">
                                                            <media:credit><![CDATA[Nvidia]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[A representation of the Rubin HPC GPU]]></media:description>                                                            <media:text><![CDATA[A representation of the Rubin HPC GPU]]></media:text>
                                <media:title type="plain"><![CDATA[A representation of the Rubin HPC GPU]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/wJ2qnaAhPDGJQVsKBgLAri-1920-80.png" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>The Supercomputing 2026 conference is under way, and despite continuing turmoil elsewhere in the company, Nvidia has lots to reveal at the show. Foremost among its announcements today, Nvidia took the wraps off a new member of the Rubin GPU family <a href="https://blogs.nvidia.com/blog/rubin-hpc/" target="_blank">called Rubin HPC</a>, a chip that the company says is meant to accelerate agentic workflows and simulation for scientists, engineers, and analysts working with large data sets.</p><p>To accelerate those most demanding workloads, Nvidia says Rubin HPC has the highest number of CUDA cores on any of its GPUs - "more than 50,000" - which might imply that this chip uses a configuration similar to the four-die Rubin Ultra GPU, which links those dies together using NVLink-C2C. </p><p>Today's Rubin GPU has 224 SMs and 28,672 CUDA cores, so scaling compute resources to the degree Rubin HPC does would necessarily require more dies on a package. Nvidia didn't reveal more details of the chip's physical layout alongside the announcement. (The representative image shared by the company wouldn't appear to take this into account.)</p><p>To keep data moving between Rubin HPC GPUs, Nvidia will use its NVLink 7 interconnect. The company didn't share any new details of the bandwidth or physical layer of this interconnect within the rack, <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-demonstrates-rubin-ultra-tray-worlds-1st-ai-gpu-with-1tb-of-hbm4e" target="_blank">but past disclosures about the tech</a> have indicated that it retains the 3.6 TB/s of per-GPU bandwidth that NVLink 6 delivers while enlarging the size of its scale-up domain.</p><p>Nvidia didn't share any details of the Rubin HPC GPU's HBM capacity or bandwidth, and it also didn't share whether this part includes a rebalancing of compute resources towards native FP64, as AMD and Intel are touting at the conference with their own GPUs this week.</p><p>Nvidia instead touts the suitability of this chip for scientists and engineers using agentic workflows, for which the company expects faster simulation performance and therefore faster opportunities for discovery in those fields thanks to better agent performance in those domains.</p><p>In fields like climate research, genomic analysis, and financial simulation, Nvidia says the Rubin HPC GPU's "balanced design," with both higher compute resources and (presumably) faster data movement between nodes, will help feed the chip with the massive data sets typical of those workloads, minimizing idle compute resources and achieving better utilization.</p><p>Researchers, engineers, and analysts marshaling agent swarms to explore the problem spaces in their domains will also be able to take advantage of Nvidia's CUDA-X domain-specific libraries to boost performance on Rubin HPC, tightening the reasoning loop from hypothesis to simulation to refinement and hastening the pursuit of the most promising results.</p><p>Nvidia says Rubin HPC will arrive sometime in 2028.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Intel bets on Classiq acquisition to turn Quantum efforts into commercially viable business  ]]></title>
                                                                                                <dc:content><![CDATA[ <p>After having spent more than a decade pursuing quantum computing research, Intel is making its first real push at commercializing its efforts. First, with the pending acquisition of quantum computing company Classiq, which has an existing list of enterprise customers. Intel says it'll be the only company that can produce qubit chips and cryogenic control electronics at scale; the deal would give it a complete top-to-bottom hardware-to-software stack, and Intel will also gain the chance to become a leading vendor of quantum computing products. Intel's acquisition of Classiq is expected to complete in the first half of 2027. </p><p>The key ingredient of that strategy is the pending <a href="https://www.intel.com/content/www/us/en/newsroom/news/corporate/intel-classiq-accelerate-scalable-quantum-computing.html" target="_blank">acquisition of Classiq</a>, which offers a hardware-agnostic quantum software development platform that enables engineers to design, synthesize, optimize, simulate, and execute quantum algorithms using high-level programming tools and AI-assisted automation, essentially untying software design from the complexity of quantum hardware. Meanwhile, Intel also plans to manufacture quantum processors using specialized versions of its <a href="https://www.tomshardware.com/tech-industry/semiconductors/intels-performance-enhanced-18a-p-process-enters-risk-production-enhanced-node-promises-9-percent-performance-improvement-at-iso-power">18A process </a>technology, offer foundry services, and eventually sell complete quantum computers. Although the company expects its quantum business to become profitable, it acknowledges that commercial success will take time.</p><p>Intel has spent more than a decade developing quantum computing technologies, including silicon spin-qubit processors, cryogenic control electronics, and software for hybrid quantum-classical computing. However, the company has largely treated these activities as research and development projects rather than an actual business. The acquisition of Classiq, which Intel expects to complete in the first half of 2027, marks a shift toward commercialization as Classiq already has many clients and established revenue streams.</p><p>"Up until now […] our quantum computing group for the last ten years was […] purely a research and development project," said Pushkar Ranade, chief technology officer at Intel, during a media briefing. "It was not really viewed as a commercial activity."</p><h2 id="intel-39-s-quantum-hardware-the-journey-so-far">Intel's quantum hardware: The journey so far</h2><p>Intel's quantum computing journey <a href="https://www.intc.com/news-events/press-releases/detail/316/intel-invests-us50-million-to-advance-quantum-computing">kicked off in 2015</a> with a research collaboration with QuTech — the quantum research institute of TU Delft and TNO — to develop quantum computing technologies, including silicon spin qubits. The idea was to leverage Intel's extensive expertise in silicon transistor development and its high-volume semiconductor manufacturing infrastructure to produce quantum processors in the same fabs that make conventional CPUs, overcoming the manufacturing scalability challenges alternative quantum computing architectures faced at the time. Intel subsequently developed <a href="https://www.tomshardware.com/news/intel-details-its-cryogenic-horse-ridge-quantum-control-chip">Horse Ridge</a>, a cryogenic control SoC manufactured using a 22nm process technology in 2020. Three years later, the company introduced <a href="https://www.tomshardware.com/news/intel-announce-tunnel-falls-quantum-research-chip">Tunnel Falls</a>, a 12-qubit silicon spin processor manufactured on a 300mm wafer.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1600px;"><p class="vanilla-image-block" style="padding-top:56.19%;"><img id="Qgqxx54Kpwyo7WwSahsFvC" name="Intel-Quantum-Engineers.jpg" alt="Intel silicon spin qubit progress" src="https://cdn.mos.cms.futurecdn.net/Qgqxx54Kpwyo7WwSahsFvC-1920-80.jpg" mos="" align="middle" fullscreen="" width="1600" height="899" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Intel )</span></figcaption></figure><p>Intel's quantum computing effort focuses on silicon spin qubits, which use spin states of individual electrons to store and manipulate quantum information. The company chose this approach partly because silicon spin qubits can be manufactured using equipment and processes similar to those employed for conventional semiconductor devices. This compatibility enables Intel to re-use most of its existing manufacturing infrastructure and expertise rather than build dedicated semiconductor fabrication facilities for quantum processors that yet have to become commercial products.  </p><p>Despite the <a href="https://www.tomshardware.com/tech-industry/policy/intel-layoffs-to-hit-data-center-group-division-focused-on-server-cpus-ai-chips-and-data-center-architecture-to-be-hit-by-an-unknown-number-of-cuts">recent cuts at Intel</a>, the company sustained its quantum computing project. Intel has developed two specialized versions of its 18A manufacturing technology for quantum computing: 18A-Q for silicon spin-qubit processors and 18A-CC for cryogenic control and readout electronics. To define/build 18A-Q, Intel has collaborated with Hitachi, AIST, and NEDO (three major supporters of quantum computing from Japan), and there appears to be a design kit intended to support large quantum computing systems. Without a doubt, 18A-Q and 18A-CC put Intel on the map of quantum computing chips in a big way. A couple of things should be kept in mind, though.</p><p>"The underlying device that we use, underlying process that we use to make the actual qubit device is a unique process, but it fits into those design rules," Ranade explained. "We can make these chips […] in the same fab that we use to make our CPUs."</p><p>Although Intel 18AQ uses BEOL interconnect geometry and design rules derived from Intel 18A, its underlying qubit devices require a different manufacturing process. According to Ranade, the technology requires a specialized thin-film stack and several additional manufacturing tools. Nevertheless, more than 90% of the manufacturing equipment is identical to that used for conventional semiconductor production, according to Intel. </p><p>"Intel 18A-CC, meanwhile, uses 18A-based geometry but requires additional characterization for operation at cryogenic temperatures," Ranade said. "Crucially, Intel can manufacture quantum processors alongside conventional CPUs in existing semiconductor fabs, which could reduce the capital expenditure required to scale production. […] More than 90% of the process equipment is exactly the same."</p><h2 id="a-major-piece-in-the-quantum-computing-puzzle-classiq">A major piece in the quantum computing puzzle: Classiq</h2><p>For conventional x86 architectures and IA-64 (while it was active), Intel has historically offered compilers that made software run better on its hardware (and sometimes on AMD hardware too). By contrast, with its quantum computing efforts, Intel has generally focused on hardware and while it has its own Quantum SDK that includes a compiler, this is not enough in the quantum computing realm as Classiq's offering operates at a higher abstraction level. With the acquisition of Classiq, the company intends to fill that gap.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2284px;"><p class="vanilla-image-block" style="padding-top:75.00%;"><img id="22SNGmTTWsPq3frQrgvRWe" name="IBM Quantum Loon wafer" alt="An IBM researcher holding a 300mm IBM Quantum Loon wafer" src="https://cdn.mos.cms.futurecdn.net/22SNGmTTWsPq3frQrgvRWe-1920-80.jpg" mos="" align="middle" fullscreen="" width="2284" height="1713" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: IBM)</span></figcaption></figure><p>Classiq develops a hardware-agnostic quantum software platform that applies electronic design automation (EDA) principles to quantum programming, which lets developers design, synthesize, and optimize quantum algorithms without manually constructing circuits from individual quantum gates. The company's software provides high-level modeling, automated algorithm synthesis, optimization, debugging, and execution capabilities across different quantum computing architectures, which somewhat resembles Intel's approach to its x86 compilers, albeit working at a different level.</p><p>Intel already has its Quantum SDK, which includes a C++-based compiler, runtime environment, and simulator for hybrid quantum-classical applications. Classiq's platform will sit above these tools and provide an application-development environment for programmers who lack extensive expertise in quantum computing. </p><p>Intel compares the potential significance of Classiq's software layer to Microsoft's MS-DOS, which helped establish the personal computer software ecosystem; Nvidia's CUDA, which enabled developers to exploit GPUs for general-purpose accelerated computing; and EDA tools, which abstracted high-level design descriptions from the complexities of transistor-level integrated circuit implementation and essentially set the stage for the multi-billion foundry industry today.</p><p>Furthermore, unlike Intel's quantum hardware business, Classiq already has commercial customers and generates revenue. The company employs approximately 130 people and serves more than 50 enterprise customers, including BMW, Comcast, Lufthansa, Rolls-Royce, and Toshiba. It also participates in more than 10 national quantum computing programs and supports over 50 quantum processors and simulators. </p><p>As existing quantum processors remain insufficient for many practical applications, companies can develop and optimize software ahead of hardware availability. As a result, these customers primarily use Classiq products to develop quantum algorithms, train engineers, and create intellectual property in preparation for commercially useful quantum computers. </p><p>Intel's CTO described Classiq's revenue and anticipated growth as healthy, but did not disclose actual financial figures. </p><p>"They have a very healthy revenue this year and a very healthy revenue CAGR over the coming years," Ranade said. "We hope that… we want that to continue." </p><p>It is necessary to note that Intel intends to preserve Classiq's hardware-agnostic approach, which means allowing the company to continue supporting quantum processors and simulators from competing vendors. This is essential to maintain Classiq's existing customer relationships and expand its revenue opportunities, as its clients can/will be able to develop quantum applications without committing to a particular hardware architecture. Obviously, Intel can generate software revenue regardless of which quantum computing technologies ultimately become commercially successful. </p><p>"We very much want to maintain that hardware-agnostic platform that they have developed to attract more and more customers," Ranade said.  </p><h2 id="from-quantum-chips-to-complete-systems">From quantum chips to complete systems</h2><p>When it comes to quantum computing, Intel's ambitions extend beyond manufacturing silicon spin-qubit processors. The company intends to become a supplier of complete quantum computing systems, which would combine quantum processors, cryogenic control electronics, classical computing resources, and software. For now, Intel has most of the building blocks required to develop complete quantum computing systems, but it still has to demonstrate that it can integrate them into commercially viable quantum computing at scale.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:4910px;"><p class="vanilla-image-block" style="padding-top:66.66%;"><img id="fVuF6LJBxPGNZ7j2uMyRk8" name="Intel-Quantum-17-Qubit-2.jpg" alt="Quantum" src="https://cdn.mos.cms.futurecdn.net/fVuF6LJBxPGNZ7j2uMyRk8-1920-80.jpg" mos="" align="middle" fullscreen="" width="4910" height="3273" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Intel)</span></figcaption></figure><p>"We want to be in the business of building quantum foundry components, but also complete quantum systems," said Pushkar Ranade, chief technology officer at Intel.</p><p>Intel is also considering opportunities to provide quantum foundry services to other hardware developers. While its internal processor development will remain focused on silicon spin qubits, the company is open to manufacturing components for alternative quantum computing technologies, which somewhat resembles its approach to its x86 CPU and foundry businesses. </p><p>"We are committed to spin qubits," Ranade said. "But we are also open to providing foundry services to other modalities." </p><p>Intel believes that its capabilities in photonics, advanced packaging, and semiconductor manufacturing could become commercially relevant even to companies developing quantum processors based on different physical principles than silicon qubits. </p><h2 id="can-quantum-computing-become-profitable">Can quantum computing become profitable?</h2><p>Intel's decision to expand its quantum computing activities with the takeover of Classiq comes as chief executive Lip-Bu Tan emphasized profitability, financial discipline, and high-margin products across the company’s established businesses. Given that Intel's quantum computing initiative barely generates any revenue, let alone profits, the CEO stance raises questions about the viability of quantum computing for Intel, especially in a situation where the industry has yet to demonstrate widespread practical applications that justify the cost of large-scale quantum hardware.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2048px;"><p class="vanilla-image-block" style="padding-top:66.70%;"><img id="iWjCt8g7ndNtZ2cWPevjSh" name="devcomp.jpeg" alt="Quantum computer in Devs" src="https://cdn.mos.cms.futurecdn.net/iWjCt8g7ndNtZ2cWPevjSh-1920-80.jpeg" mos="" align="middle" fullscreen="" width="2048" height="1366" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: DNA TV/ FXP)</span></figcaption></figure><p>When asked how Intel’s quantum computing strategy fits Tan's profitability requirements, Ranade explained that the company's immediate financial priorities primarily concern its established client and server processor businesses. Nevertheless, Intel must continue investing in emerging technologies that could shape future computing markets. </p><p>"This is really a bet on the future," Ranade said. "This is something we also have to do in addition to our mainstream businesses today."</p><p>The Intel CTO acknowledged that quantum computing is not set to become a high-margin business in the near future. Meanwhile, the acquisition of Classiq gives Intel an established commercial software operation and revenue stream here and now, as well as an opportunity to expand customer relationships well before its own quantum processors reach commercial maturity. Intel also hopes that combining its hardware and software capabilities (probably including Classiq) will attract customers interested in complete quantum computing solutions.</p><p>When asked whether vertical integration would make Intel's quantum computing operations profitable, Ranade acknowledged that he could not predict gross margins of the quantum computing business because the industry is still in its infancy. Nevertheless, he expressed confidence in the commercial prospects of Intel's quantum hardware and software.</p><p>"I expect it to become a profitable business very quickly on the hardware side, but too early to tell," Ranade said.</p><h2 id="quantum-computing-will-remain-hybrid">Quantum computing will remain hybrid</h2><p>Even though quantum processors evolve at a rapid pace, the technology is unlikely to replace existing CPUs and GPUs for conventional computing any time soon. Instead, Intel expects quantum computing systems to operate alongside the usual processors in data centers and supercomputing environments. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1280px;"><p class="vanilla-image-block" style="padding-top:50.00%;"><img id="aKKUfnAx4WFf5kQpoFLYAS" name="two-gate-quantum-japan.jpg" alt="A diagram of the system." src="https://cdn.mos.cms.futurecdn.net/aKKUfnAx4WFf5kQpoFLYAS-1920-80.jpg" mos="" align="middle" fullscreen="" width="1280" height="640" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Institute for Molecular Science)</span></figcaption></figure><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:534px;"><p class="vanilla-image-block" style="padding-top:56.18%;"><img id="4HmU6fopBGxQaVbCNqNY7U" name="researchers-set-record.jpeg" alt="The actual size of the quantum computing chip used by the researchers" src="https://cdn.mos.cms.futurecdn.net/4HmU6fopBGxQaVbCNqNY7U-1920-80.jpeg" mos="" align="middle" fullscreen="" width="534" height="300" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: David Awschalom / University of Chicago)</span></figcaption></figure><p>Classiq's software platform will provide high-level quantum algorithm development capabilities, while Intel Quantum SDK will handle compilation, simulation, and hybrid quantum-classical execution. This approach differs from Nvidia's CUDA-Q, which is designed to integrate quantum computing workflows with existing accelerated computing infrastructure. According to Classiq chief executive Nir Minerbi, Classiq operates at a higher level of abstraction and can complement CUDA-Q rather than compete with it. </p><p>"CUDA Q is very much low-level; it is gate-level. It is a good platform, by the way, but it is very low-level," said Minerbi. "The reason companies like Rolls-Royce and BMW and others use the Classiq platform on top of CUDA Q is due to the level of abstraction. It is much, much easier to develop quantum programs with Classiq and then to simulate it with CUDA Q rather than only with CUDA Q." </p><p>Meanwhile, commercial success of quantum computing still depends on substantial advances in quantum hardware, including qubit scalability, error correction, control electronics, and system integration. For now, many Classiq customers are developing applications for quantum computers that have yet to reach the scale required for commercially valuable workloads.</p><h2 id="summary">Summary</h2><p>Intel expects to complete its acquisition of Classiq in the first half of 2027, subject to regulatory approvals and customary closing conditions. The company has not disclosed the purchase price or other financial terms.</p><p>The transaction will bring Classiq's software development capabilities and established customer relationships into Intel's quantum computing organization, but will preserve the platform's compatibility with competing hardware technologies. For Intel, the acquisition represents an opportunity to turn more than a decade of quantum computing research into a commercial business unit as it will have all the building blocks required to create quantum systems. Furthermore, Intel's quantum computing business will be open to third parties in the sense that Intel will be able to build quantum processors for partners using its 18AQ and 18ACC process technologies. </p><p>Whether Intel's quantum computing business eventually delivers the high margins Intel seeks will depend not only on Intel's ability to build systems, produce quantum chips for others, or Classiq's software revenue, but also on the industry's ability to develop quantum processors and systems capable of solving real-world problems that people are willing to pay for.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/quantum-computing/intel-bets-on-classiq-acquisition-to-turn-quantum-efforts-into-commercially-viable-business-deal-positions-company-as-leading-vendor-of-qubit-chips-acquisition-to-close-by-the-first-half-of-2027</link>
                                                                            <description>
                            <![CDATA[ Intel aims to turn its quantum computing research effort into a commercial business by combining its silicon spin-qubit technology and manufacturing capabilities with Classiq’s established software platform and customer base. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">z24ujNhrLngueVacZcZvLg</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/jFuLeDoS3yepBpNT3uuUaN-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Mon, 16 Nov 2026 16:30:00 +0000</pubDate>                                                                                                                                <updated>Tue, 17 Nov 2026 16:41:34 +0000</updated>
                                                                                                                                            <category><![CDATA[Quantum Computing]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                <author><![CDATA[ ashilov@gmail.com (Anton Shilov) ]]></author>                    <dc:creator><![CDATA[ Anton Shilov ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/uMZ5kNphxA2Ut6whdLaSQV-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Anton Shilov has been in the PC industry since 1990s playing games, building PCs, and writing stories about pretty much everything that relates to PCs, Macs, smartphones, tablets, and even fab equipment. Over his career, he has worked at a variety of high-ranking websites, including AnandTech, EE Times, TechRadar, X-bit Labs, and now Tom&#039;s Hardware. He is also a regular features contributor to Tom&#039;s Hardware Premium, writing about the latest developments in the semiconductor industry and related tech news and roadmaps. When Anton is not reading or writing about something high-tech, he is probably watching a good movie, playing a video game, or spending time with his family.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/jFuLeDoS3yepBpNT3uuUaN-1920-80.jpg">
                                                            <media:credit><![CDATA[Microsoft]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Microsoft Quantum materials]]></media:description>                                                            <media:text><![CDATA[Microsoft Quantum materials]]></media:text>
                                <media:title type="plain"><![CDATA[Microsoft Quantum materials]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/jFuLeDoS3yepBpNT3uuUaN-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>After having spent more than a decade pursuing quantum computing research, Intel is making its first real push at commercializing its efforts. First, with the pending acquisition of quantum computing company Classiq, which has an existing list of enterprise customers. Intel says it'll be the only company that can produce qubit chips and cryogenic control electronics at scale; the deal would give it a complete top-to-bottom hardware-to-software stack, and Intel will also gain the chance to become a leading vendor of quantum computing products. Intel's acquisition of Classiq is expected to complete in the first half of 2027. </p><p>The key ingredient of that strategy is the pending <a href="https://www.intel.com/content/www/us/en/newsroom/news/corporate/intel-classiq-accelerate-scalable-quantum-computing.html" target="_blank">acquisition of Classiq</a>, which offers a hardware-agnostic quantum software development platform that enables engineers to design, synthesize, optimize, simulate, and execute quantum algorithms using high-level programming tools and AI-assisted automation, essentially untying software design from the complexity of quantum hardware. Meanwhile, Intel also plans to manufacture quantum processors using specialized versions of its <a href="https://www.tomshardware.com/tech-industry/semiconductors/intels-performance-enhanced-18a-p-process-enters-risk-production-enhanced-node-promises-9-percent-performance-improvement-at-iso-power">18A process </a>technology, offer foundry services, and eventually sell complete quantum computers. Although the company expects its quantum business to become profitable, it acknowledges that commercial success will take time.</p><p>Intel has spent more than a decade developing quantum computing technologies, including silicon spin-qubit processors, cryogenic control electronics, and software for hybrid quantum-classical computing. However, the company has largely treated these activities as research and development projects rather than an actual business. The acquisition of Classiq, which Intel expects to complete in the first half of 2027, marks a shift toward commercialization as Classiq already has many clients and established revenue streams.</p><p>"Up until now […] our quantum computing group for the last ten years was […] purely a research and development project," said Pushkar Ranade, chief technology officer at Intel, during a media briefing. "It was not really viewed as a commercial activity."</p><h2 id="intel-39-s-quantum-hardware-the-journey-so-far">Intel's quantum hardware: The journey so far</h2><p>Intel's quantum computing journey <a href="https://www.intc.com/news-events/press-releases/detail/316/intel-invests-us50-million-to-advance-quantum-computing">kicked off in 2015</a> with a research collaboration with QuTech — the quantum research institute of TU Delft and TNO — to develop quantum computing technologies, including silicon spin qubits. The idea was to leverage Intel's extensive expertise in silicon transistor development and its high-volume semiconductor manufacturing infrastructure to produce quantum processors in the same fabs that make conventional CPUs, overcoming the manufacturing scalability challenges alternative quantum computing architectures faced at the time. Intel subsequently developed <a href="https://www.tomshardware.com/news/intel-details-its-cryogenic-horse-ridge-quantum-control-chip">Horse Ridge</a>, a cryogenic control SoC manufactured using a 22nm process technology in 2020. Three years later, the company introduced <a href="https://www.tomshardware.com/news/intel-announce-tunnel-falls-quantum-research-chip">Tunnel Falls</a>, a 12-qubit silicon spin processor manufactured on a 300mm wafer.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1600px;"><p class="vanilla-image-block" style="padding-top:56.19%;"><img id="Qgqxx54Kpwyo7WwSahsFvC" name="Intel-Quantum-Engineers.jpg" alt="Intel silicon spin qubit progress" src="https://cdn.mos.cms.futurecdn.net/Qgqxx54Kpwyo7WwSahsFvC-1920-80.jpg" mos="" align="middle" fullscreen="" width="1600" height="899" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Intel )</span></figcaption></figure><p>Intel's quantum computing effort focuses on silicon spin qubits, which use spin states of individual electrons to store and manipulate quantum information. The company chose this approach partly because silicon spin qubits can be manufactured using equipment and processes similar to those employed for conventional semiconductor devices. This compatibility enables Intel to re-use most of its existing manufacturing infrastructure and expertise rather than build dedicated semiconductor fabrication facilities for quantum processors that yet have to become commercial products.  </p><p>Despite the <a href="https://www.tomshardware.com/tech-industry/policy/intel-layoffs-to-hit-data-center-group-division-focused-on-server-cpus-ai-chips-and-data-center-architecture-to-be-hit-by-an-unknown-number-of-cuts">recent cuts at Intel</a>, the company sustained its quantum computing project. Intel has developed two specialized versions of its 18A manufacturing technology for quantum computing: 18A-Q for silicon spin-qubit processors and 18A-CC for cryogenic control and readout electronics. To define/build 18A-Q, Intel has collaborated with Hitachi, AIST, and NEDO (three major supporters of quantum computing from Japan), and there appears to be a design kit intended to support large quantum computing systems. Without a doubt, 18A-Q and 18A-CC put Intel on the map of quantum computing chips in a big way. A couple of things should be kept in mind, though.</p><p>"The underlying device that we use, underlying process that we use to make the actual qubit device is a unique process, but it fits into those design rules," Ranade explained. "We can make these chips […] in the same fab that we use to make our CPUs."</p><p>Although Intel 18AQ uses BEOL interconnect geometry and design rules derived from Intel 18A, its underlying qubit devices require a different manufacturing process. According to Ranade, the technology requires a specialized thin-film stack and several additional manufacturing tools. Nevertheless, more than 90% of the manufacturing equipment is identical to that used for conventional semiconductor production, according to Intel. </p><p>"Intel 18A-CC, meanwhile, uses 18A-based geometry but requires additional characterization for operation at cryogenic temperatures," Ranade said. "Crucially, Intel can manufacture quantum processors alongside conventional CPUs in existing semiconductor fabs, which could reduce the capital expenditure required to scale production. […] More than 90% of the process equipment is exactly the same."</p><h2 id="a-major-piece-in-the-quantum-computing-puzzle-classiq">A major piece in the quantum computing puzzle: Classiq</h2><p>For conventional x86 architectures and IA-64 (while it was active), Intel has historically offered compilers that made software run better on its hardware (and sometimes on AMD hardware too). By contrast, with its quantum computing efforts, Intel has generally focused on hardware and while it has its own Quantum SDK that includes a compiler, this is not enough in the quantum computing realm as Classiq's offering operates at a higher abstraction level. With the acquisition of Classiq, the company intends to fill that gap.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2284px;"><p class="vanilla-image-block" style="padding-top:75.00%;"><img id="22SNGmTTWsPq3frQrgvRWe" name="IBM Quantum Loon wafer" alt="An IBM researcher holding a 300mm IBM Quantum Loon wafer" src="https://cdn.mos.cms.futurecdn.net/22SNGmTTWsPq3frQrgvRWe-1920-80.jpg" mos="" align="middle" fullscreen="" width="2284" height="1713" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: IBM)</span></figcaption></figure><p>Classiq develops a hardware-agnostic quantum software platform that applies electronic design automation (EDA) principles to quantum programming, which lets developers design, synthesize, and optimize quantum algorithms without manually constructing circuits from individual quantum gates. The company's software provides high-level modeling, automated algorithm synthesis, optimization, debugging, and execution capabilities across different quantum computing architectures, which somewhat resembles Intel's approach to its x86 compilers, albeit working at a different level.</p><p>Intel already has its Quantum SDK, which includes a C++-based compiler, runtime environment, and simulator for hybrid quantum-classical applications. Classiq's platform will sit above these tools and provide an application-development environment for programmers who lack extensive expertise in quantum computing. </p><p>Intel compares the potential significance of Classiq's software layer to Microsoft's MS-DOS, which helped establish the personal computer software ecosystem; Nvidia's CUDA, which enabled developers to exploit GPUs for general-purpose accelerated computing; and EDA tools, which abstracted high-level design descriptions from the complexities of transistor-level integrated circuit implementation and essentially set the stage for the multi-billion foundry industry today.</p><p>Furthermore, unlike Intel's quantum hardware business, Classiq already has commercial customers and generates revenue. The company employs approximately 130 people and serves more than 50 enterprise customers, including BMW, Comcast, Lufthansa, Rolls-Royce, and Toshiba. It also participates in more than 10 national quantum computing programs and supports over 50 quantum processors and simulators. </p><p>As existing quantum processors remain insufficient for many practical applications, companies can develop and optimize software ahead of hardware availability. As a result, these customers primarily use Classiq products to develop quantum algorithms, train engineers, and create intellectual property in preparation for commercially useful quantum computers. </p><p>Intel's CTO described Classiq's revenue and anticipated growth as healthy, but did not disclose actual financial figures. </p><p>"They have a very healthy revenue this year and a very healthy revenue CAGR over the coming years," Ranade said. "We hope that… we want that to continue." </p><p>It is necessary to note that Intel intends to preserve Classiq's hardware-agnostic approach, which means allowing the company to continue supporting quantum processors and simulators from competing vendors. This is essential to maintain Classiq's existing customer relationships and expand its revenue opportunities, as its clients can/will be able to develop quantum applications without committing to a particular hardware architecture. Obviously, Intel can generate software revenue regardless of which quantum computing technologies ultimately become commercially successful. </p><p>"We very much want to maintain that hardware-agnostic platform that they have developed to attract more and more customers," Ranade said.  </p><h2 id="from-quantum-chips-to-complete-systems">From quantum chips to complete systems</h2><p>When it comes to quantum computing, Intel's ambitions extend beyond manufacturing silicon spin-qubit processors. The company intends to become a supplier of complete quantum computing systems, which would combine quantum processors, cryogenic control electronics, classical computing resources, and software. For now, Intel has most of the building blocks required to develop complete quantum computing systems, but it still has to demonstrate that it can integrate them into commercially viable quantum computing at scale.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:4910px;"><p class="vanilla-image-block" style="padding-top:66.66%;"><img id="fVuF6LJBxPGNZ7j2uMyRk8" name="Intel-Quantum-17-Qubit-2.jpg" alt="Quantum" src="https://cdn.mos.cms.futurecdn.net/fVuF6LJBxPGNZ7j2uMyRk8-1920-80.jpg" mos="" align="middle" fullscreen="" width="4910" height="3273" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Intel)</span></figcaption></figure><p>"We want to be in the business of building quantum foundry components, but also complete quantum systems," said Pushkar Ranade, chief technology officer at Intel.</p><p>Intel is also considering opportunities to provide quantum foundry services to other hardware developers. While its internal processor development will remain focused on silicon spin qubits, the company is open to manufacturing components for alternative quantum computing technologies, which somewhat resembles its approach to its x86 CPU and foundry businesses. </p><p>"We are committed to spin qubits," Ranade said. "But we are also open to providing foundry services to other modalities." </p><p>Intel believes that its capabilities in photonics, advanced packaging, and semiconductor manufacturing could become commercially relevant even to companies developing quantum processors based on different physical principles than silicon qubits. </p><h2 id="can-quantum-computing-become-profitable">Can quantum computing become profitable?</h2><p>Intel's decision to expand its quantum computing activities with the takeover of Classiq comes as chief executive Lip-Bu Tan emphasized profitability, financial discipline, and high-margin products across the company’s established businesses. Given that Intel's quantum computing initiative barely generates any revenue, let alone profits, the CEO stance raises questions about the viability of quantum computing for Intel, especially in a situation where the industry has yet to demonstrate widespread practical applications that justify the cost of large-scale quantum hardware.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2048px;"><p class="vanilla-image-block" style="padding-top:66.70%;"><img id="iWjCt8g7ndNtZ2cWPevjSh" name="devcomp.jpeg" alt="Quantum computer in Devs" src="https://cdn.mos.cms.futurecdn.net/iWjCt8g7ndNtZ2cWPevjSh-1920-80.jpeg" mos="" align="middle" fullscreen="" width="2048" height="1366" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: DNA TV/ FXP)</span></figcaption></figure><p>When asked how Intel’s quantum computing strategy fits Tan's profitability requirements, Ranade explained that the company's immediate financial priorities primarily concern its established client and server processor businesses. Nevertheless, Intel must continue investing in emerging technologies that could shape future computing markets. </p><p>"This is really a bet on the future," Ranade said. "This is something we also have to do in addition to our mainstream businesses today."</p><p>The Intel CTO acknowledged that quantum computing is not set to become a high-margin business in the near future. Meanwhile, the acquisition of Classiq gives Intel an established commercial software operation and revenue stream here and now, as well as an opportunity to expand customer relationships well before its own quantum processors reach commercial maturity. Intel also hopes that combining its hardware and software capabilities (probably including Classiq) will attract customers interested in complete quantum computing solutions.</p><p>When asked whether vertical integration would make Intel's quantum computing operations profitable, Ranade acknowledged that he could not predict gross margins of the quantum computing business because the industry is still in its infancy. Nevertheless, he expressed confidence in the commercial prospects of Intel's quantum hardware and software.</p><p>"I expect it to become a profitable business very quickly on the hardware side, but too early to tell," Ranade said.</p><h2 id="quantum-computing-will-remain-hybrid">Quantum computing will remain hybrid</h2><p>Even though quantum processors evolve at a rapid pace, the technology is unlikely to replace existing CPUs and GPUs for conventional computing any time soon. Instead, Intel expects quantum computing systems to operate alongside the usual processors in data centers and supercomputing environments. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1280px;"><p class="vanilla-image-block" style="padding-top:50.00%;"><img id="aKKUfnAx4WFf5kQpoFLYAS" name="two-gate-quantum-japan.jpg" alt="A diagram of the system." src="https://cdn.mos.cms.futurecdn.net/aKKUfnAx4WFf5kQpoFLYAS-1920-80.jpg" mos="" align="middle" fullscreen="" width="1280" height="640" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Institute for Molecular Science)</span></figcaption></figure><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:534px;"><p class="vanilla-image-block" style="padding-top:56.18%;"><img id="4HmU6fopBGxQaVbCNqNY7U" name="researchers-set-record.jpeg" alt="The actual size of the quantum computing chip used by the researchers" src="https://cdn.mos.cms.futurecdn.net/4HmU6fopBGxQaVbCNqNY7U-1920-80.jpeg" mos="" align="middle" fullscreen="" width="534" height="300" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: David Awschalom / University of Chicago)</span></figcaption></figure><p>Classiq's software platform will provide high-level quantum algorithm development capabilities, while Intel Quantum SDK will handle compilation, simulation, and hybrid quantum-classical execution. This approach differs from Nvidia's CUDA-Q, which is designed to integrate quantum computing workflows with existing accelerated computing infrastructure. According to Classiq chief executive Nir Minerbi, Classiq operates at a higher level of abstraction and can complement CUDA-Q rather than compete with it. </p><p>"CUDA Q is very much low-level; it is gate-level. It is a good platform, by the way, but it is very low-level," said Minerbi. "The reason companies like Rolls-Royce and BMW and others use the Classiq platform on top of CUDA Q is due to the level of abstraction. It is much, much easier to develop quantum programs with Classiq and then to simulate it with CUDA Q rather than only with CUDA Q." </p><p>Meanwhile, commercial success of quantum computing still depends on substantial advances in quantum hardware, including qubit scalability, error correction, control electronics, and system integration. For now, many Classiq customers are developing applications for quantum computers that have yet to reach the scale required for commercially valuable workloads.</p><h2 id="summary">Summary</h2><p>Intel expects to complete its acquisition of Classiq in the first half of 2027, subject to regulatory approvals and customary closing conditions. The company has not disclosed the purchase price or other financial terms.</p><p>The transaction will bring Classiq's software development capabilities and established customer relationships into Intel's quantum computing organization, but will preserve the platform's compatibility with competing hardware technologies. For Intel, the acquisition represents an opportunity to turn more than a decade of quantum computing research into a commercial business unit as it will have all the building blocks required to create quantum systems. Furthermore, Intel's quantum computing business will be open to third parties in the sense that Intel will be able to build quantum processors for partners using its 18AQ and 18ACC process technologies. </p><p>Whether Intel's quantum computing business eventually delivers the high margins Intel seeks will depend not only on Intel's ability to build systems, produce quantum chips for others, or Classiq's software revenue, but also on the industry's ability to develop quantum processors and systems capable of solving real-world problems that people are willing to pay for.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Anthropic says Claude Mythos Preview cut AMD MI355X onboarding from up to 12 months to around three days ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Anthropic told IPO investors that its Claude Mythos Preview model did the software work required to get its AI running on AMD’s GPUs in about three days, with “basically no human intervention,” a job it says used to take six to 12 months. The three days cover prompt processing for one Anthropic model on AMD MI355X GPUs “at a useful level of performance,” with Mythos porting or writing “almost 100% of the kernels.” That kind of low-level code, which ties an AI model to one company’s chips, has long been part of Nvidia’s CUDA moat, and an AI that can write it in days would suggest that moat is eroding.</p><p>Tom Brown, Anthropic’s co-founder and chief compute officer, made the time comparison, noting <a href="https://ipo.anthropic.com/us" target="_blank">to investors </a>that the original meant using a team of “dozens of engineers.” The company already runs its models on multiple kinds of chips — on GPUs from Nvidia and AMD, Google’s TPUs, and AWS Trainium — and its fleet is managed as “one fungible pool of compute.” Brown called the early multi-chip decision “an unconventional bet.” It may have been a wise one, given that Google’s <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/googles-tpus-for-ai-workloads-now-run-in-at-metas-data-centers-tpus-sold-to-external-customers-for-the-first-time-threaten-nvidias-market-dominance">TPUs can now natively run AI models built in PyTorch</a>, and Meta uses them in production.</p><p>Anthropic's claim that the job used to take six to 12 months comes from past work, from its “historical experience bringing up new accelerators using traditional engineering approaches,” where tailoring the code to each hardware platform was a challenge. Removing that onboarding barrier could reduce things to a purchase decision, and Anthropic started working with AMD GPUs for the first time this year, with <a href="https://www.tomshardware.com/tech-industry/amd-to-supply-anthropic-with-2-gigawatts-of-instinct-mi450-gpus">an agreement for up to 2 gigawatts</a> now in place.</p><p>CUDA, Nvidia’s programming platform, is very mature at nearly two decades old. AMD has an equivalent, ROCm, with HIP as its programming layer. Code written for CUDA has to be transformed to work there, and AMD even built a tool called HIPify for the process. There is still a conversion cost to switching, and AMD’s vice president of AI software, Anush Elangovan, told <a href="https://www.eetimes.com/taking-on-cuda-with-rocm-one-step-after-another/" target="_blank"><em>EE Times</em></a> in April that “Claude is better than HIPify because it has web search built in.”</p><p>AMD has been making the argument that AI can write this type of code since August of last year, presenting a research agent that fell short of human experts at the time. In September of that year, AMD CEO Lisa Su said on the <a href="https://www.possible.fm/podcasts/lisa/" target="_blank"><em>Possible</em> podcast</a> that much of AI software is written in CUDA, which needs to be “translated to AMD,” and “a ton of that can now be done in software.”</p><p>Moving forward to July earlier this year, AMD’s senior vice president of AI, Vamsi Boppana, said its engineers were using AI models to generate GPU kernels, and Su said that when she asked whether her team was helping the Anthropic engineer working on the MI355X, she was told they didn’t need any help. Boppana said that in some cases, the resulting kernels were even “better than the most manually tuned versions.” Under its July agreement with Anthropic, AMD is also using Claude to optimize workloads for its Instinct GPUs and to speed up ROCm development.</p><p>At AMD’s<a href="https://www.tomshardware.com/pc-components/gpus/amd-takes-the-wraps-off-its-instinct-mi455x-ai-accelerator-cdna-5-and-helios-rack-scale-architecture-combine-to-take-the-fight-to-nvidia-in-the-data-center"> Advancing AI event in July</a>, OpenAI’s Philippe Tillet said that once OpenAI got its Helios racks, a single engineer confirmed everything worked within a few days, and that “agents are taking over our jobs, in a sense, as kernel engineers.” Before the event, AMD’s Andrew Dieckman, corporate vice president and general manager of the company’s data center GPU business, told reporters that he has had practically zero conversations with AMD’s customers about CUDA.</p><div ><table><thead><tr><th class="firstcol " ><p>Date</p></th><th  ><p>Who</p></th><th  ><p>What they said or showed</p></th><th  ><p>Where</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Aug. 1, 2025</p></td><td  ><p>AMD researchers</p></td><td  ><p>Published GEAK, an AI agent that writes GPU kernels; against 30 expert-written kernels it got 19 right, at 0.92 times the experts’ speed on average</p></td><td  ><p>AMD’s ROCm blog</p></td></tr><tr><td class="firstcol " ><p>Sept. 24, 2025</p></td><td  ><p>Lisa Su, AMD chair and CEO</p></td><td  ><p>Asked about AI coding tools, said much of the code written for CUDA can now be translated to AMD in software: “We’re using software to significantly optimize our customers’ code so that they can run on AMD”</p></td><td  ><p><em>Possible</em> podcast</p></td></tr><tr><td class="firstcol " ><p>April 1, 2026</p></td><td  ><p>Anush Elangovan, AMD vice president of AI software</p></td><td  ><p>Compared Claude with HIPify, AMD’s tool for converting CUDA code: “Claude is better than HIPify because it has web search built in”</p></td><td  ><p><em>EE Times</em> interview</p></td></tr><tr><td class="firstcol " ><p>April 15, 2026</p></td><td  ><p>Jensen Huang, Nvidia CEO</p></td><td  ><p>“We use a ton of AI to create the kernels that we have”; “nobody knows our architecture better than we do”</p></td><td  ><p>Dwarkesh Patel interview</p></td></tr><tr><td class="firstcol " ><p>July 23, 2026</p></td><td  ><p>Andrew Dieckman, AMD corporate vice president and general manager, data center GPU business</p></td><td  ><p>“I have almost zero conversations with our customers about CUDA at this point in time”</p></td><td  ><p><em>PC Gamer</em>, from a pre-event AMD briefing</p></td></tr><tr><td class="firstcol " ><p>July 23, 2026</p></td><td  ><p>Vamsi Boppana, AMD senior vice president of AI</p></td><td  ><p>AI-assisted GPU programming will “remove any remaining barriers to broad adoption”</p></td><td  ><p>AMD Advancing AI keynote</p></td></tr><tr><td class="firstcol " ><p>July 23, 2026</p></td><td  ><p>Philippe Tillet, OpenAI</p></td><td  ><p>“agents are taking over our jobs, in a sense, as kernel engineers”</p></td><td  ><p>AMD Advancing AI keynote</p></td></tr><tr><td class="firstcol " ><p>July 23, 2026</p></td><td  ><p>Tom Brown, Anthropic co-founder and chief compute officer</p></td><td  ><p>One engineer asked Claude to bring up an MI355X rack and left it running “over the weekend”</p></td><td  ><p>AMD Advancing AI keynote</p></td></tr><tr><td class="firstcol " ><p>Nov. 10, 2026</p></td><td  ><p>Tom Brown, Anthropic co-founder and chief compute officer</p></td><td  ><p>Mythos “got Claude running on AMD in just three days with basically no human intervention”; the slide footnote limits the three days to prompt processing for one model</p></td><td  ><p>Anthropic’s IPO roadshow video</p></td></tr></tbody></table></div><p>Until this week, though, all of this came from AMD or its stage, not from Anthropic to its own investors. Anthropic buys the chips rather than selling them, and it’s heading into <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-ipo-filing-shows-most-of-its-compute-goes-to-r-and-d-not-serving-customers-company-details-massive-usd800-billion-in-tpu-and-gpu-contracts">its IPO</a> next week, so its multi-chip approach — now with at least some numbers to back it up — carries more weight than a chipmaker’s claims. Its prospectus says it in writing, too, though without a number: “We can onboard new chip platforms in a fraction of the time it once took.”</p><div ><table><thead><tr><th class="firstcol " ><p>Where Anthropic said it</p></th><th  ><p>Who did the work</p></th><th  ><p>How long</p></th><th  ><p>What it covered</p></th><th  ><p>What it replaced</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>AMD’s Advancing AI keynote, July 23</p></td><td  ><p>One engineer and Claude</p></td><td  ><p>Not stated; left running “over the weekend”</p></td><td  ><p>Its “leading model” on an MI355X rack, with performance climbing</p></td><td  ><p>What it called “a big effort”</p></td></tr><tr><td class="firstcol " ><p>IPO roadshow video, narration, Nov. 10</p></td><td  ><p>Mythos, with “basically no human intervention”</p></td><td  ><p>Three days (“in just three days”)</p></td><td  ><p>Getting “Claude running on AMD”; Claude “ported or wrote almost 100% of the kernels”</p></td><td  ><p>Work that “took dozens of engineers months”</p></td></tr><tr><td class="firstcol " ><p>IPO roadshow video, slide footnote, Nov. 10</p></td><td  ><p>Claude Mythos Preview</p></td><td  ><p>About three days (“~3 day timeline”)</p></td><td  ><p>Bringing up “prompt processing for one Anthropic model on AMD MI355X GPUs at a useful level of performance”</p></td><td  ><p>Six to 12 months (“~6-12-month estimate”), from “historical experience bringing up new accelerators”</p></td></tr><tr><td class="firstcol " ><p>Prospectus (S-1/A No. 2), Nov. 10</p></td><td  ><p>Its engineering teams, “leveraging Claude”</p></td><td  ><p>No figure; “more quickly”</p></td><td  ><p>Bringing “new chip platforms into production”; no chip named</p></td><td  ><p>What it calls “intense engineering effort over many months”</p></td></tr></tbody></table></div><p>The prompt processing (prefill) part of the workload is the step where a model reads and processes what it’s been given before it starts replying, and Anthropic hasn’t said whether the replying step (decode) was included as well. Anthropic’s chart, with its axis labeled “Time to onboard new silicon,” shows both lines reaching the same “~100%” mark without saying 100% of what. The company also hasn’t said how fast the model ran on AMD’s GPUs versus Nvidia’s, and that’s the comparison we want to see to show whether AMD’s chips, running AI-written code, can match Nvidia silicon.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:53.59%;"><img id="qkDmAUP4gEEK88Sc6vyaQV" name="anthropic-roadshow-amd-bringup-chart" alt="A chart from Anthropic’s roadshow video comparing about three days with Mythos to about six to 12 months with a traditional approach to onboard AMD GPUs." src="https://cdn.mos.cms.futurecdn.net/qkDmAUP4gEEK88Sc6vyaQV-1920-80.jpg" mos="" align="middle" fullscreen="" width="2560" height="1372" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Anthropic)</span></figcaption></figure><p>On the other hand, Nvidia CEO Jensen Huang said in an April interview with Dwarkesh Patel that Nvidia already uses AI on its own kernels and knows its chips best — “nobody knows our architecture better than we do” — and can get the most out of them. Asked why labs like Anthropic use other accelerators, he called Anthropic “a unique instance, not a trend.” Nvidia and AMD, among seven companies, are buying $26.0 billion of Anthropic’s stock at the IPO price, so Nvidia is investing in the very company Huang called the exception.</p><p>In late July, <em>SemiAnalysis</em> published an article, “Can AMD break the CUDA Moat?,” which essentially said that software creation is not the same moat it was even a year ago. That said, the catch for AMD is that with agentic coding, “each agent requires GPUs to test their code against,” which means that the AI writing the code still needs the hardware. It has found a constant lack of GPUs inside AMD for this type of development, which has held AMD back and may be a new bottleneck.</p><p>That leaves CUDA’s moat somewhere in between. Getting a model running on a new chip, once the expensive part, appears to be getting cheaper, and better AI may wear down Nvidia’s claim that it knows its hardware best. As for Anthropic’s fungible pool, the chips in it aren’t fully interchangeable — Google’s TPUs are <a href="https://www.tomshardware.com/tech-industry/semiconductors/google-splits-its-tpu-into-two-chips-for-the-first-time-with-training-and-inference-variants">increasingly specialized</a>, for example — and getting the most out of any of them still takes the hardware to test on. The moat is narrowing, but is not gone.</p><p>Anthropic’s first AMD MI450 chips, <a href="https://www.tomshardware.com/pc-components/gpus/amd-takes-the-wraps-off-its-instinct-mi455x-ai-accelerator-cdna-5-and-helios-rack-scale-architecture-combine-to-take-the-fight-to-nvidia-in-the-data-center">in Helios racks</a>, are expected in the first half of 2027. That’s a new generation and an opportunity to see if the three-day bring-up can be repeated. Anthropic’s IPO comes sooner, after which it will report as a public company.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropic-says-claude-mythos-slashed-software-work-to-onboard-amd-gpus-from-up-to-12-months-to-around-three-days-writing-almost-100-percent-of-kernels-autonomously-chips-away-at-nvidias-cuda-moat</link>
                                                                            <description>
                            <![CDATA[ Anthropic says Claude Mythos Preview brought up prompt processing for one model on AMD Instinct MI355X GPUs in about three days. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">ABaxPgN6KrSWsZKG6bL7jm</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/jcG9LwCNh8XUGiSJ9Pxw7Z-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Mon, 16 Nov 2026 15:00:00 +0000</pubDate>                                                                                                                                <updated>Mon, 16 Nov 2026 16:13:44 +0000</updated>
                                                                                                                                            <category><![CDATA[Artificial Intelligence]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/jcG9LwCNh8XUGiSJ9Pxw7Z-1920-80.jpg">
                                                            <media:credit><![CDATA[AMD]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[An AMD Instinct MI350 Series accelerator module seen from above, with its chip package exposed at the center of a black circuit board.]]></media:description>                                                            <media:text><![CDATA[An AMD Instinct MI350 Series accelerator module seen from above, with its chip package exposed at the center of a black circuit board.]]></media:text>
                                <media:title type="plain"><![CDATA[An AMD Instinct MI350 Series accelerator module seen from above, with its chip package exposed at the center of a black circuit board.]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/jcG9LwCNh8XUGiSJ9Pxw7Z-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Anthropic told IPO investors that its Claude Mythos Preview model did the software work required to get its AI running on AMD’s GPUs in about three days, with “basically no human intervention,” a job it says used to take six to 12 months. The three days cover prompt processing for one Anthropic model on AMD MI355X GPUs “at a useful level of performance,” with Mythos porting or writing “almost 100% of the kernels.” That kind of low-level code, which ties an AI model to one company’s chips, has long been part of Nvidia’s CUDA moat, and an AI that can write it in days would suggest that moat is eroding.</p><p>Tom Brown, Anthropic’s co-founder and chief compute officer, made the time comparison, noting <a href="https://ipo.anthropic.com/us" target="_blank">to investors </a>that the original meant using a team of “dozens of engineers.” The company already runs its models on multiple kinds of chips — on GPUs from Nvidia and AMD, Google’s TPUs, and AWS Trainium — and its fleet is managed as “one fungible pool of compute.” Brown called the early multi-chip decision “an unconventional bet.” It may have been a wise one, given that Google’s <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/googles-tpus-for-ai-workloads-now-run-in-at-metas-data-centers-tpus-sold-to-external-customers-for-the-first-time-threaten-nvidias-market-dominance">TPUs can now natively run AI models built in PyTorch</a>, and Meta uses them in production.</p><p>Anthropic's claim that the job used to take six to 12 months comes from past work, from its “historical experience bringing up new accelerators using traditional engineering approaches,” where tailoring the code to each hardware platform was a challenge. Removing that onboarding barrier could reduce things to a purchase decision, and Anthropic started working with AMD GPUs for the first time this year, with <a href="https://www.tomshardware.com/tech-industry/amd-to-supply-anthropic-with-2-gigawatts-of-instinct-mi450-gpus">an agreement for up to 2 gigawatts</a> now in place.</p><p>CUDA, Nvidia’s programming platform, is very mature at nearly two decades old. AMD has an equivalent, ROCm, with HIP as its programming layer. Code written for CUDA has to be transformed to work there, and AMD even built a tool called HIPify for the process. There is still a conversion cost to switching, and AMD’s vice president of AI software, Anush Elangovan, told <a href="https://www.eetimes.com/taking-on-cuda-with-rocm-one-step-after-another/" target="_blank"><em>EE Times</em></a> in April that “Claude is better than HIPify because it has web search built in.”</p><p>AMD has been making the argument that AI can write this type of code since August of last year, presenting a research agent that fell short of human experts at the time. In September of that year, AMD CEO Lisa Su said on the <a href="https://www.possible.fm/podcasts/lisa/" target="_blank"><em>Possible</em> podcast</a> that much of AI software is written in CUDA, which needs to be “translated to AMD,” and “a ton of that can now be done in software.”</p><p>Moving forward to July earlier this year, AMD’s senior vice president of AI, Vamsi Boppana, said its engineers were using AI models to generate GPU kernels, and Su said that when she asked whether her team was helping the Anthropic engineer working on the MI355X, she was told they didn’t need any help. Boppana said that in some cases, the resulting kernels were even “better than the most manually tuned versions.” Under its July agreement with Anthropic, AMD is also using Claude to optimize workloads for its Instinct GPUs and to speed up ROCm development.</p><p>At AMD’s<a href="https://www.tomshardware.com/pc-components/gpus/amd-takes-the-wraps-off-its-instinct-mi455x-ai-accelerator-cdna-5-and-helios-rack-scale-architecture-combine-to-take-the-fight-to-nvidia-in-the-data-center"> Advancing AI event in July</a>, OpenAI’s Philippe Tillet said that once OpenAI got its Helios racks, a single engineer confirmed everything worked within a few days, and that “agents are taking over our jobs, in a sense, as kernel engineers.” Before the event, AMD’s Andrew Dieckman, corporate vice president and general manager of the company’s data center GPU business, told reporters that he has had practically zero conversations with AMD’s customers about CUDA.</p><div ><table><thead><tr><th class="firstcol " ><p>Date</p></th><th  ><p>Who</p></th><th  ><p>What they said or showed</p></th><th  ><p>Where</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Aug. 1, 2025</p></td><td  ><p>AMD researchers</p></td><td  ><p>Published GEAK, an AI agent that writes GPU kernels; against 30 expert-written kernels it got 19 right, at 0.92 times the experts’ speed on average</p></td><td  ><p>AMD’s ROCm blog</p></td></tr><tr><td class="firstcol " ><p>Sept. 24, 2025</p></td><td  ><p>Lisa Su, AMD chair and CEO</p></td><td  ><p>Asked about AI coding tools, said much of the code written for CUDA can now be translated to AMD in software: “We’re using software to significantly optimize our customers’ code so that they can run on AMD”</p></td><td  ><p><em>Possible</em> podcast</p></td></tr><tr><td class="firstcol " ><p>April 1, 2026</p></td><td  ><p>Anush Elangovan, AMD vice president of AI software</p></td><td  ><p>Compared Claude with HIPify, AMD’s tool for converting CUDA code: “Claude is better than HIPify because it has web search built in”</p></td><td  ><p><em>EE Times</em> interview</p></td></tr><tr><td class="firstcol " ><p>April 15, 2026</p></td><td  ><p>Jensen Huang, Nvidia CEO</p></td><td  ><p>“We use a ton of AI to create the kernels that we have”; “nobody knows our architecture better than we do”</p></td><td  ><p>Dwarkesh Patel interview</p></td></tr><tr><td class="firstcol " ><p>July 23, 2026</p></td><td  ><p>Andrew Dieckman, AMD corporate vice president and general manager, data center GPU business</p></td><td  ><p>“I have almost zero conversations with our customers about CUDA at this point in time”</p></td><td  ><p><em>PC Gamer</em>, from a pre-event AMD briefing</p></td></tr><tr><td class="firstcol " ><p>July 23, 2026</p></td><td  ><p>Vamsi Boppana, AMD senior vice president of AI</p></td><td  ><p>AI-assisted GPU programming will “remove any remaining barriers to broad adoption”</p></td><td  ><p>AMD Advancing AI keynote</p></td></tr><tr><td class="firstcol " ><p>July 23, 2026</p></td><td  ><p>Philippe Tillet, OpenAI</p></td><td  ><p>“agents are taking over our jobs, in a sense, as kernel engineers”</p></td><td  ><p>AMD Advancing AI keynote</p></td></tr><tr><td class="firstcol " ><p>July 23, 2026</p></td><td  ><p>Tom Brown, Anthropic co-founder and chief compute officer</p></td><td  ><p>One engineer asked Claude to bring up an MI355X rack and left it running “over the weekend”</p></td><td  ><p>AMD Advancing AI keynote</p></td></tr><tr><td class="firstcol " ><p>Nov. 10, 2026</p></td><td  ><p>Tom Brown, Anthropic co-founder and chief compute officer</p></td><td  ><p>Mythos “got Claude running on AMD in just three days with basically no human intervention”; the slide footnote limits the three days to prompt processing for one model</p></td><td  ><p>Anthropic’s IPO roadshow video</p></td></tr></tbody></table></div><p>Until this week, though, all of this came from AMD or its stage, not from Anthropic to its own investors. Anthropic buys the chips rather than selling them, and it’s heading into <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-ipo-filing-shows-most-of-its-compute-goes-to-r-and-d-not-serving-customers-company-details-massive-usd800-billion-in-tpu-and-gpu-contracts">its IPO</a> next week, so its multi-chip approach — now with at least some numbers to back it up — carries more weight than a chipmaker’s claims. Its prospectus says it in writing, too, though without a number: “We can onboard new chip platforms in a fraction of the time it once took.”</p><div ><table><thead><tr><th class="firstcol " ><p>Where Anthropic said it</p></th><th  ><p>Who did the work</p></th><th  ><p>How long</p></th><th  ><p>What it covered</p></th><th  ><p>What it replaced</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>AMD’s Advancing AI keynote, July 23</p></td><td  ><p>One engineer and Claude</p></td><td  ><p>Not stated; left running “over the weekend”</p></td><td  ><p>Its “leading model” on an MI355X rack, with performance climbing</p></td><td  ><p>What it called “a big effort”</p></td></tr><tr><td class="firstcol " ><p>IPO roadshow video, narration, Nov. 10</p></td><td  ><p>Mythos, with “basically no human intervention”</p></td><td  ><p>Three days (“in just three days”)</p></td><td  ><p>Getting “Claude running on AMD”; Claude “ported or wrote almost 100% of the kernels”</p></td><td  ><p>Work that “took dozens of engineers months”</p></td></tr><tr><td class="firstcol " ><p>IPO roadshow video, slide footnote, Nov. 10</p></td><td  ><p>Claude Mythos Preview</p></td><td  ><p>About three days (“~3 day timeline”)</p></td><td  ><p>Bringing up “prompt processing for one Anthropic model on AMD MI355X GPUs at a useful level of performance”</p></td><td  ><p>Six to 12 months (“~6-12-month estimate”), from “historical experience bringing up new accelerators”</p></td></tr><tr><td class="firstcol " ><p>Prospectus (S-1/A No. 2), Nov. 10</p></td><td  ><p>Its engineering teams, “leveraging Claude”</p></td><td  ><p>No figure; “more quickly”</p></td><td  ><p>Bringing “new chip platforms into production”; no chip named</p></td><td  ><p>What it calls “intense engineering effort over many months”</p></td></tr></tbody></table></div><p>The prompt processing (prefill) part of the workload is the step where a model reads and processes what it’s been given before it starts replying, and Anthropic hasn’t said whether the replying step (decode) was included as well. Anthropic’s chart, with its axis labeled “Time to onboard new silicon,” shows both lines reaching the same “~100%” mark without saying 100% of what. The company also hasn’t said how fast the model ran on AMD’s GPUs versus Nvidia’s, and that’s the comparison we want to see to show whether AMD’s chips, running AI-written code, can match Nvidia silicon.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:53.59%;"><img id="qkDmAUP4gEEK88Sc6vyaQV" name="anthropic-roadshow-amd-bringup-chart" alt="A chart from Anthropic’s roadshow video comparing about three days with Mythos to about six to 12 months with a traditional approach to onboard AMD GPUs." src="https://cdn.mos.cms.futurecdn.net/qkDmAUP4gEEK88Sc6vyaQV-1920-80.jpg" mos="" align="middle" fullscreen="" width="2560" height="1372" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Anthropic)</span></figcaption></figure><p>On the other hand, Nvidia CEO Jensen Huang said in an April interview with Dwarkesh Patel that Nvidia already uses AI on its own kernels and knows its chips best — “nobody knows our architecture better than we do” — and can get the most out of them. Asked why labs like Anthropic use other accelerators, he called Anthropic “a unique instance, not a trend.” Nvidia and AMD, among seven companies, are buying $26.0 billion of Anthropic’s stock at the IPO price, so Nvidia is investing in the very company Huang called the exception.</p><p>In late July, <em>SemiAnalysis</em> published an article, “Can AMD break the CUDA Moat?,” which essentially said that software creation is not the same moat it was even a year ago. That said, the catch for AMD is that with agentic coding, “each agent requires GPUs to test their code against,” which means that the AI writing the code still needs the hardware. It has found a constant lack of GPUs inside AMD for this type of development, which has held AMD back and may be a new bottleneck.</p><p>That leaves CUDA’s moat somewhere in between. Getting a model running on a new chip, once the expensive part, appears to be getting cheaper, and better AI may wear down Nvidia’s claim that it knows its hardware best. As for Anthropic’s fungible pool, the chips in it aren’t fully interchangeable — Google’s TPUs are <a href="https://www.tomshardware.com/tech-industry/semiconductors/google-splits-its-tpu-into-two-chips-for-the-first-time-with-training-and-inference-variants">increasingly specialized</a>, for example — and getting the most out of any of them still takes the hardware to test on. The moat is narrowing, but is not gone.</p><p>Anthropic’s first AMD MI450 chips, <a href="https://www.tomshardware.com/pc-components/gpus/amd-takes-the-wraps-off-its-instinct-mi455x-ai-accelerator-cdna-5-and-helios-rack-scale-architecture-combine-to-take-the-fight-to-nvidia-in-the-data-center">in Helios racks</a>, are expected in the first half of 2027. That’s a new generation and an opportunity to see if the three-day bring-up can be repeated. Anthropic’s IPO comes sooner, after which it will report as a public company.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ How Shanghai fund manager Kang Wei introduced his wife Alice Huang to Jensen and built $7.5 billion Nvidia AI GPU smuggling operation ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Jensen Huang had a stunning fall from grace — from being one of the richest people on Earth and one of the primary drivers of the AI boom to being implicated in <a href="https://www.tomshardware.com/tech-industry/jensen-huang-now-accused-of-affair-with-nvidia-executive-who-reported-to-him-alleged-cover-up-attempt-new-indiscretion-comes-in-wake-of-admitted-affair-with-alleged-chinese-spy">two extramarital affair scandals</a> and <a href="https://www.tomshardware.com/tech-industry/nvidia-confirms-jensen-huang-has-resigned-as-ceo-colette-kress-executive-vice-president-and-chief-financial-officer-appointed-interim-ceo">resigning from Nvidia</a> in under two weeks. While many are focused on the former Nvidia chief’s relationship with Alice Huang, who was <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/us-intelligence-believes-jensen-huang-had-romantic-relationship-with-chinese-spy-in-honeypot-operation-spy-suspect-linked-to-company-accused-of-smuggling-usd2-billion-in-ai-chips">allegedly passing on information to Chinese intelligence</a>, the <em>Wall Street Journal</em> dug deeper into the person who introduced her to Jensen — Kang Wei, a star fund manager in Shanghai who also happened to be Ms. Huang’s husband.</p><p>There have been reports that Kang Wei met Jensen Huang in the 2010s, when his investment firm was heavily involved in Chinese gaming companies. It’s not clear if this introduction predated Kang’s relationship and marriage to Alice, but sources close to Ms. Huang said that she was able to get close to Jensen by the 2020s. Despite being a couple, the <em>WSJ</em> reported that Kang seemingly accepted Alice’s relationship with Jensen, with people close to the couple saying that he would often joke that she was a personal assistant to the billionaire.</p><p>Kang Wei is not just an average business executive, though. He made a name for himself in the finance world as a strategist for Shenyin Wanguo before taking over the Zhonghai Fund in 2007. This made him one of the youngest mutual fund chiefs in China at that time, before finally setting out on his own and founding Shanghai Hexi in 2010. </p><p>It was also around this time that Alice Huang, whose Chinese name is Huang Le, entered the financial industry, first working at HSBC’s private banking arm in Hong Kong, before moving to CITIC Securities, where she attained a vice president position. It was during her time at CITIC that Alice crossed paths with Kang. Shanghai Hexi signed a deal with the securities firm in 2014 and 2015, and by 2016, corporate registry records show that Ms. Huang had joined the investment firm.</p><p>The <a href="https://economictimes.indiatimes.com/news/international/us/who-is-kang-wei-the-chinese-fund-executive-reported-to-be-alice-huangs-husband-and-now-under-police-investigation-in-shanghai/articleshow/135740485.cms"><em>Economic Times</em></a><em> </em>reports Alice Huang’s move into Shanghai Hexi resulted in her getting a 25% stake in the company, all of which came from Kang. The remaining shares were shared between Kang Wei, with 48.75%, and his business partner, Zhu Xiaoming, with 26.25%, with Alice Huang’s 25% ownership stake being held through Shanghai Yuxi Investment Management Consulting Co., Ltd. Meanwhile, Kang also started introducing Alice Huang as his wife to his business associates and investors. </p><p>One year later, Zhu and Kang had a falling out when one of their investments under a Hexi fund resulted in a major windfall, with the two partners disagreeing on how the profits were to be shared. Nevertheless, the massive profit generated by the deal allowed Shanghai Hexi to purchase Hangzhou Shengfeng, a video game company, for around 500 million yuan from Chinese conglomerate Century Huatong. </p><p>Two years after the transaction, in 2019, Hangzhou Shengfeng’s ownership was transferred to an offshore entity, which was eventually consolidated under Osmanthus Vale Holdings — a holding company incorporated in the British Virgin Islands and listed Huang Le, Alice Huang’s Chinese name, as its sole owner. It was then purchased by 7Road Holdings less than two months after it was founded. While Ms. Huang was the signatory in this deal, it’s believed that she served as a “white glove,” a person who operates on behalf of another to conceal their identity — but it remains unclear if this was for Shanghai Hexi and Kang Wei or some other third party.</p><p>There is not much information on Alice Huang’s activities between 2019 and 2021. However, Shanghai Hexi invested in Shanghai Longjing Information Technology, a data center-focused subsidiary of Century Huatong in 2020, where Ms. Huang served as a director between 2021 and 2022, according to a report by <a href="https://www.thewirechina.com/2026/11/11/the-rise-and-rise-of-alice-huang/"><em>The Wire China</em></a>. At about the same time, 7Road Holdings expanded into cloud computing, just before ChatGPT brought AI LLMs to the forefront. The U.S. government’s ban on the export of Nvidia chips to China put this investment in jeopardy, though, forcing it to spin off the business as 7Road International in Singapore.</p><p>We don’t know when 7Road Holdings sold its Singapore subsidiary, but corporate records show that by the time the company was renamed Megaspeed in March 2023, it was already under the ownership of Huang Le Limited and led by Ms. Huang. It was sold again in November 2023 to Swiftdata, a company owned by Singaporean Tan Yong Pong. <a href="https://www.axios.com/2026/11/09/alice-huang-nvidia-jensen-megaspeed"><em>Axios</em></a> said that she left the firm as a managing director soon after this acquisition, but sources said that she kept working with the company afterward, with Kang Wei helping her get it off the ground.</p><p>In the same year, Shanghai Century Litong was founded to build AI data centers in Shanghai and Shenzhen, with Jiangsu Lettall Electronic Co., an electronic components manufacturer, holding a 90% stake, while the remaining 10% is owned by Century Longteng, as reported by <a href="https://finance.biggo.com/news/ac2bef77-1a31-4236-af0b-8cb60cb6a38a"><em>BigGo Finance</em></a>. It should be noted that Century Longteng was renamed from Shanghai Yuxi Investment Management, with the firm pivoting its business from finance to cloud computing, and that even though Alice Huang has already exited the company by this time, it’s under the control of Huang Yu, who some sources say is Ms. Huang’s sister.</p><p>A year after these data center acquisitions, Alice Huang began to be spotted in public with the former Nvidia CEO. A <a href="https://www.linkedin.com/posts/computex-ai-ugcPost-7204717553166798849-PpUx/?rcm=ACoAABpx3FcBKAV76kUVefLtYTzSNLBlXQ3BPNk">LinkedIn post</a> showed her having drinks with a leather-clad Jensen Huang alongside other tech executives from NetThunder and Gigabyte at a post-Computex 2024 event. She was once again seen having dinner with him in a Taipei restaurant as reporters tried to interview him while one kid demanded his attention off camera. </p><p>But more than these public sightings, Alice herself boasted about her “closeness” to the then CEO. People close to her said that she regularly posted photos with Jensen on her WeChat account, and even shared others with potential customers to help establish her credibility. She even reportedly claimed that Mr. Huang would celebrate her birthdays with restaurant dinners and gifts, with one Nvidia China sales employee also saying that she was a familiar face at the company’s annual event for China partners.</p><p>An investigation revealed that from 2023 to 2025, Megaspeed imported $4.6 billion worth of Nvidia AI chips, while <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/singapore-company-alleged-to-have-helped-china-get-usd2-billion-worth-of-nvidia-ai-processors-report-claims-nvidia-denies-that-the-accused-has-any-china-ties-but-a-u-s-investigation-is-underway">another $2 billion was routed through Speedmatrix</a>, Megaspeed’s subsidiary with offices in Johor, Malaysia. Shanghai Century Litong also reportedly bought AI GPUs from other sources, with estimates putting the total number of banned chips moved through Alice Huang’s assistance at around 240,000, putting the <a href="https://www.tomshardware.com/tech-industry/policy/alice-huang-reportedly-instrumental-in-securing-over-usd7-5-billion-of-nvidia-ai-gpu-imports-during-three-year-relationship-with-jensen-huang-helped-move-240-000-banned-chips-into-china-mostly-h100-and-h200">shipments’ worth at around $7.5 billion.</a></p><p>But even as Ms. Huang was making her international moves, Kang Wei was facing trouble at home. Chinese regulators fined him 2.5 million yuan, or approximately $380,000, in 2025, for violating sections of China’s Securities Law. This also led to a warning letter against him, issued in August 2026 by the China Securities Regulatory Commission. Despite running into trouble with the authorities, this massive paper trail showed how Kang Wei actively made moves that put Alice Huang close to Jensen Huang. </p><p>Some questions remain, though, about the U.S. intelligence community’s claims that she worked with the Ministry of State Security (MSS), China’s primary intelligence and counterintelligence agency. If it’s true that Alice Huang worked with the MSS and passed on information to the agency, did she specifically target Jensen Huang, or did the spy agency simply discover their indiscretion and use that to force her to cooperate? Was all this planned by the MSS, or were they just shrewd business moves that allowed Kang to make more money while going under the radar of regulators? These are questions that only Kang Wei and Alice Huang (or the Chinese spy agency) can answer. But no matter the case, the fact remains that China acquired advanced AI chips that were supposed to be out of its reach because of this couple’s connections.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/how-shanghai-fund-manager-kang-wei-introduced-his-wife-alice-huang-to-jensen-and-built-usd7-5-billion-nvidia-ai-gpu-smuggling-operation-mistresss-husband-joked-she-was-the-ceos-personal-assistant</link>
                                                                            <description>
                            <![CDATA[ While Alice Huang lives on the headlines, it was actually her husband who put her in the same room as Jensen Huang. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">ec73eTgZXxwMCJrb3TgQTo</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/j4BGW2BkHVr8ACoUoyotye-1920-80.png" type="image/png" length="0"></enclosure>
                                                                        <pubDate>Fri, 13 Nov 2026 11:00:00 +0000</pubDate>                                                                                                                                <updated>Fri, 13 Nov 2026 13:28:17 +0000</updated>
                                                                                                                                            <category><![CDATA[Tech Industry]]></category>
                                                                                                <author><![CDATA[ editors@tomshardware.com (Jowi Morales) ]]></author>                    <dc:creator><![CDATA[ Jowi Morales ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/gM7E2WSDg2wgCFoaDPz9yK-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Jowi Morales is a writer and journalist covering the tech beat since 2021. However, he’s been interested in technology far earlier than that. He started discovering desktop computers when his father brought home a Windows 95 PC, but his first real experience working under the hood of the PC was when the old computer’s hard drive was filled to the brim in the year 2000. He deleted the Windows folder to attempt to rectify the situation, which led to his dad buying a new desktop PC. Since then, he learned a lot more about computers, and he’s always been the go-to tech expert for his family and friends.&lt;/p&gt;&lt;p&gt;Jowi primarily uses a Windows workstation and an Android phone, but he also bought into the Apple ecosystem with the 6th-gen iPad, iPhone 14 Pro Max, and the M1 MacBook Air. Today, Jowi covers hardware and software from Redmond and Cupertino, while also looking at the tech industry in general.&lt;/p&gt;&lt;p&gt;Aside from covering technology, Jowi is an avid photographer and writes about automobiles, aviation, and tanks. You can find his bylines at &lt;a href=&quot;https://www.makeuseof.com/author/jowi-morales/&quot;&gt;MakeUseOf&lt;/a&gt;, &lt;a href=&quot;https://www.slashgear.com/author/jowimorales/&quot;&gt;SlashGear&lt;/a&gt;, and, of course, &lt;a href=&quot;https://www.tomshardware.com/author/jowi-morales&quot;&gt;Tom’s Hardware&lt;/a&gt;.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/png" url="https://cdn.mos.cms.futurecdn.net/j4BGW2BkHVr8ACoUoyotye-1920-80.png">
                                                            <media:credit><![CDATA[Getty Images]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[the Shanghai skyline ]]></media:description>                                                            <media:text><![CDATA[the Shanghai skyline ]]></media:text>
                                <media:title type="plain"><![CDATA[the Shanghai skyline ]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/j4BGW2BkHVr8ACoUoyotye-1920-80.png" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Jensen Huang had a stunning fall from grace — from being one of the richest people on Earth and one of the primary drivers of the AI boom to being implicated in <a href="https://www.tomshardware.com/tech-industry/jensen-huang-now-accused-of-affair-with-nvidia-executive-who-reported-to-him-alleged-cover-up-attempt-new-indiscretion-comes-in-wake-of-admitted-affair-with-alleged-chinese-spy">two extramarital affair scandals</a> and <a href="https://www.tomshardware.com/tech-industry/nvidia-confirms-jensen-huang-has-resigned-as-ceo-colette-kress-executive-vice-president-and-chief-financial-officer-appointed-interim-ceo">resigning from Nvidia</a> in under two weeks. While many are focused on the former Nvidia chief’s relationship with Alice Huang, who was <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/us-intelligence-believes-jensen-huang-had-romantic-relationship-with-chinese-spy-in-honeypot-operation-spy-suspect-linked-to-company-accused-of-smuggling-usd2-billion-in-ai-chips">allegedly passing on information to Chinese intelligence</a>, the <em>Wall Street Journal</em> dug deeper into the person who introduced her to Jensen — Kang Wei, a star fund manager in Shanghai who also happened to be Ms. Huang’s husband.</p><p>There have been reports that Kang Wei met Jensen Huang in the 2010s, when his investment firm was heavily involved in Chinese gaming companies. It’s not clear if this introduction predated Kang’s relationship and marriage to Alice, but sources close to Ms. Huang said that she was able to get close to Jensen by the 2020s. Despite being a couple, the <em>WSJ</em> reported that Kang seemingly accepted Alice’s relationship with Jensen, with people close to the couple saying that he would often joke that she was a personal assistant to the billionaire.</p><p>Kang Wei is not just an average business executive, though. He made a name for himself in the finance world as a strategist for Shenyin Wanguo before taking over the Zhonghai Fund in 2007. This made him one of the youngest mutual fund chiefs in China at that time, before finally setting out on his own and founding Shanghai Hexi in 2010. </p><p>It was also around this time that Alice Huang, whose Chinese name is Huang Le, entered the financial industry, first working at HSBC’s private banking arm in Hong Kong, before moving to CITIC Securities, where she attained a vice president position. It was during her time at CITIC that Alice crossed paths with Kang. Shanghai Hexi signed a deal with the securities firm in 2014 and 2015, and by 2016, corporate registry records show that Ms. Huang had joined the investment firm.</p><p>The <a href="https://economictimes.indiatimes.com/news/international/us/who-is-kang-wei-the-chinese-fund-executive-reported-to-be-alice-huangs-husband-and-now-under-police-investigation-in-shanghai/articleshow/135740485.cms"><em>Economic Times</em></a><em> </em>reports Alice Huang’s move into Shanghai Hexi resulted in her getting a 25% stake in the company, all of which came from Kang. The remaining shares were shared between Kang Wei, with 48.75%, and his business partner, Zhu Xiaoming, with 26.25%, with Alice Huang’s 25% ownership stake being held through Shanghai Yuxi Investment Management Consulting Co., Ltd. Meanwhile, Kang also started introducing Alice Huang as his wife to his business associates and investors. </p><p>One year later, Zhu and Kang had a falling out when one of their investments under a Hexi fund resulted in a major windfall, with the two partners disagreeing on how the profits were to be shared. Nevertheless, the massive profit generated by the deal allowed Shanghai Hexi to purchase Hangzhou Shengfeng, a video game company, for around 500 million yuan from Chinese conglomerate Century Huatong. </p><p>Two years after the transaction, in 2019, Hangzhou Shengfeng’s ownership was transferred to an offshore entity, which was eventually consolidated under Osmanthus Vale Holdings — a holding company incorporated in the British Virgin Islands and listed Huang Le, Alice Huang’s Chinese name, as its sole owner. It was then purchased by 7Road Holdings less than two months after it was founded. While Ms. Huang was the signatory in this deal, it’s believed that she served as a “white glove,” a person who operates on behalf of another to conceal their identity — but it remains unclear if this was for Shanghai Hexi and Kang Wei or some other third party.</p><p>There is not much information on Alice Huang’s activities between 2019 and 2021. However, Shanghai Hexi invested in Shanghai Longjing Information Technology, a data center-focused subsidiary of Century Huatong in 2020, where Ms. Huang served as a director between 2021 and 2022, according to a report by <a href="https://www.thewirechina.com/2026/11/11/the-rise-and-rise-of-alice-huang/"><em>The Wire China</em></a>. At about the same time, 7Road Holdings expanded into cloud computing, just before ChatGPT brought AI LLMs to the forefront. The U.S. government’s ban on the export of Nvidia chips to China put this investment in jeopardy, though, forcing it to spin off the business as 7Road International in Singapore.</p><p>We don’t know when 7Road Holdings sold its Singapore subsidiary, but corporate records show that by the time the company was renamed Megaspeed in March 2023, it was already under the ownership of Huang Le Limited and led by Ms. Huang. It was sold again in November 2023 to Swiftdata, a company owned by Singaporean Tan Yong Pong. <a href="https://www.axios.com/2026/11/09/alice-huang-nvidia-jensen-megaspeed"><em>Axios</em></a> said that she left the firm as a managing director soon after this acquisition, but sources said that she kept working with the company afterward, with Kang Wei helping her get it off the ground.</p><p>In the same year, Shanghai Century Litong was founded to build AI data centers in Shanghai and Shenzhen, with Jiangsu Lettall Electronic Co., an electronic components manufacturer, holding a 90% stake, while the remaining 10% is owned by Century Longteng, as reported by <a href="https://finance.biggo.com/news/ac2bef77-1a31-4236-af0b-8cb60cb6a38a"><em>BigGo Finance</em></a>. It should be noted that Century Longteng was renamed from Shanghai Yuxi Investment Management, with the firm pivoting its business from finance to cloud computing, and that even though Alice Huang has already exited the company by this time, it’s under the control of Huang Yu, who some sources say is Ms. Huang’s sister.</p><p>A year after these data center acquisitions, Alice Huang began to be spotted in public with the former Nvidia CEO. A <a href="https://www.linkedin.com/posts/computex-ai-ugcPost-7204717553166798849-PpUx/?rcm=ACoAABpx3FcBKAV76kUVefLtYTzSNLBlXQ3BPNk">LinkedIn post</a> showed her having drinks with a leather-clad Jensen Huang alongside other tech executives from NetThunder and Gigabyte at a post-Computex 2024 event. She was once again seen having dinner with him in a Taipei restaurant as reporters tried to interview him while one kid demanded his attention off camera. </p><p>But more than these public sightings, Alice herself boasted about her “closeness” to the then CEO. People close to her said that she regularly posted photos with Jensen on her WeChat account, and even shared others with potential customers to help establish her credibility. She even reportedly claimed that Mr. Huang would celebrate her birthdays with restaurant dinners and gifts, with one Nvidia China sales employee also saying that she was a familiar face at the company’s annual event for China partners.</p><p>An investigation revealed that from 2023 to 2025, Megaspeed imported $4.6 billion worth of Nvidia AI chips, while <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/singapore-company-alleged-to-have-helped-china-get-usd2-billion-worth-of-nvidia-ai-processors-report-claims-nvidia-denies-that-the-accused-has-any-china-ties-but-a-u-s-investigation-is-underway">another $2 billion was routed through Speedmatrix</a>, Megaspeed’s subsidiary with offices in Johor, Malaysia. Shanghai Century Litong also reportedly bought AI GPUs from other sources, with estimates putting the total number of banned chips moved through Alice Huang’s assistance at around 240,000, putting the <a href="https://www.tomshardware.com/tech-industry/policy/alice-huang-reportedly-instrumental-in-securing-over-usd7-5-billion-of-nvidia-ai-gpu-imports-during-three-year-relationship-with-jensen-huang-helped-move-240-000-banned-chips-into-china-mostly-h100-and-h200">shipments’ worth at around $7.5 billion.</a></p><p>But even as Ms. Huang was making her international moves, Kang Wei was facing trouble at home. Chinese regulators fined him 2.5 million yuan, or approximately $380,000, in 2025, for violating sections of China’s Securities Law. This also led to a warning letter against him, issued in August 2026 by the China Securities Regulatory Commission. Despite running into trouble with the authorities, this massive paper trail showed how Kang Wei actively made moves that put Alice Huang close to Jensen Huang. </p><p>Some questions remain, though, about the U.S. intelligence community’s claims that she worked with the Ministry of State Security (MSS), China’s primary intelligence and counterintelligence agency. If it’s true that Alice Huang worked with the MSS and passed on information to the agency, did she specifically target Jensen Huang, or did the spy agency simply discover their indiscretion and use that to force her to cooperate? Was all this planned by the MSS, or were they just shrewd business moves that allowed Kang to make more money while going under the radar of regulators? These are questions that only Kang Wei and Alice Huang (or the Chinese spy agency) can answer. But no matter the case, the fact remains that China acquired advanced AI chips that were supposed to be out of its reach because of this couple’s connections.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Google reworks TPU silicon into specialized systems for agents and security ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Google's eighth-generation TPU family has now expanded further. Back in April, the company announced TPU 8t for training AI models and TPU 8i for general-purpose inference workloads. Google has now introduced TPU 8tx, built on the same silicon as the previously announced 8t, for confidential, hardware-isolated inference aimed at regulated enterprises and governments, and TPU 8i Max, which combines two 8i chips on one package for long-context agentic work. </p><p>TPU 8tx swaps the Arm-based Axion hosts of the 8t and 8i for a 5th Gen AMD EPYC CPU, adopting a Boardfly network, and the 8i Max doubles the compute, SRAM, and HBM bandwidth of the original 8i. <a href="https://www.tomshardware.com/tech-industry/semiconductors/google-splits-its-tpu-into-two-chips-for-the-first-time-with-training-and-inference-variants">The original split</a>, the TPU 8t and 8i, made natural sense. The 8t targets large-scale training, while the 8i leans toward low-latency inference and reasoning. TPUs have become a fast-growing business, with the company claiming nearly 100 customers, including <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/googles-tpus-for-ai-workloads-now-run-in-at-metas-data-centers-tpus-sold-to-external-customers-for-the-first-time-threaten-nvidias-market-dominance">Meta’s recent move to TPUs</a>. The more modular approach is flexible and can deliver specialized systems “in a matter of months rather than years,” helping that popularity along.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2000px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="KcaSgrdyhPSQKD5Y38ihWB" name="google-tpu-8tx-8i-max-body" alt="Renderings of Google’s four eighth-generation TPU systems, labeled TPU 8t, TPU 8i, TPU 8tx and TPU 8i Max, with Google’s captions for training, reasoning, confidential computing and context." src="https://cdn.mos.cms.futurecdn.net/KcaSgrdyhPSQKD5Y38ihWB-1920-80.jpg" mos="" align="middle" fullscreen="" width="2000" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Google)</span></figcaption></figure><div ><table><caption>Google’s eighth-generation TPU systems</caption><thead><tr><th class="firstcol empty" ></th><th  ><p>TPU 8t</p></th><th  ><p>TPU 8i</p></th><th  ><p>TPU 8tx</p></th><th  ><p>TPU 8i Max</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Built for</p></td><td  ><p>Large-scale pre-training</p></td><td  ><p>Sampling, serving and reasoning</p></td><td  ><p>Secure, confidential inference; small-scale training</p></td><td  ><p>Long-context agent work</p></td></tr><tr><td class="firstcol " ><p>Silicon</p></td><td  ><p>TPU 8t</p></td><td  ><p>TPU 8i</p></td><td  ><p>Same silicon as TPU 8t</p></td><td  ><p>Two TPU 8i chips on one package</p></td></tr><tr><td class="firstcol " ><p>Host CPU</p></td><td  ><p>Arm Axion</p></td><td  ><p>Arm Axion</p></td><td  ><p>5th Gen AMD EPYC</p></td><td  ><p>Not stated</p></td></tr><tr><td class="firstcol " ><p>Network</p></td><td  ><p>3D torus</p></td><td  ><p>Boardfly</p></td><td  ><p>Boardfly</p></td><td  ><p>Not stated</p></td></tr><tr><td class="firstcol " ><p>Peak FP4 PFLOPS</p></td><td  ><p>12.6</p></td><td  ><p>10.1</p></td><td  ><p>Not stated</p></td><td  ><p>About 20.2 (derived)</p></td></tr><tr><td class="firstcol " ><p>HBM bandwidth (GB/s)</p></td><td  ><p>6,528</p></td><td  ><p>8,601</p></td><td  ><p>Not stated</p></td><td  ><p>About 17,202 (derived)</p></td></tr><tr><td class="firstcol " ><p>HBM capacity (GB)</p></td><td  ><p>216</p></td><td  ><p>288</p></td><td  ><p>Not stated</p></td><td  ><p>Not stated</p></td></tr><tr><td class="firstcol " ><p>On-chip SRAM (MB)</p></td><td  ><p>128</p></td><td  ><p>384</p></td><td  ><p>Not stated</p></td><td  ><p>768 (derived)</p></td></tr><tr><td class="firstcol " ><p>Availability</p></td><td  ><p>Coming soon</p></td><td  ><p>Coming soon</p></td><td  ><p>Coming soon</p></td><td  ><p>Coming soon</p></td></tr></tbody></table></div><p>For the new parts, the TPU 8i Max isn't innovative on the surface, as it’s just two TPU 8i chips combined into a single package. The result is twice the theoretical compute, but more importantly, it fits twice the silicon within the “existing rack footprint,” and Google claims more compute per square foot and per watt.</p><p>Doubling both SRAM and HBM bandwidth helps keep things scaling, which would put the 8i Max at about 20.2 FP4 PFLOPS, 768 MB SRAM, and about 17.2 TB/s HBM bandwidth. This compares to <a href="https://tomshardware.com/pc-components/gpus/nvidia-launches-vera-rubin-nvl72-ai-supercomputer-at-ces-promises-up-to-5x-greater-inference-performance-and-10x-lower-cost-per-token-than-blackwell-coming-2h-2026">Nvidia’s Rubin R200 GPU</a> at a rated 50 FP4 PFLOPS for inference and 22 TB/s of HBM4 per socket, or<a href="https://www.tomshardware.com/pc-components/gpus/amd-takes-the-wraps-off-its-instinct-mi455x-ai-accelerator-cdna-5-and-helios-rack-scale-architecture-combine-to-take-the-fight-to-nvidia-in-the-data-center"> AMD’s MI455X</a> with 40 FP4 PFLOPS. The compute gap is closing.</p><p>The memory bandwidth doubling is also potentially one way around moving to HBM4. The memory wall, as Google calls it, or the bandwidth bottleneck, is a real concern “when dozens of AI agents work together across an entire software codebase.” Larger context windows — with larger models targeting 1 million tokens — mean larger KV caches, partly held in SRAM, and each new token has to read that cache back from memory, which is where the extra HBM bandwidth helps. It’s a different route from <a href="https://www.tomshardware.com/tech-industry/semiconductors/nvidia-presents-groq-3-lpx-architecture-and-unveils-its-first-third-party-inference-benchmark">Nvidia’s LPX</a>: more bandwidth in one package, rather than a second chip type for decoding and a storage tier for context.</p><p>Google also provides useful software for agents to get the most out of its hardware. TPU Sync is an open-source “model-state layer for TPU clusters” that moves KV caches off the chip’s HBM, “automatically tiering massive KV caches off the chip in milliseconds” down to host DRAM and, experimentally, shared storage. Google says this can cut KV transfer latency by “more than 90%,” with time-to-first-token gains that vary by workload. The company has also announced the Agent Compute Engine, a Google Cloud service that pauses idle agent sessions and moves their working memory aside until they’re needed again, another memory-saving effort.</p><p>Google credits its vLLM-TorchTPU stack with “state-of-the-art price-performance marks” on InferenceX and AgentX, which test realistic agentic coding and long-context work. However, the only published third-party numbers so far are for the last-generation Ironwood TPU. In <a href="https://newsletter.semianalysis.com/p/tpu-strikes-the-king" target="_blank"><em>SemiAnalysis</em>’s InferenceX tests in September</a>, running Qwen3.5 397B in FP8, Ironwood delivered 50.4% more tokens per dollar than Nvidia’s B200 and 96.0% more than the B300 at 20 tokens per second per user. One issue with that, however, is that Ironwood lacks native FP4, although newer November results show its FP8 beating Nvidia’s FP4 GB300 NVL72 on the same model, at 1.67 times the performance per dollar. There is also no eighth-gen result yet to compare how the new generation differs.</p><p>In contrast to the 8i Max, which is designed for long-context work, the new TPU 8tx, “built on the same silicon as TPU 8t,” is “optimized for secure and confidential inference” as well as small-scale training. The security is built into the hardware with isolation to prevent snooping on sensitive data, including model weights. Normally, data has to be unlocked in memory during processing, which is the step this isolation is meant to cover.</p><p>This chip is intended for areas where security and confidentiality are high priorities, which means regulated enterprises and governments, particularly those who are using frontier or proprietary models. In that case, the operators may want to “fine-tune models using their private data, distill large models into smaller ones, and train their agents,” without having to move sensitive data between clusters. Agents and security meet here for “secure agent execution.” Back in April, the Pentagon and Google <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/google-and-pentagon-in-talks-to-run-tpus-inside-classified-environments">reportedly held discussions</a> for TPU usage in classified settings, although the 8tx chip is not specifically named.</p><p>Google’s Private AI Compute, its “hardened TPU platform” with protections “comparable to a typical TEE,” per its September technical brief, already runs confidential serving on TPUs. The 8tx appears to be that capability turned into a product for customers. Alternatives include Nvidia’s NVL72, with one security domain across 72 Rubin GPUs, 36 Vera CPUs, and the interconnects, and <a href="https://www.tomshardware.com/pc-components/cpus/intel-to-phase-out-sgx-after-xeon-diamond-rapids-in-favor-of-tdx-confidential-computing-pivot-adds-trusted-i-o-and-post-quantum-crypto-for-agentic-ai">Intel’s TDX</a>, which is adding trusted I/O to cover data moving to accelerators.</p><p>The TPU 8tx hosts are based on 5th Gen AMD EPYC boards, instead of the Axion hosts on the 8t, which replaced x86 in April. <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/google-reportedly-taps-amd-to-design-next-generation-tpu-hybrid-ai-asic-could-integrate-on-package-cpu-cores-for-reinforcement-learning">There's one Axion for every two TPUs</a> on the 8i, compared with one Xeon per four on Ironwood. Google gives no reason for the change, but its Confidential VM configuration list has both AMD and Intel series chips, with no Arm-based ones listed. Private AI Compute uses “an AMD-based hardware Trusted Execution Environment” for its CPU workloads, suggesting Google swapped something it already runs on its confidential stack.</p><p>The 8tx also adopts the 8i’s Boardfly network over the 8t’s torus, which Google claims can deliver “up to 2x lower latency compared to TPU 8t” with almost no performance penalty, although it’s not clear what measure Google is using. The alternative torus network suits training, since chips mostly swap data with neighbors in a predictable manner, but inference on large mixture-of-experts and reasoning models can send smaller messages between distant chips, with every hop adding delay. With Boardfly, it’s at most seven hops against 16 at 1,024 chips.</p><p>What is lacking is a full specifications table, the amount of HBM capacity, pod sizes, an availability date, pricing, and customers for either system. Details on attestation, that is, the ability for the customer to verify confidentiality, are also missing for 8tx. The 8tx and 8i Max are currently listed, along with the 8i and 8t, as “coming soon” on Google’s Cloud TPU page.</p><p>Delivering two new TPU systems nearly seven months after April’s pair shows Google wants to turn out specialized systems much more quickly than has been the case in the past. Now, it just needs to prove them, through benchmarks and an availability date when the hardware is actually in customers’ hands.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/artificial-intelligence/google-adds-two-new-specialized-chips-for-ai-agents-and-security-to-8th-gen-tpu-lineup-tpu-8i-max-puts-two-chips-on-one-package-for-long-context-work-tpu-8tx-swaps-arm-hosts-for-amd-epyc</link>
                                                                            <description>
                            <![CDATA[ Google’s TPU 8i Max and TPU 8tx reuse TPU 8i and TPU 8t silicon to target long-context agents and confidential inference. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">TMNyhd3roQRhtcXcyzUMzM</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/Ut57hXr6UTF3v4LJmkuGSi-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Thu, 12 Nov 2026 11:00:00 +0000</pubDate>                                                                                                                                <updated>Fri, 13 Nov 2026 13:22:29 +0000</updated>
                                                                                                                                            <category><![CDATA[Artificial Intelligence]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/Ut57hXr6UTF3v4LJmkuGSi-1920-80.jpg">
                                                            <media:credit><![CDATA[Google]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Renderings of a Google TPU 8tx board with four chips and four TPU 8i Max boards, each carrying two chips on one package, labeled TPU 8tx and TPU 8i Max.]]></media:description>                                                            <media:text><![CDATA[Renderings of a Google TPU 8tx board with four chips and four TPU 8i Max boards, each carrying two chips on one package, labeled TPU 8tx and TPU 8i Max.]]></media:text>
                                <media:title type="plain"><![CDATA[Renderings of a Google TPU 8tx board with four chips and four TPU 8i Max boards, each carrying two chips on one package, labeled TPU 8tx and TPU 8i Max.]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/Ut57hXr6UTF3v4LJmkuGSi-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Google's eighth-generation TPU family has now expanded further. Back in April, the company announced TPU 8t for training AI models and TPU 8i for general-purpose inference workloads. Google has now introduced TPU 8tx, built on the same silicon as the previously announced 8t, for confidential, hardware-isolated inference aimed at regulated enterprises and governments, and TPU 8i Max, which combines two 8i chips on one package for long-context agentic work. </p><p>TPU 8tx swaps the Arm-based Axion hosts of the 8t and 8i for a 5th Gen AMD EPYC CPU, adopting a Boardfly network, and the 8i Max doubles the compute, SRAM, and HBM bandwidth of the original 8i. <a href="https://www.tomshardware.com/tech-industry/semiconductors/google-splits-its-tpu-into-two-chips-for-the-first-time-with-training-and-inference-variants">The original split</a>, the TPU 8t and 8i, made natural sense. The 8t targets large-scale training, while the 8i leans toward low-latency inference and reasoning. TPUs have become a fast-growing business, with the company claiming nearly 100 customers, including <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/googles-tpus-for-ai-workloads-now-run-in-at-metas-data-centers-tpus-sold-to-external-customers-for-the-first-time-threaten-nvidias-market-dominance">Meta’s recent move to TPUs</a>. The more modular approach is flexible and can deliver specialized systems “in a matter of months rather than years,” helping that popularity along.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2000px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="KcaSgrdyhPSQKD5Y38ihWB" name="google-tpu-8tx-8i-max-body" alt="Renderings of Google’s four eighth-generation TPU systems, labeled TPU 8t, TPU 8i, TPU 8tx and TPU 8i Max, with Google’s captions for training, reasoning, confidential computing and context." src="https://cdn.mos.cms.futurecdn.net/KcaSgrdyhPSQKD5Y38ihWB-1920-80.jpg" mos="" align="middle" fullscreen="" width="2000" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Google)</span></figcaption></figure><div ><table><caption>Google’s eighth-generation TPU systems</caption><thead><tr><th class="firstcol empty" ></th><th  ><p>TPU 8t</p></th><th  ><p>TPU 8i</p></th><th  ><p>TPU 8tx</p></th><th  ><p>TPU 8i Max</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Built for</p></td><td  ><p>Large-scale pre-training</p></td><td  ><p>Sampling, serving and reasoning</p></td><td  ><p>Secure, confidential inference; small-scale training</p></td><td  ><p>Long-context agent work</p></td></tr><tr><td class="firstcol " ><p>Silicon</p></td><td  ><p>TPU 8t</p></td><td  ><p>TPU 8i</p></td><td  ><p>Same silicon as TPU 8t</p></td><td  ><p>Two TPU 8i chips on one package</p></td></tr><tr><td class="firstcol " ><p>Host CPU</p></td><td  ><p>Arm Axion</p></td><td  ><p>Arm Axion</p></td><td  ><p>5th Gen AMD EPYC</p></td><td  ><p>Not stated</p></td></tr><tr><td class="firstcol " ><p>Network</p></td><td  ><p>3D torus</p></td><td  ><p>Boardfly</p></td><td  ><p>Boardfly</p></td><td  ><p>Not stated</p></td></tr><tr><td class="firstcol " ><p>Peak FP4 PFLOPS</p></td><td  ><p>12.6</p></td><td  ><p>10.1</p></td><td  ><p>Not stated</p></td><td  ><p>About 20.2 (derived)</p></td></tr><tr><td class="firstcol " ><p>HBM bandwidth (GB/s)</p></td><td  ><p>6,528</p></td><td  ><p>8,601</p></td><td  ><p>Not stated</p></td><td  ><p>About 17,202 (derived)</p></td></tr><tr><td class="firstcol " ><p>HBM capacity (GB)</p></td><td  ><p>216</p></td><td  ><p>288</p></td><td  ><p>Not stated</p></td><td  ><p>Not stated</p></td></tr><tr><td class="firstcol " ><p>On-chip SRAM (MB)</p></td><td  ><p>128</p></td><td  ><p>384</p></td><td  ><p>Not stated</p></td><td  ><p>768 (derived)</p></td></tr><tr><td class="firstcol " ><p>Availability</p></td><td  ><p>Coming soon</p></td><td  ><p>Coming soon</p></td><td  ><p>Coming soon</p></td><td  ><p>Coming soon</p></td></tr></tbody></table></div><p>For the new parts, the TPU 8i Max isn't innovative on the surface, as it’s just two TPU 8i chips combined into a single package. The result is twice the theoretical compute, but more importantly, it fits twice the silicon within the “existing rack footprint,” and Google claims more compute per square foot and per watt.</p><p>Doubling both SRAM and HBM bandwidth helps keep things scaling, which would put the 8i Max at about 20.2 FP4 PFLOPS, 768 MB SRAM, and about 17.2 TB/s HBM bandwidth. This compares to <a href="https://tomshardware.com/pc-components/gpus/nvidia-launches-vera-rubin-nvl72-ai-supercomputer-at-ces-promises-up-to-5x-greater-inference-performance-and-10x-lower-cost-per-token-than-blackwell-coming-2h-2026">Nvidia’s Rubin R200 GPU</a> at a rated 50 FP4 PFLOPS for inference and 22 TB/s of HBM4 per socket, or<a href="https://www.tomshardware.com/pc-components/gpus/amd-takes-the-wraps-off-its-instinct-mi455x-ai-accelerator-cdna-5-and-helios-rack-scale-architecture-combine-to-take-the-fight-to-nvidia-in-the-data-center"> AMD’s MI455X</a> with 40 FP4 PFLOPS. The compute gap is closing.</p><p>The memory bandwidth doubling is also potentially one way around moving to HBM4. The memory wall, as Google calls it, or the bandwidth bottleneck, is a real concern “when dozens of AI agents work together across an entire software codebase.” Larger context windows — with larger models targeting 1 million tokens — mean larger KV caches, partly held in SRAM, and each new token has to read that cache back from memory, which is where the extra HBM bandwidth helps. It’s a different route from <a href="https://www.tomshardware.com/tech-industry/semiconductors/nvidia-presents-groq-3-lpx-architecture-and-unveils-its-first-third-party-inference-benchmark">Nvidia’s LPX</a>: more bandwidth in one package, rather than a second chip type for decoding and a storage tier for context.</p><p>Google also provides useful software for agents to get the most out of its hardware. TPU Sync is an open-source “model-state layer for TPU clusters” that moves KV caches off the chip’s HBM, “automatically tiering massive KV caches off the chip in milliseconds” down to host DRAM and, experimentally, shared storage. Google says this can cut KV transfer latency by “more than 90%,” with time-to-first-token gains that vary by workload. The company has also announced the Agent Compute Engine, a Google Cloud service that pauses idle agent sessions and moves their working memory aside until they’re needed again, another memory-saving effort.</p><p>Google credits its vLLM-TorchTPU stack with “state-of-the-art price-performance marks” on InferenceX and AgentX, which test realistic agentic coding and long-context work. However, the only published third-party numbers so far are for the last-generation Ironwood TPU. In <a href="https://newsletter.semianalysis.com/p/tpu-strikes-the-king" target="_blank"><em>SemiAnalysis</em>’s InferenceX tests in September</a>, running Qwen3.5 397B in FP8, Ironwood delivered 50.4% more tokens per dollar than Nvidia’s B200 and 96.0% more than the B300 at 20 tokens per second per user. One issue with that, however, is that Ironwood lacks native FP4, although newer November results show its FP8 beating Nvidia’s FP4 GB300 NVL72 on the same model, at 1.67 times the performance per dollar. There is also no eighth-gen result yet to compare how the new generation differs.</p><p>In contrast to the 8i Max, which is designed for long-context work, the new TPU 8tx, “built on the same silicon as TPU 8t,” is “optimized for secure and confidential inference” as well as small-scale training. The security is built into the hardware with isolation to prevent snooping on sensitive data, including model weights. Normally, data has to be unlocked in memory during processing, which is the step this isolation is meant to cover.</p><p>This chip is intended for areas where security and confidentiality are high priorities, which means regulated enterprises and governments, particularly those who are using frontier or proprietary models. In that case, the operators may want to “fine-tune models using their private data, distill large models into smaller ones, and train their agents,” without having to move sensitive data between clusters. Agents and security meet here for “secure agent execution.” Back in April, the Pentagon and Google <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/google-and-pentagon-in-talks-to-run-tpus-inside-classified-environments">reportedly held discussions</a> for TPU usage in classified settings, although the 8tx chip is not specifically named.</p><p>Google’s Private AI Compute, its “hardened TPU platform” with protections “comparable to a typical TEE,” per its September technical brief, already runs confidential serving on TPUs. The 8tx appears to be that capability turned into a product for customers. Alternatives include Nvidia’s NVL72, with one security domain across 72 Rubin GPUs, 36 Vera CPUs, and the interconnects, and <a href="https://www.tomshardware.com/pc-components/cpus/intel-to-phase-out-sgx-after-xeon-diamond-rapids-in-favor-of-tdx-confidential-computing-pivot-adds-trusted-i-o-and-post-quantum-crypto-for-agentic-ai">Intel’s TDX</a>, which is adding trusted I/O to cover data moving to accelerators.</p><p>The TPU 8tx hosts are based on 5th Gen AMD EPYC boards, instead of the Axion hosts on the 8t, which replaced x86 in April. <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/google-reportedly-taps-amd-to-design-next-generation-tpu-hybrid-ai-asic-could-integrate-on-package-cpu-cores-for-reinforcement-learning">There's one Axion for every two TPUs</a> on the 8i, compared with one Xeon per four on Ironwood. Google gives no reason for the change, but its Confidential VM configuration list has both AMD and Intel series chips, with no Arm-based ones listed. Private AI Compute uses “an AMD-based hardware Trusted Execution Environment” for its CPU workloads, suggesting Google swapped something it already runs on its confidential stack.</p><p>The 8tx also adopts the 8i’s Boardfly network over the 8t’s torus, which Google claims can deliver “up to 2x lower latency compared to TPU 8t” with almost no performance penalty, although it’s not clear what measure Google is using. The alternative torus network suits training, since chips mostly swap data with neighbors in a predictable manner, but inference on large mixture-of-experts and reasoning models can send smaller messages between distant chips, with every hop adding delay. With Boardfly, it’s at most seven hops against 16 at 1,024 chips.</p><p>What is lacking is a full specifications table, the amount of HBM capacity, pod sizes, an availability date, pricing, and customers for either system. Details on attestation, that is, the ability for the customer to verify confidentiality, are also missing for 8tx. The 8tx and 8i Max are currently listed, along with the 8i and 8t, as “coming soon” on Google’s Cloud TPU page.</p><p>Delivering two new TPU systems nearly seven months after April’s pair shows Google wants to turn out specialized systems much more quickly than has been the case in the past. Now, it just needs to prove them, through benchmarks and an availability date when the hardware is actually in customers’ hands.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ AMD shares how it can retain 95% of server performance with just half the DRAM, sidestepping memory shortages with MEXT  ]]></title>
                                                                                                <dc:content><![CDATA[ <p>In June, <a href="https://www.tomshardware.com/pc-components/cpus/amd-takes-over-mext-to-address-growing-memory-constraints-in-the-data-center-memory-tiering-technology-enables-flash-to-appear-as-dram-to-applications"><u>AMD overtook the predictive memory engine</u></a> known as MEXT, which claims to offer upwards of 99% of the performance of a system with full DRAM capacity while running with as little as 25% of the physical DIMMs. The technology offloads cold memory pages to flash storage and uses an AI model to serve pages before they’re needed, and AMD says it plans to extend the tech across a wide swath of its IT infrastructure next year. </p><p>The pitch behind MEXT seems too good to be true. According to AMD and MEXT founder Gary Smerdon, the technology can cut the physical DRAM required in a server by 20% to 75% depending on the application and performance target. That missing memory still registers as installed DRAM to the OS. Instead of DRAM, however, MEXT uses relatively inexpensive flash storage to offload cold pages, flying them back in as they’re needed with an AI model. In some of AMD’s internal workloads, such as Synopsys PrimeTime, the company says it’s able to achieve 95% of the performance of a full server while running at 50% DRAM capacity with MEXT. </p><p>AMD has deployed MEXT internally on 145 servers, which it says will expand to 2,000 servers next year. In addition, the company has announced a partnership with HPE to bring MEXT to customers. Currently, we’re in phase 1 of the rollout, with some <a href="https://www.tomshardware.com/pc-components/cpus/amd-drops-an-epyc-usd15-000-256-core-bomb-epyc-9006-zen-6-venice-cpus-get-full-spec-and-pricing-treatment-from-usd700-up-to-usd14-904">EPYC Venice servers</a> carrying a “MEXT Ready” badge and a dedicated NVMe drive or bay for the offload. Phase 2 brings MEXT licenses out to these servers, as well as the RAM View application. AMD hasn’t said when Phase 2 will begin. </p><p>The AI-driven DRAM pricing surge impacts more than just gaming PCs, with enterprises also feeling the brunt of price increases, even if they aren’t running AI workloads. If MEXT is able to achieve what AMD claims, it could be transformative given the current market. Because of that, it’s worth understanding what exactly MEXT is, as well as what AMD claims it can do. </p><p>Although MEXT is exciting in concept, it’s obvious that AMD is taking a conservative approach to the rollout. Extraordinary claims require extraordinary evidence, after all. The applications of MEXT are currently narrow, though AMD says it’s expanding the workloads it’s using MEXT on with promising early results. </p><h2 id="dram-flash-and-an-llm">DRAM, Flash and an LLM</h2><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:56.28%;"><img id="szd4pFW7dS3x2zDzur5WBD" name="image5" alt="AMD MEXT Slides" src="https://cdn.mos.cms.futurecdn.net/szd4pFW7dS3x2zDzur5WBD-1920-80.jpg" mos="" align="middle" fullscreen="" width="1999" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: AMD)</span></figcaption></figure><p>The basic setup of MEXT is simple in concept: cold pages go to flash, AI brings those pages back as needed, and it’s as if the memory was never missing in the first place. The main problem should be obvious. Flash is orders of magnitude slower than DRAM, so streaming pages back into memory from flash as they’re needed isn’t feasible. MEXT’s trick is using an AI model to <em>predict </em>the pages you’ll need. And, as long as the model is able to stay ahead by at least the transfer latency from the SSD, it’s as if the physical DRAM was installed the whole time. </p><p>This isn’t a crazy concept considering how a transformer-based LLM, which MEXT uses, operates. If you’re unfamiliar with the inner workings, <a href="https://poloclub.github.io/transformer-explainer/"><u>this open-source visualization</u></a> breaks down the concept of a transformer AI model cleanly. The foundational concept is next token prediction, which underpins the design of modern LLMs. Given a set of tokens, the LLM determines which token (word, portion of a word, etc.) is most likely to come next. It makes a prediction. In the case of MEXT, it’s predicting what page is needed next, not what word comes next. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:56.28%;"><img id="u6qu2xxWYwZ24zermszh8D" name="image4" alt="AMD MEXT Slides" src="https://cdn.mos.cms.futurecdn.net/u6qu2xxWYwZ24zermszh8D-1920-80.jpg" mos="" align="middle" fullscreen="" width="1999" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: AMD)</span></figcaption></figure><p>Memory <a href="https://engineering.fb.com/2022/06/20/data-infrastructure/transparent-memory-offloading-more-memory-at-a-fraction-of-the-cost-and-power/"><u>offloading to flash isn’t a new concept</u></a>, but MEXT founder Gary Smerdon says that AI is the critical difference in this technology. MEXT itself is deployed as a Linux kernel module that works with any Linux distro. It uses an “ensemble of AI models,” with the primary transformer-based LLM driving the technology. And critically, it doesn’t come pre-trained. </p><p>Instead, AMD says that MEXT trains in under a second when first deployed, operates on one EPYC core without a GPU, and continually trains as the workload progresses. This approach makes sense given the varied workloads that MEXT may apply to. The idea is that it would learn the access patterns of the application in order to drive that prediction engine. </p><p>MEXT isn’t perfect, but it doesn’t need to be perfect in every situation. For instance, the order of stored pages isn’t important. With text, an LLM needs to predict the correct text in the correct order, but with MEXT, the order doesn’t matter. As long as the page is in DRAM when it’s needed, that’s fine. </p><p>In order to understand the application of MEXT to different workloads, AMD has developed RAM View to go alongside the tech. It’s a memory utilization tool that shows which pages are hot and which have gone cold. As you can see from the example above, a system with 1.5 TB of memory shows near-full utilization across the full workload, but around half of the pages stored in memory haven’t been touched in five minutes or more at any given time. </p><p>AMD provided this generalized example and didn’t clarify what workload it was testing. That’s very important, as the utilization breakdown will look vastly different depending on the application(s) running on the server. Initially, AMD is applying MEXT to critical EDA applications it uses internally, such as Synopsys PrimeTime, but the company says it has focused on optimizing agentic AI, database, and media and entertainment workloads as well. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:56.28%;"><img id="VZRn2Xu94NrsX3DYokQTKD" name="image8" alt="AMD MEXT Slides" src="https://cdn.mos.cms.futurecdn.net/VZRn2Xu94NrsX3DYokQTKD-1920-80.jpg" mos="" align="middle" fullscreen="" width="1999" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: AMD)</span></figcaption></figure><p>This is important not only because memory access patterns change between applications, but also because performance in those applications reacts differently when a system doesn’t have enough memory. The “memory wall,” as AMD calls it, should be a familiar concept regardless of whether you’ve ever touched a server. With MEXT, AMD says it’s trying to extend this memory wall, which may look different for different applications. </p><p>For instance, the company is confident that many workloads can reduce capacity by around 20% to 25%. As the reduction increases, however, the applications become more specialized, and the performance delta becomes greater. </p><h2 id="amd-s-internal-mext-results">AMD’s internal MEXT results</h2><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:56.28%;"><img id="LXsDefHdYwv8g5uWxdPcCD" name="image6" alt="AMD MEXT Slides" src="https://cdn.mos.cms.futurecdn.net/LXsDefHdYwv8g5uWxdPcCD-1920-80.jpg" mos="" align="middle" fullscreen="" width="1999" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: AMD)</span></figcaption></figure><p>AMD has shared some internal MEXT results, which look impressive, though as usual, it’s always smart to handle first-party benchmarks like this with skepticism. Starting with PrimeTime, where AMD has focused much of its attention on MEXT up to this point, it says servers running at 50% capacity retain 95% of the performance of a full server, and going down to as little as 25% capacity still maintains 80% of the application’s performance. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:56.28%;"><img id="4XCC4EAKicwHNCfqGB8hyC" name="image1" alt="AMD MEXT Slides" src="https://cdn.mos.cms.futurecdn.net/4XCC4EAKicwHNCfqGB8hyC-1920-80.jpg" mos="" align="middle" fullscreen="" width="1999" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: AMD)</span></figcaption></figure><p>PrimeTime demands a lot of memory for large designs. There are some applications where performance is nearly identical, according to AMD. In a 3-node cluster running in Azure with 5th-gen EPYC CPUs, AMD says MEXT offers 99% of the performance of the full cluster while running at 50% DRAM capacity with the Spark SQL workload from the TPC-DS benchmark. SQL databases try to cache as much data as possible in memory to prevent I/O bottlenecks, creating a golden opportunity for tech like MEXT that wants to free cold pages. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:56.28%;"><img id="xFg3RWrLQKaZA3mmDPZcCD" name="image3" alt="AMD MEXT Slides" src="https://cdn.mos.cms.futurecdn.net/xFg3RWrLQKaZA3mmDPZcCD-1920-80.jpg" mos="" align="middle" fullscreen="" width="1999" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: AMD)</span></figcaption></figure><p>Perhaps the most applicable workload at the current moment, however, is agentic AI, and AMD says MEXT has similar applications there. Across a five-agent concurrent geomean, AMD claims 99% of the performance of a full server with 50% memory capacity, and 85% of the performance with 25% capacity. With a “memory-hungry” agent, AMD claims 90% and 88% performance for 50% and 25% capacity, respectively. </p><p>The slide you can see above has footnote tags. However, AMD has not shared the footnotes with us yet, so we don’t have a view into how these tests were run. Naturally, we assume AMD has, at minimum, represented the performance of MEXT in the best light possible. We can’t do a full analysis without knowing how the tests were run, however. </p><p>What is interesting is AMD’s cautious approach to rolling out MEXT. It’s an EPYC-exclusive feature right now, though AMD left the door open for support on future products (consumer devices, such as <a href="https://www.tomshardware.com/pc-components/air-cooling/deepcool-ak620-and-ak400-g2-review">Strix Halo</a>) far down the road. It will be sold as a license, bundled with the RAM View software, but we don’t have a timeline on when that will happen. And, despite a partnership with HPE to push out “MEXT Ready” servers, we still don’t know when MEXT itself will be ready for deployment beyond AMD’s walls. </p><p>Given that conservative approach, there are likely pitfalls to MEXT that we’re not seeing yet, and it’s hard to imagine it will be broadly applicable to every server, at least at the 50% and 25% DRAM capacity levels that AMD showed off with its benchmarks. </p><p>However, even being able to trim 25% of a server’s DRAM capacity represents a massive cost savings in the current supply-constrained environment. If some workloads can push that to 50% capacity, that’s even better. Of course, we’ll need to see how MEXT holds up once it’s available broadly, but if AMD’s claims hold water, it could become a defining technology in data centers that can’t seem to get enough memory.  </p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/pc-components/cpus/amd-shares-how-it-can-retain-95-percent-of-server-performance-with-just-half-the-dram-sidestepping-memory-shortages-with-mext-tech-powers-over-100-servers-already-running-with-plans-to-expand-to-2-000-servers-in-2027</link>
                                                                            <description>
                            <![CDATA[ Following AMD’s acquisition of MEXT earlier this year, the company has revealed its own IT infrastructure is using the technology to run servers at 50% DRAM capacity while achieving 95% of the performance of a full-DRAM system in certain applications. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">MckFLx3NUAQjroR83AvKob</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/Yr76o6YXagRHfHkjYNTT3D-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Wed, 11 Nov 2026 12:00:00 +0000</pubDate>                                                                                                                                <updated>Wed, 11 Nov 2026 14:10:45 +0000</updated>
                                                                                                                                            <category><![CDATA[CPUs]]></category>
                                                    <category><![CDATA[PC Components]]></category>
                                                                                                                    <dc:creator><![CDATA[ Jake Roach ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/h6PRM8bTimCTnNfoAYfjAi-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Jake Roach has been bending pins and busting solder joints since the mid-2000s. From trying to run scratched CDs of &lt;em&gt;Delta Force &lt;/em&gt;and &lt;em&gt;Unreal Tournament &lt;/em&gt;to spitting out virtual machines on a Threadripper, Jake has been on the hunt for the latest hardware and highest performance for decades. That eventually spun up a career, with Jake serving as Lead Reporter at Digital Trends, as well as contributing to outlets like XDA, PC Invasion, Business Insider, and WIRED. At Tom’s Hardware, Jake is focused on consumer and workstation CPUs. Outside working hours, you’ll find him knee-deep in the latest roguelite taking over Steam, spending way too much money on &lt;em&gt;Magic: The Gathering, &lt;/em&gt;or forcing his lazy corgi onto walks.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/Yr76o6YXagRHfHkjYNTT3D-1920-80.jpg">
                                                            <media:credit><![CDATA[AMD]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[AMD MEXT Slides]]></media:description>                                                            <media:text><![CDATA[AMD MEXT Slides]]></media:text>
                                <media:title type="plain"><![CDATA[AMD MEXT Slides]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/Yr76o6YXagRHfHkjYNTT3D-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>In June, <a href="https://www.tomshardware.com/pc-components/cpus/amd-takes-over-mext-to-address-growing-memory-constraints-in-the-data-center-memory-tiering-technology-enables-flash-to-appear-as-dram-to-applications"><u>AMD overtook the predictive memory engine</u></a> known as MEXT, which claims to offer upwards of 99% of the performance of a system with full DRAM capacity while running with as little as 25% of the physical DIMMs. The technology offloads cold memory pages to flash storage and uses an AI model to serve pages before they’re needed, and AMD says it plans to extend the tech across a wide swath of its IT infrastructure next year. </p><p>The pitch behind MEXT seems too good to be true. According to AMD and MEXT founder Gary Smerdon, the technology can cut the physical DRAM required in a server by 20% to 75% depending on the application and performance target. That missing memory still registers as installed DRAM to the OS. Instead of DRAM, however, MEXT uses relatively inexpensive flash storage to offload cold pages, flying them back in as they’re needed with an AI model. In some of AMD’s internal workloads, such as Synopsys PrimeTime, the company says it’s able to achieve 95% of the performance of a full server while running at 50% DRAM capacity with MEXT. </p><p>AMD has deployed MEXT internally on 145 servers, which it says will expand to 2,000 servers next year. In addition, the company has announced a partnership with HPE to bring MEXT to customers. Currently, we’re in phase 1 of the rollout, with some <a href="https://www.tomshardware.com/pc-components/cpus/amd-drops-an-epyc-usd15-000-256-core-bomb-epyc-9006-zen-6-venice-cpus-get-full-spec-and-pricing-treatment-from-usd700-up-to-usd14-904">EPYC Venice servers</a> carrying a “MEXT Ready” badge and a dedicated NVMe drive or bay for the offload. Phase 2 brings MEXT licenses out to these servers, as well as the RAM View application. AMD hasn’t said when Phase 2 will begin. </p><p>The AI-driven DRAM pricing surge impacts more than just gaming PCs, with enterprises also feeling the brunt of price increases, even if they aren’t running AI workloads. If MEXT is able to achieve what AMD claims, it could be transformative given the current market. Because of that, it’s worth understanding what exactly MEXT is, as well as what AMD claims it can do. </p><p>Although MEXT is exciting in concept, it’s obvious that AMD is taking a conservative approach to the rollout. Extraordinary claims require extraordinary evidence, after all. The applications of MEXT are currently narrow, though AMD says it’s expanding the workloads it’s using MEXT on with promising early results. </p><h2 id="dram-flash-and-an-llm">DRAM, Flash and an LLM</h2><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:56.28%;"><img id="szd4pFW7dS3x2zDzur5WBD" name="image5" alt="AMD MEXT Slides" src="https://cdn.mos.cms.futurecdn.net/szd4pFW7dS3x2zDzur5WBD-1920-80.jpg" mos="" align="middle" fullscreen="" width="1999" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: AMD)</span></figcaption></figure><p>The basic setup of MEXT is simple in concept: cold pages go to flash, AI brings those pages back as needed, and it’s as if the memory was never missing in the first place. The main problem should be obvious. Flash is orders of magnitude slower than DRAM, so streaming pages back into memory from flash as they’re needed isn’t feasible. MEXT’s trick is using an AI model to <em>predict </em>the pages you’ll need. And, as long as the model is able to stay ahead by at least the transfer latency from the SSD, it’s as if the physical DRAM was installed the whole time. </p><p>This isn’t a crazy concept considering how a transformer-based LLM, which MEXT uses, operates. If you’re unfamiliar with the inner workings, <a href="https://poloclub.github.io/transformer-explainer/"><u>this open-source visualization</u></a> breaks down the concept of a transformer AI model cleanly. The foundational concept is next token prediction, which underpins the design of modern LLMs. Given a set of tokens, the LLM determines which token (word, portion of a word, etc.) is most likely to come next. It makes a prediction. In the case of MEXT, it’s predicting what page is needed next, not what word comes next. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:56.28%;"><img id="u6qu2xxWYwZ24zermszh8D" name="image4" alt="AMD MEXT Slides" src="https://cdn.mos.cms.futurecdn.net/u6qu2xxWYwZ24zermszh8D-1920-80.jpg" mos="" align="middle" fullscreen="" width="1999" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: AMD)</span></figcaption></figure><p>Memory <a href="https://engineering.fb.com/2022/06/20/data-infrastructure/transparent-memory-offloading-more-memory-at-a-fraction-of-the-cost-and-power/"><u>offloading to flash isn’t a new concept</u></a>, but MEXT founder Gary Smerdon says that AI is the critical difference in this technology. MEXT itself is deployed as a Linux kernel module that works with any Linux distro. It uses an “ensemble of AI models,” with the primary transformer-based LLM driving the technology. And critically, it doesn’t come pre-trained. </p><p>Instead, AMD says that MEXT trains in under a second when first deployed, operates on one EPYC core without a GPU, and continually trains as the workload progresses. This approach makes sense given the varied workloads that MEXT may apply to. The idea is that it would learn the access patterns of the application in order to drive that prediction engine. </p><p>MEXT isn’t perfect, but it doesn’t need to be perfect in every situation. For instance, the order of stored pages isn’t important. With text, an LLM needs to predict the correct text in the correct order, but with MEXT, the order doesn’t matter. As long as the page is in DRAM when it’s needed, that’s fine. </p><p>In order to understand the application of MEXT to different workloads, AMD has developed RAM View to go alongside the tech. It’s a memory utilization tool that shows which pages are hot and which have gone cold. As you can see from the example above, a system with 1.5 TB of memory shows near-full utilization across the full workload, but around half of the pages stored in memory haven’t been touched in five minutes or more at any given time. </p><p>AMD provided this generalized example and didn’t clarify what workload it was testing. That’s very important, as the utilization breakdown will look vastly different depending on the application(s) running on the server. Initially, AMD is applying MEXT to critical EDA applications it uses internally, such as Synopsys PrimeTime, but the company says it has focused on optimizing agentic AI, database, and media and entertainment workloads as well. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:56.28%;"><img id="VZRn2Xu94NrsX3DYokQTKD" name="image8" alt="AMD MEXT Slides" src="https://cdn.mos.cms.futurecdn.net/VZRn2Xu94NrsX3DYokQTKD-1920-80.jpg" mos="" align="middle" fullscreen="" width="1999" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: AMD)</span></figcaption></figure><p>This is important not only because memory access patterns change between applications, but also because performance in those applications reacts differently when a system doesn’t have enough memory. The “memory wall,” as AMD calls it, should be a familiar concept regardless of whether you’ve ever touched a server. With MEXT, AMD says it’s trying to extend this memory wall, which may look different for different applications. </p><p>For instance, the company is confident that many workloads can reduce capacity by around 20% to 25%. As the reduction increases, however, the applications become more specialized, and the performance delta becomes greater. </p><h2 id="amd-s-internal-mext-results">AMD’s internal MEXT results</h2><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:56.28%;"><img id="LXsDefHdYwv8g5uWxdPcCD" name="image6" alt="AMD MEXT Slides" src="https://cdn.mos.cms.futurecdn.net/LXsDefHdYwv8g5uWxdPcCD-1920-80.jpg" mos="" align="middle" fullscreen="" width="1999" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: AMD)</span></figcaption></figure><p>AMD has shared some internal MEXT results, which look impressive, though as usual, it’s always smart to handle first-party benchmarks like this with skepticism. Starting with PrimeTime, where AMD has focused much of its attention on MEXT up to this point, it says servers running at 50% capacity retain 95% of the performance of a full server, and going down to as little as 25% capacity still maintains 80% of the application’s performance. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:56.28%;"><img id="4XCC4EAKicwHNCfqGB8hyC" name="image1" alt="AMD MEXT Slides" src="https://cdn.mos.cms.futurecdn.net/4XCC4EAKicwHNCfqGB8hyC-1920-80.jpg" mos="" align="middle" fullscreen="" width="1999" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: AMD)</span></figcaption></figure><p>PrimeTime demands a lot of memory for large designs. There are some applications where performance is nearly identical, according to AMD. In a 3-node cluster running in Azure with 5th-gen EPYC CPUs, AMD says MEXT offers 99% of the performance of the full cluster while running at 50% DRAM capacity with the Spark SQL workload from the TPC-DS benchmark. SQL databases try to cache as much data as possible in memory to prevent I/O bottlenecks, creating a golden opportunity for tech like MEXT that wants to free cold pages. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:56.28%;"><img id="xFg3RWrLQKaZA3mmDPZcCD" name="image3" alt="AMD MEXT Slides" src="https://cdn.mos.cms.futurecdn.net/xFg3RWrLQKaZA3mmDPZcCD-1920-80.jpg" mos="" align="middle" fullscreen="" width="1999" height="1125" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: AMD)</span></figcaption></figure><p>Perhaps the most applicable workload at the current moment, however, is agentic AI, and AMD says MEXT has similar applications there. Across a five-agent concurrent geomean, AMD claims 99% of the performance of a full server with 50% memory capacity, and 85% of the performance with 25% capacity. With a “memory-hungry” agent, AMD claims 90% and 88% performance for 50% and 25% capacity, respectively. </p><p>The slide you can see above has footnote tags. However, AMD has not shared the footnotes with us yet, so we don’t have a view into how these tests were run. Naturally, we assume AMD has, at minimum, represented the performance of MEXT in the best light possible. We can’t do a full analysis without knowing how the tests were run, however. </p><p>What is interesting is AMD’s cautious approach to rolling out MEXT. It’s an EPYC-exclusive feature right now, though AMD left the door open for support on future products (consumer devices, such as <a href="https://www.tomshardware.com/pc-components/air-cooling/deepcool-ak620-and-ak400-g2-review">Strix Halo</a>) far down the road. It will be sold as a license, bundled with the RAM View software, but we don’t have a timeline on when that will happen. And, despite a partnership with HPE to push out “MEXT Ready” servers, we still don’t know when MEXT itself will be ready for deployment beyond AMD’s walls. </p><p>Given that conservative approach, there are likely pitfalls to MEXT that we’re not seeing yet, and it’s hard to imagine it will be broadly applicable to every server, at least at the 50% and 25% DRAM capacity levels that AMD showed off with its benchmarks. </p><p>However, even being able to trim 25% of a server’s DRAM capacity represents a massive cost savings in the current supply-constrained environment. If some workloads can push that to 50% capacity, that’s even better. Of course, we’ll need to see how MEXT holds up once it’s available broadly, but if AMD’s claims hold water, it could become a defining technology in data centers that can’t seem to get enough memory.  </p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ New framework grades AI jailbreaks across capability gains, scope, and how easy they are to deploy ]]></title>
                                                                                                <dc:content><![CDATA[ <p>In recent months, high-profile headlines about AI models letting loose from their safeguards dominated headlines. Now, the Frontier Model Forum (FMF) has released a new <a href="https://www.frontiermodelforum.org/updates/introducing-the-fmf-jailbreak-rating-framework-for-frontier-cyber-threats/">Jailbreak Rating Framework</a> that aims to track how easy it might be for a given AI model to get around safeguards. The framework also estimates the scale of the collateral damage, should an AI model wreak cyber-havoc across the internet.</p><p>The project started in 2023 with the support of four founding companies: Anthropic, Google, Microsoft, and OpenAI. The published Jailbreak Rating Framework is a rough analog to the well-known <a href="https://www.first.org/cvss/" target="_blank">CVSS security rating system,</a> but specifically targeted at classifying the type and severity of LLM jailbreaks across various axes and multiple frontier LLM providers.</p><p>Many AI companies and evaluation entities already have their own rating systems, but the FMF intends for its scale to be a cross-vendor standard that teams developing patches can rely on.</p><p>The FMF's approach is to have a classification method for jailbreaks that can make a model perform cyber-warfare operations beyond what's already out, across both the regular and black market. Version 1.0 of the proposal is a work-in-progress, and the group is taking comments and suggestions on it until December 21, noting that relying on its system in the meantime is inadvisable.</p><p>Right now, the framework has four scoring components, split across two main categories: severity, measuring how serious a jailbreak technique is, and likelihood, or how easy it is to find and weaponize it.</p><p>A combined final rating adds the scores obtained from the two categories into a final low-to-critical rating. The scientists note that the capability gain is the most important dimension and acts as both a floor and a ceiling for the remaining scores — a jailbreak that doesn't offer results beyond what can already be obtained via other means is meaningless within the purview of the grading system.</p><p>The individual ratings pertain to capability gains, capability scope, ease of deployment, and ease of discovery. They're all graded from zero to three, though the FMF stresses that a rating of zero doesn't mean a jailbreak is harmless or ought to be ignored — rather, it means the flaw should be addressed by standard practices like the platform's policies and standard cybersecurity ratings.</p><p>An important feature is that the FMF's baseline implicitly changes as models evolve, meaning that a grade of zero for a jailbreak in 2026 likely isn't going to mean the same thing as it would in 2028, or isn't better than alternative tools. A capability-gain rating of zero means that the jailbreak under analysis doesn't offer any improvement beyond what can already be done with existing models or tooling right off the bat.</p><p>For example, jailbreaking a model to perform a DDoS on a site isn't all that helpful to an attacker, since there are a number of black-market sites for that — or even other LLMs with fewer safeguards. This doesn't mean the jailbreak isn't handy, only that it's no better than the alternatives. However, the rating can be higher if the attacker can leverage the jailbreak to get a better DDoS than they could otherwise shop for.</p><p>On the other end of the capability-gain scale, three means that a jailbroken model can perform cyber-attacks with capabilities similar to those of a model operating without safeguards.</p><p>The capability scope measurement is easier to interpret: zero means that the jailbreak only works for one specific action or target and can't be tweaked for other applications, or in other words, is a one-trick pony. A jailbreak gets a rating of three when it can, without specific redesign, make a model perform work across at least two entirely unrelated <em>categories</em> of cyber-warfare — like getting the model to both autonomously discover vulnerabilities or, separately, write malware.</p><p>For ease of deployment, the FMF considers the amount of human interaction, or the degree of automation that the jailbreak allows. A zero means that the jailbreak must be steered by humans with a high degree of knowledge, providing additional instructions to be effective, while a three means that, with little effort, it can be used by a novice and results in an automated cyber-warfare task. </p><p>Last but by no means least, ease-of-discovery is how hard it is to find the jailbreak. At a zero, the attacker must spend copious amounts of money on inference tokens, need elevated access (like a privileged account given to a white-hat hacker), or specialized infrastructure like a legion of AI servers. A rating of three simply means that anyone can look up the technique and use it.</p><p>If all of this sounds familiar, you may be thinking of Anthropic's safeguard-breakage classifiers <a href="https://www.anthropic.com/news/fable-safeguards-jailbreak-framework">the company proposed</a> in July 2026, after an incident that led to Claude Fable 5 being temporarily unavailable. Much like the FMF's effort, Anthropic's system would also use a moving baseline and only consider jailbreaks that improve on cyber-warfare capabilities already available.</p><p>That document is strikingly similar to FMF's proposal. Anthropic is an FMF founding member, and as such may have had a hand or two in developing the new standard — though we have no confirmation either way. </p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/artificial-intelligence/new-framework-grades-ai-jailbreaks-across-capability-gains-scope-and-how-easy-they-are-to-deploy-frontier-model-forum-proposes-cvss-style-rating-system-for-new-ai-models</link>
                                                                            <description>
                            <![CDATA[ Frontier Model Forum releases its Jailbreak Rating Framework to estimate how easy it is for models to escape their cages and how much damage they can do. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">JHyamxoM5gXbySzbBzcm7V</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/PsE5GF3sUyv6c6FQPvP9j-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Tue, 10 Nov 2026 13:40:00 +0000</pubDate>                                                                                                                                <updated>Tue, 10 Nov 2026 13:44:49 +0000</updated>
                                                                                                                                            <category><![CDATA[Artificial Intelligence]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                <author><![CDATA[ editors@tomshardware.com (Bruno Ferreira) ]]></author>                    <dc:creator><![CDATA[ Bruno Ferreira ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/ZQiPPaXaAuQ4VrVEYnnR7G-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Bruno Ferreira&#039;s journey kicked off with the venerable ZX Spectrum, a cassette player, and his hopes and dreams. He quickly realized he had more fun figuring out how computers work than he did actually using the things. Kicking off a developer career with C and Assembly before moving to scripting languages, he&#039;s worn many hats, including both database architect and systems administration. As a teen, Bruno co-founded a web development outfit where he was for 17 years before moving on to spend nearly a decade at The Tech Report as a writer, editor, and (of course) developer. In this decade, he&#039;s been at Asus, MLCommons, and HotHardware, among others. When not fiddling with computers and games, his love for music and production sends him off to live shows and festivals. Occasionally, he pretends he can play the guitar and bass.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/PsE5GF3sUyv6c6FQPvP9j-1920-80.jpg">
                                                            <media:credit><![CDATA[Getty Images]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Render of a Chain ]]></media:description>                                                            <media:text><![CDATA[Render of a Chain ]]></media:text>
                                <media:title type="plain"><![CDATA[Render of a Chain ]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/PsE5GF3sUyv6c6FQPvP9j-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>In recent months, high-profile headlines about AI models letting loose from their safeguards dominated headlines. Now, the Frontier Model Forum (FMF) has released a new <a href="https://www.frontiermodelforum.org/updates/introducing-the-fmf-jailbreak-rating-framework-for-frontier-cyber-threats/">Jailbreak Rating Framework</a> that aims to track how easy it might be for a given AI model to get around safeguards. The framework also estimates the scale of the collateral damage, should an AI model wreak cyber-havoc across the internet.</p><p>The project started in 2023 with the support of four founding companies: Anthropic, Google, Microsoft, and OpenAI. The published Jailbreak Rating Framework is a rough analog to the well-known <a href="https://www.first.org/cvss/" target="_blank">CVSS security rating system,</a> but specifically targeted at classifying the type and severity of LLM jailbreaks across various axes and multiple frontier LLM providers.</p><p>Many AI companies and evaluation entities already have their own rating systems, but the FMF intends for its scale to be a cross-vendor standard that teams developing patches can rely on.</p><p>The FMF's approach is to have a classification method for jailbreaks that can make a model perform cyber-warfare operations beyond what's already out, across both the regular and black market. Version 1.0 of the proposal is a work-in-progress, and the group is taking comments and suggestions on it until December 21, noting that relying on its system in the meantime is inadvisable.</p><p>Right now, the framework has four scoring components, split across two main categories: severity, measuring how serious a jailbreak technique is, and likelihood, or how easy it is to find and weaponize it.</p><p>A combined final rating adds the scores obtained from the two categories into a final low-to-critical rating. The scientists note that the capability gain is the most important dimension and acts as both a floor and a ceiling for the remaining scores — a jailbreak that doesn't offer results beyond what can already be obtained via other means is meaningless within the purview of the grading system.</p><p>The individual ratings pertain to capability gains, capability scope, ease of deployment, and ease of discovery. They're all graded from zero to three, though the FMF stresses that a rating of zero doesn't mean a jailbreak is harmless or ought to be ignored — rather, it means the flaw should be addressed by standard practices like the platform's policies and standard cybersecurity ratings.</p><p>An important feature is that the FMF's baseline implicitly changes as models evolve, meaning that a grade of zero for a jailbreak in 2026 likely isn't going to mean the same thing as it would in 2028, or isn't better than alternative tools. A capability-gain rating of zero means that the jailbreak under analysis doesn't offer any improvement beyond what can already be done with existing models or tooling right off the bat.</p><p>For example, jailbreaking a model to perform a DDoS on a site isn't all that helpful to an attacker, since there are a number of black-market sites for that — or even other LLMs with fewer safeguards. This doesn't mean the jailbreak isn't handy, only that it's no better than the alternatives. However, the rating can be higher if the attacker can leverage the jailbreak to get a better DDoS than they could otherwise shop for.</p><p>On the other end of the capability-gain scale, three means that a jailbroken model can perform cyber-attacks with capabilities similar to those of a model operating without safeguards.</p><p>The capability scope measurement is easier to interpret: zero means that the jailbreak only works for one specific action or target and can't be tweaked for other applications, or in other words, is a one-trick pony. A jailbreak gets a rating of three when it can, without specific redesign, make a model perform work across at least two entirely unrelated <em>categories</em> of cyber-warfare — like getting the model to both autonomously discover vulnerabilities or, separately, write malware.</p><p>For ease of deployment, the FMF considers the amount of human interaction, or the degree of automation that the jailbreak allows. A zero means that the jailbreak must be steered by humans with a high degree of knowledge, providing additional instructions to be effective, while a three means that, with little effort, it can be used by a novice and results in an automated cyber-warfare task. </p><p>Last but by no means least, ease-of-discovery is how hard it is to find the jailbreak. At a zero, the attacker must spend copious amounts of money on inference tokens, need elevated access (like a privileged account given to a white-hat hacker), or specialized infrastructure like a legion of AI servers. A rating of three simply means that anyone can look up the technique and use it.</p><p>If all of this sounds familiar, you may be thinking of Anthropic's safeguard-breakage classifiers <a href="https://www.anthropic.com/news/fable-safeguards-jailbreak-framework">the company proposed</a> in July 2026, after an incident that led to Claude Fable 5 being temporarily unavailable. Much like the FMF's effort, Anthropic's system would also use a moving baseline and only consider jailbreaks that improve on cyber-warfare capabilities already available.</p><p>That document is strikingly similar to FMF's proposal. Anthropic is an FMF founding member, and as such may have had a hand or two in developing the new standard — though we have no confirmation either way. </p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Nvidia cracks down on overseas cloud AI GPU rentals for China ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Nvidia-based compute is getting scarce and expensive for Chinese AI labs on both of their main routes, buying export-compliant cards and renting abroad, and the <a href="https://www.ft.com/content/1902df70-1380-4588-bad5-171352b575f2"><em>Financial Times</em> reports</a> that Nvidia’s own crackdowns are part of the reason. Nvidia has told cloud partners worldwide that <a href="https://www.tomshardware.com/tech-industry/data-centers/nvidia-tells-data-center-operators-to-limit-business-from-chinese-customers-with-new-global-restrictions-report-claims-allegedly-restricted-to-49-percent-of-the-total-ensuring-51-percent-are-non-chinese-outfits">at least 51% of their services</a> must go to non-Chinese firms, while Chinese companies have bought more than 1 million Nvidia gaming and workstation cards this year to cover the shortfall. Meanwhile, a B300 rack that sells for about $600,000 internationally reportedly fetched $2.8 million in China last month.</p><p>Chinese customers have ordered more than 6 million of Nvidia’s export-compliant RTX Pro 5500 cards. That is at least twice the 3 million that the company could reportedly manufacture in 2027. One source told <em>CNBC</em> that it was proposed that Moonshot pay 30% to 50% more upfront, in comparison to a non-Chinese entity.</p><p>Many Chinese groups have relied on overseas leases for advanced AI work, which can include model training, and these tech groups have been the biggest clients for data centers in Southeast Asia. Data centers unable to find non-Chinese clients could be forced to pull back supply, and it’s “difficult to find enough non-Chinese clients outside the US,” one Singapore-based group told the <em>FT</em>. While Nvidia hasn’t confirmed or denied its new rule, implementing it would ease criticism from U.S. lawmakers just as Chinese demand is growing.</p><p>What Nvidia told partners has come through verbal instructions, via sales and regional teams, and applies globally rather than just in Asia. Existing deals and leases are not changed, but new ones must comply. Nvidia’s whitelist more than halved its authorized Asian buyer count, which encouraged more Chinese groups to lease over other options, including smuggling. On top of this, <a href="https://www.tomshardware.com/tech-industry/big-tech/nvidia-tightens-tracking-of-ai-gpus-in-southeast-asia-as-chinese-chip-smuggling-draws-scrutiny-report-claims-cloud-providers-must-join-monitoring-program-to-qualify-for-official-partner-status">Fleet Intelligence tracking</a> is now required for Southeast Asian providers that want to become Nvidia Cloud Partners. All of these combined — whitelist, cap, and tracking — may act to alleviate U.S. concerns.</p><p>Among companies that have moved to renting, Moonshot has held discussions to <a href="https://www.tomshardware.com/tech-industry/policy/chinese-ai-lab-moonshot-reportedly-looked-at-european-data-centers-to-access-nvidia-chips-report-claims-kimi-maker-primarily-interested-in-inferencing-capacity-but-also-asked-about-training-clusters">rent capacity from neoclouds that host in Europe</a> with Nvidia GPUs. Specifically, it’s looking for Nvidia Blackwell systems, which export controls bar from China; one source said the talks covered clusters of 8,000 or more Blackwell GPUs in Europe. These clusters would handle inference, serving models to end users, and the European talks for them happened after the White House said Moonshot <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/chinas-moonshot-ai-reportedly-used-nvidia-blackwell-chips-for-training-kimi-k3-company-circumvented-both-u-s-export-and-chinese-import-controls-to-acquire-compute">accessed Blackwell chips through a facility in Thailand</a>. The firm has been fervent about its intentions to secure more compute.</p><p>Moonshot isn't the only Chinese company looking to secure Nvidia chips for AI training. Tencent has a deal with Oracle to <a href="https://www.tomshardware.com/tech-industry/data-centers/tencent-scores-100-000-offshore-ai-chip-deal-with-oracle-for-usd7-billion-despite-climbing-prices-per-hour-costs-estimated-to-be-43-percent-under-standard-h100-rental-rates">lease 100,000 GPUs</a> in Southeast Asian data centers. ByteDance and Alibaba are also pursuing similar long-term leases, and most of these overseas data centers run chips like the forbidden B300. MiniMax, like Moonshot, also leases significant compute both at home and abroad.</p><p>The big reason is that domestic AI chip supply is simply not up to the task, although not for a lack of trying. Huawei is expected to hit 1.6 million chips this year, with about 1 million from Alibaba and Baidu, and 300,000 from Cambricon. Total supply this year will be up to 3.5 million, including about 1 million made overseas, against a demand that far exceeds 5 million. That number could more than double next year.</p><p>Part of the problem is that the <a href="https://www.tomshardware.com/pc-components/gpus/chinese-ai-companies-buy-1-million-nvidia-gpus-in-a-single-year-report-claims-export-control-compliant-chips-flood-the-market-to-bridge-gap-in-compute-demand">export-compliant cards Chinese firms have been buying</a>, mostly the RTX Pro 6000D, RTX Pro 5000, and RTX 5090, are mostly useful for simpler work such as prefill. Even so, the discontinued 6000D’s pricing in China has reportedly doubled. Huawei’s Ascend 910C outperforms them in most areas, but these lower-end options are still coveted, partly because of a preference for CUDA over Huawei’s CANN. Nvidia’s programming platform is supported by most AI frameworks and libraries, and much in-house code is optimized for it.</p><p>Then there’s the <a href="https://www.tomshardware.com/pc-components/gpus/gaming-takes-a-backseat-as-nvidia-overhauls-the-rtx-5090-for-maximum-ai-margins-rtx-pro-5500-delivers-2-6x-vram-at-matching-specs">RTX Pro 5500</a>, with 84GB of GDDR7 with ECC and 1,398 GB/s of memory bandwidth. The card runs on a full PCIe Gen 5 x16 interface, draws up to 600 W, and can be split into up to two MIG instances. That still leaves it well short of a B300, which has about 3.4 times the memory and 5.7 times the bandwidth, plus NVLink, which the 5500 lacks.</p><div ><table><caption>What Chinese labs can buy vs. what they rent</caption><tbody><tr><td class="firstcol " ><p>GPU</p></td><td  ><p>Memory</p></td><td  ><p>Bandwidth</p></td><td  ><p>Interconnect</p></td><td  ><p>Status for Chinese buyers</p></td></tr><tr><td class="firstcol " ><p>RTX Pro 5500</p></td><td  ><p>84GB GDDR7</p></td><td  ><p>1.4 TB/s</p></td><td  ><p>PCIe Gen 5 x16</p></td><td  ><p>Export-compliant; samples in December, deliveries from Q1 2027</p></td></tr><tr><td class="firstcol " ><p>H200</p></td><td  ><p>141GB HBM3E</p></td><td  ><p>4.8 TB/s</p></td><td  ><p>NVLink</p></td><td  ><p>Licensed by the U.S. case by case; Beijing withholding approval for large-scale imports</p></td></tr><tr><td class="firstcol " ><p>B300</p></td><td  ><p>288GB HBM3E</p></td><td  ><p>8 TB/s</p></td><td  ><p>NVLink (1.8 TB/s)</p></td><td  ><p>Banned for sale to China; rented abroad</p></td></tr></tbody></table></div><p>Indicated orders from China are not precise, but the total of over six million includes more than three million from ByteDance, Alibaba, and Tencent. The big three ordered at about $11,400 each, for a total sticker price over $35 billion. This is difficult to fill when there may be fewer than three million produced next year, and partners might only actually get 5% to 10% of the original ask. Starting in 2027, much of Nvidia’s TSMC capacity may shift to Rubin, taking more production, while the RTX Pro 5500’s early samples are due in December.</p><p>On the legal side, it’s currently easier to rent than to own as current U.S. law does not give regulators broad authority over overseas cloud access. The Remote Access Security Act (RASA), which would extend U.S. export controls to that area, has sailed through the House but has not yet passed the Senate. Those directly affected — neoclouds and related financiers — price in a possible ban regardless, which drives the upfront premium.</p><p>China’s side is not completely clean, either, with about 600,000 H200s reportedly stuck in storage in Hong Kong awaiting shipment to the mainland. Beijing is withholding approvals for large-volume imports such as this. Only relatively small quantities have come in during recent months, for some big internet companies, with others coming through intermediaries and indirect routes in small quantities. The price: over $900,000 for an eight-card H200 server in China at the end of last month, about 3x the cost a year ago. This is despite the fact that H200 sales are already under U.S. license.</p><p>The first test that will show how many cards will actually land occurs in Q1 2027, when the RTX Pro 5500 is expected in commercial deliveries. For rentals, whether or not the Remote Access Security Act is a factor depends on a Senate vote, which is not yet scheduled. Nvidia’s Q3 results come in soon, on Nov. 17, which could bring some illumination.</p><p>With CEO Jensen Huang under scrutiny lately over reports that he was <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-ceo-jensen-huang-admits-to-having-an-affair-with-suspected-chinese-spy-jensen-apologizes-in-letter-to-staff-company-board-forms-independent-committee-to-investigate">targeted by an alleged Chinese spy</a>, which Nvidia’s board is now investigating, he’s going to feel the heat even more just a day later, when lawmakers have asked him to attend a closed-door briefing. He’s been pushing for laxer controls and more access, so the assumed 51% rule will also be put to the test.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/policy/nvidia-cracks-down-on-overseas-cloud-ai-gpu-rentals-for-china-partners-face-51-percent-non-chinese-quota-as-black-market-b300-rack-prices-surge-to-usd2-8-million</link>
                                                                            <description>
                            <![CDATA[ Nvidia may cap cloud partners’ Chinese business at 49% as Chinese firms buy up more than 1 million gaming and workstation GPUs. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">zvygF64D8r4TjYHZWj6rzK</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/n3JSWEEYTE8ywDfC8rkMiS-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Mon, 09 Nov 2026 17:48:17 +0000</pubDate>                                                                                                                                <updated>Tue, 10 Nov 2026 13:26:47 +0000</updated>
                                                                                                                                            <category><![CDATA[Policy]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/n3JSWEEYTE8ywDfC8rkMiS-1920-80.jpg">
                                                            <media:credit><![CDATA[Nvidia]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Nvidia RTX Pro 5500 workstation graphics card with a ribbed black shroud and gold Nvidia logo, on a dark, wave-patterned background.]]></media:description>                                                            <media:text><![CDATA[Nvidia RTX Pro 5500 workstation graphics card with a ribbed black shroud and gold Nvidia logo, on a dark, wave-patterned background.]]></media:text>
                                <media:title type="plain"><![CDATA[Nvidia RTX Pro 5500 workstation graphics card with a ribbed black shroud and gold Nvidia logo, on a dark, wave-patterned background.]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/n3JSWEEYTE8ywDfC8rkMiS-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Nvidia-based compute is getting scarce and expensive for Chinese AI labs on both of their main routes, buying export-compliant cards and renting abroad, and the <a href="https://www.ft.com/content/1902df70-1380-4588-bad5-171352b575f2"><em>Financial Times</em> reports</a> that Nvidia’s own crackdowns are part of the reason. Nvidia has told cloud partners worldwide that <a href="https://www.tomshardware.com/tech-industry/data-centers/nvidia-tells-data-center-operators-to-limit-business-from-chinese-customers-with-new-global-restrictions-report-claims-allegedly-restricted-to-49-percent-of-the-total-ensuring-51-percent-are-non-chinese-outfits">at least 51% of their services</a> must go to non-Chinese firms, while Chinese companies have bought more than 1 million Nvidia gaming and workstation cards this year to cover the shortfall. Meanwhile, a B300 rack that sells for about $600,000 internationally reportedly fetched $2.8 million in China last month.</p><p>Chinese customers have ordered more than 6 million of Nvidia’s export-compliant RTX Pro 5500 cards. That is at least twice the 3 million that the company could reportedly manufacture in 2027. One source told <em>CNBC</em> that it was proposed that Moonshot pay 30% to 50% more upfront, in comparison to a non-Chinese entity.</p><p>Many Chinese groups have relied on overseas leases for advanced AI work, which can include model training, and these tech groups have been the biggest clients for data centers in Southeast Asia. Data centers unable to find non-Chinese clients could be forced to pull back supply, and it’s “difficult to find enough non-Chinese clients outside the US,” one Singapore-based group told the <em>FT</em>. While Nvidia hasn’t confirmed or denied its new rule, implementing it would ease criticism from U.S. lawmakers just as Chinese demand is growing.</p><p>What Nvidia told partners has come through verbal instructions, via sales and regional teams, and applies globally rather than just in Asia. Existing deals and leases are not changed, but new ones must comply. Nvidia’s whitelist more than halved its authorized Asian buyer count, which encouraged more Chinese groups to lease over other options, including smuggling. On top of this, <a href="https://www.tomshardware.com/tech-industry/big-tech/nvidia-tightens-tracking-of-ai-gpus-in-southeast-asia-as-chinese-chip-smuggling-draws-scrutiny-report-claims-cloud-providers-must-join-monitoring-program-to-qualify-for-official-partner-status">Fleet Intelligence tracking</a> is now required for Southeast Asian providers that want to become Nvidia Cloud Partners. All of these combined — whitelist, cap, and tracking — may act to alleviate U.S. concerns.</p><p>Among companies that have moved to renting, Moonshot has held discussions to <a href="https://www.tomshardware.com/tech-industry/policy/chinese-ai-lab-moonshot-reportedly-looked-at-european-data-centers-to-access-nvidia-chips-report-claims-kimi-maker-primarily-interested-in-inferencing-capacity-but-also-asked-about-training-clusters">rent capacity from neoclouds that host in Europe</a> with Nvidia GPUs. Specifically, it’s looking for Nvidia Blackwell systems, which export controls bar from China; one source said the talks covered clusters of 8,000 or more Blackwell GPUs in Europe. These clusters would handle inference, serving models to end users, and the European talks for them happened after the White House said Moonshot <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/chinas-moonshot-ai-reportedly-used-nvidia-blackwell-chips-for-training-kimi-k3-company-circumvented-both-u-s-export-and-chinese-import-controls-to-acquire-compute">accessed Blackwell chips through a facility in Thailand</a>. The firm has been fervent about its intentions to secure more compute.</p><p>Moonshot isn't the only Chinese company looking to secure Nvidia chips for AI training. Tencent has a deal with Oracle to <a href="https://www.tomshardware.com/tech-industry/data-centers/tencent-scores-100-000-offshore-ai-chip-deal-with-oracle-for-usd7-billion-despite-climbing-prices-per-hour-costs-estimated-to-be-43-percent-under-standard-h100-rental-rates">lease 100,000 GPUs</a> in Southeast Asian data centers. ByteDance and Alibaba are also pursuing similar long-term leases, and most of these overseas data centers run chips like the forbidden B300. MiniMax, like Moonshot, also leases significant compute both at home and abroad.</p><p>The big reason is that domestic AI chip supply is simply not up to the task, although not for a lack of trying. Huawei is expected to hit 1.6 million chips this year, with about 1 million from Alibaba and Baidu, and 300,000 from Cambricon. Total supply this year will be up to 3.5 million, including about 1 million made overseas, against a demand that far exceeds 5 million. That number could more than double next year.</p><p>Part of the problem is that the <a href="https://www.tomshardware.com/pc-components/gpus/chinese-ai-companies-buy-1-million-nvidia-gpus-in-a-single-year-report-claims-export-control-compliant-chips-flood-the-market-to-bridge-gap-in-compute-demand">export-compliant cards Chinese firms have been buying</a>, mostly the RTX Pro 6000D, RTX Pro 5000, and RTX 5090, are mostly useful for simpler work such as prefill. Even so, the discontinued 6000D’s pricing in China has reportedly doubled. Huawei’s Ascend 910C outperforms them in most areas, but these lower-end options are still coveted, partly because of a preference for CUDA over Huawei’s CANN. Nvidia’s programming platform is supported by most AI frameworks and libraries, and much in-house code is optimized for it.</p><p>Then there’s the <a href="https://www.tomshardware.com/pc-components/gpus/gaming-takes-a-backseat-as-nvidia-overhauls-the-rtx-5090-for-maximum-ai-margins-rtx-pro-5500-delivers-2-6x-vram-at-matching-specs">RTX Pro 5500</a>, with 84GB of GDDR7 with ECC and 1,398 GB/s of memory bandwidth. The card runs on a full PCIe Gen 5 x16 interface, draws up to 600 W, and can be split into up to two MIG instances. That still leaves it well short of a B300, which has about 3.4 times the memory and 5.7 times the bandwidth, plus NVLink, which the 5500 lacks.</p><div ><table><caption>What Chinese labs can buy vs. what they rent</caption><tbody><tr><td class="firstcol " ><p>GPU</p></td><td  ><p>Memory</p></td><td  ><p>Bandwidth</p></td><td  ><p>Interconnect</p></td><td  ><p>Status for Chinese buyers</p></td></tr><tr><td class="firstcol " ><p>RTX Pro 5500</p></td><td  ><p>84GB GDDR7</p></td><td  ><p>1.4 TB/s</p></td><td  ><p>PCIe Gen 5 x16</p></td><td  ><p>Export-compliant; samples in December, deliveries from Q1 2027</p></td></tr><tr><td class="firstcol " ><p>H200</p></td><td  ><p>141GB HBM3E</p></td><td  ><p>4.8 TB/s</p></td><td  ><p>NVLink</p></td><td  ><p>Licensed by the U.S. case by case; Beijing withholding approval for large-scale imports</p></td></tr><tr><td class="firstcol " ><p>B300</p></td><td  ><p>288GB HBM3E</p></td><td  ><p>8 TB/s</p></td><td  ><p>NVLink (1.8 TB/s)</p></td><td  ><p>Banned for sale to China; rented abroad</p></td></tr></tbody></table></div><p>Indicated orders from China are not precise, but the total of over six million includes more than three million from ByteDance, Alibaba, and Tencent. The big three ordered at about $11,400 each, for a total sticker price over $35 billion. This is difficult to fill when there may be fewer than three million produced next year, and partners might only actually get 5% to 10% of the original ask. Starting in 2027, much of Nvidia’s TSMC capacity may shift to Rubin, taking more production, while the RTX Pro 5500’s early samples are due in December.</p><p>On the legal side, it’s currently easier to rent than to own as current U.S. law does not give regulators broad authority over overseas cloud access. The Remote Access Security Act (RASA), which would extend U.S. export controls to that area, has sailed through the House but has not yet passed the Senate. Those directly affected — neoclouds and related financiers — price in a possible ban regardless, which drives the upfront premium.</p><p>China’s side is not completely clean, either, with about 600,000 H200s reportedly stuck in storage in Hong Kong awaiting shipment to the mainland. Beijing is withholding approvals for large-volume imports such as this. Only relatively small quantities have come in during recent months, for some big internet companies, with others coming through intermediaries and indirect routes in small quantities. The price: over $900,000 for an eight-card H200 server in China at the end of last month, about 3x the cost a year ago. This is despite the fact that H200 sales are already under U.S. license.</p><p>The first test that will show how many cards will actually land occurs in Q1 2027, when the RTX Pro 5500 is expected in commercial deliveries. For rentals, whether or not the Remote Access Security Act is a factor depends on a Senate vote, which is not yet scheduled. Nvidia’s Q3 results come in soon, on Nov. 17, which could bring some illumination.</p><p>With CEO Jensen Huang under scrutiny lately over reports that he was <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-ceo-jensen-huang-admits-to-having-an-affair-with-suspected-chinese-spy-jensen-apologizes-in-letter-to-staff-company-board-forms-independent-committee-to-investigate">targeted by an alleged Chinese spy</a>, which Nvidia’s board is now investigating, he’s going to feel the heat even more just a day later, when lawmakers have asked him to attend a closed-door briefing. He’s been pushing for laxer controls and more access, so the assumed 51% rule will also be put to the test.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Intel’s next-gen Nova Lake needs new motherboards — what we expect from Nova Point and Z990, Z970, and B960 chipsets ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Intel’s new <a href="https://www.tomshardware.com/pc-components/cpus/intels-core-ultra-400-nova-lake-launch-schedule-leaks-out-mass-production-in-q4-first-nova-lake-cpus-in-q1-2027"><u>Nova Lake-S CPUs are rumored to be on track to start mass production</u></a> in Q4 2026, with a launch expected in Q1 2027. Unlike previous launches, we’re supposed to see mid-range SKU(s), supposedly an unlocked 28-core product(s), first in Q1, followed later in 2027 by flagship-class models that allegedly use two compute tiles and feature up to 52 cores. </p><p><a href="https://www.tomshardware.com/pc-components/cpus/details-about-intels-next-gen-nova-lake-cpus-keep-leaking-an-attempt-to-establish-a-timeline-based-on-what-we-know-so-far"><u>Lip-Bu Tan, Intel CEO, said earlier this year</u></a> that Nova Lake is “coming at the end of 2026.” Despite the growing number of rumors and statements, we still don’t have a concrete date, but we can expect to see the CPUs soon, and with them, a brand-new socket and motherboard chipsets.<br /><br />Regardless of how many cores they have or when they arrive, you will <em>need</em> a new motherboard for Nova Lake. The LGA 1851 socket is at the end of its life with the release of Core Ultra Series 2 processors like the <a href="https://www.techpowerup.com/review/intel-core-ultra-7-270k-plus/"><u>Core Ultra 7 270K Plus</u></a>, released earlier in 2026; that means no more new processors and a dead end. A new socket, LGA 1954, is necessary because Nova Lake CPUs have a significant increase in I/O and power needs. It’s also Intel’s first bLLC (big last-level cache) intended to take on AMD’s X3D CPUs. </p><p>We haven’t seen many details about the boards themselves, like pricing, official configurations, or product stacks, and we've only caught glimpses of one or two early models at Computex. Still, we have seen a few interesting platform details and rumors. To that end, we’ve put together a table of what we know so far about the specifications. Some things aren’t official yet and are noted by the tilde (~). We’ve also listed last-gen equivalents for reference.</p><div ><table><tbody><tr><td class="firstcol " ><p><strong>900 Series Chipset, “up to”</strong></p></td><td  ><p><strong>B960</strong><br /><strong>(Mainstream)</strong></p></td><td  ><p>B860<br />(Mainstream)</p></td><td  ><p><strong>Z990</strong><br /><strong>(Enthusiast)</strong></p></td><td  ><p>Z890<br />(Enthusiast)</p></td><td  ><p><strong>Z970</strong><br /><strong>(Enthusiast)</strong></p></td></tr><tr><td class="firstcol " ><p><strong>Total PCIe Lanes</strong></p></td><td  ><p>34</p></td><td  ><p>34</p></td><td  ><p>48</p></td><td  ><p>48</p></td><td  ><p>34</p></td></tr><tr><td class="firstcol " ><p><strong>Processor TB4/USB4</strong></p></td><td  ><p>1<br />(TB5, 80Gbps)</p></td><td  ><p>1</p></td><td  ><p>2 <br />(TB5, 80Gbps)</p></td><td  ><p>2</p></td><td  ><p>1 (TB5, 80Gbps)</p></td></tr><tr><td class="firstcol " ><p><strong>DMI Gen5 Lanes</strong></p></td><td  ><p><strong>2</strong></p></td><td  ><p>0 (4x G4)</p></td><td  ><p><strong>4</strong></p></td><td  ><p>0 (8x G4)</p></td><td  ><p><strong>2</strong></p></td></tr><tr><td class="firstcol " ><p><strong>Chipset PCIE 5.0 Lanes</strong></p></td><td  ><p>0</p></td><td  ><p>0</p></td><td  ><p><strong>12</strong></p></td><td  ><p>0</p></td><td  ><p>0</p></td></tr><tr><td class="firstcol " ><p><strong>Chipset PCIe 4.0 Lanes</strong></p></td><td  ><p>14</p></td><td  ><p>14</p></td><td  ><p>12</p></td><td  ><p>24</p></td><td  ><p>14</p></td></tr><tr><td class="firstcol " ><p><strong>SATA 3.0 (6G Lanes)</strong></p></td><td  ><p>4</p></td><td  ><p>4</p></td><td  ><p>8</p></td><td  ><p>8</p></td><td  ><p>4</p></td></tr><tr><td class="firstcol " ><p><strong>USB2 Ports</strong></p></td><td  ><p>12</p></td><td  ><p>12</p></td><td  ><p>14</p></td><td  ><p>14</p></td><td  ><p>12</p></td></tr><tr><td class="firstcol " ><p><strong>USB3.2 (20G) Ports</strong></p></td><td  ><p>~2</p></td><td  ><p>2</p></td><td  ><p>~5</p></td><td  ><p>5</p></td><td  ><p>~2</p></td></tr><tr><td class="firstcol " ><p><strong>USB3.2 (10G) Ports</strong></p></td><td  ><p>~2</p></td><td  ><p>4</p></td><td  ><p>~10</p></td><td  ><p>10</p></td><td  ><p>~4</p></td></tr><tr><td class="firstcol " ><p><strong>USB3.2 (5G) Ports </strong></p></td><td  ><p>~6</p></td><td  ><p>6</p></td><td  ><p>~10</p></td><td  ><p>10</p></td><td  ><p>~6</p></td></tr><tr><td class="firstcol " ><p><strong>IA OC</strong></p></td><td  ><p>No</p></td><td  ><p>No</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td></tr><tr><td class="firstcol " ><p><strong>BCLK OC </strong></p></td><td  ><p>No</p></td><td  ><p>No (board discretion)</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>No</p></td></tr><tr><td class="firstcol " ><p><strong>Memory OC</strong></p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td></tr><tr><td class="firstcol " ><p><strong>Processor PCIE 5.0 Slot Lane Config </strong></p></td><td  ><p>1x16</p></td><td  ><p>1x16 + 1x4</p></td><td  ><p>1x16 or 1x8 + 2x4 or 2x8 or 4x4</p></td><td  ><p>1x16 + 1x4 or 2x8 + 1x4 or 1x8 or 3x4</p></td><td  ><p>1x16</p></td></tr><tr><td class="firstcol " ><p><strong>Processor PCIE 5.0 Storage Lane Config</strong></p></td><td  ><p>1x4</p></td><td  ><p>1x4</p></td><td  ><p><strong>1x8 or 2x4</strong></p></td><td  ><p>1x4</p></td><td  ><p>1x4</p></td></tr><tr><td class="firstcol " ><p><strong>ECC</strong></p></td><td  ><p>No</p></td><td  ><p>No</p></td><td  ><p>No</p></td><td  ><p>No</p></td><td  ><p>No</p></td></tr><tr><td class="firstcol " ><p><strong>Simultaneous Displays Supported</strong></p></td><td  ><p>4</p></td><td  ><p>4</p></td><td  ><p>4</p></td><td  ><p>4</p></td><td  ><p>4</p></td></tr><tr><td class="firstcol " ><p><strong>PCIE RAID 0/1/5/10 Support</strong></p></td><td  ><p>No</p></td><td  ><p>No</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>No</p></td></tr><tr><td class="firstcol " ><p><strong>SATA RAID 0/1/5/10 Support</strong></p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td></tr><tr><td class="firstcol " ><p><strong>Intel vPro + Standard Manageability</strong></p></td><td  ><p>No</p></td><td  ><p>No</p></td><td  ><p><strong>No</strong></p></td><td  ><p>Yes</p></td><td  ><p><strong>No</strong></p></td></tr></tbody></table></div><p>The new lineup consists of five different chipsets. You get Z990 and Z970 for the enthusiast class, B960 for the Mainstream, W980 for workstations, and Q970. You may notice it’s missing the ‘H910’ platform for entry-level PCs, and now we see a second enthusiast-class platform, the Z970, wedged between the B960 and Z990. Z990 is still the flagship, offering more PCIe lanes, a DMI Gen5 x4 link, and 12 Gen5 lanes from the chipset. Z970 is the less expensive enthusiast option and is reportedly built on the same silicon as B960. You get CPU multiplier overclocking, but no BCLK adjustment, and of course fewer ports and lanes than the flagship.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1920px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="hq4daWHYbpMUL4jYgLBS5F" name="chipsets" alt="Motherboard image" src="https://cdn.mos.cms.futurecdn.net/hq4daWHYbpMUL4jYgLBS5F-1920-80.jpg" mos="" align="middle" fullscreen="" width="1920" height="1080" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p>The chipset package itself, internally codenamed “Nova Point,” will also differ slightly and is smaller than Z890 (600 mm² vs 658 mm²). The die-on-package shrinks to 72.5 mm² (11.15 x 6.5mm) compared to 92.9 mm² (11.15 x 8.33 mm), or roughly 22% smaller. Base power increases to 7.9 Watts, up from 6 W on the Z890, with a maximum of 14 W while working. </p><p>That power-density increase explains Intel’s rumored active-cooling guidelines for the chipset. When you start connecting components that use the chipset, think some PCIe slots and M.2 drives, the chipset raises its working voltage to maintain signal integrity for all of the bandwidth it now has to provide for the attached devices. Hopefully this is only needed in the high-end Z990/970, where it has a lot more lanes/bandwidth to support, as some of the fans on the chipset (remember early X570?) could get loud. </p><h2 id="what-to-expect-from-nova-point-motherboards">What to expect from Nova Point motherboards</h2><p>Beyond the new LGA 1954 socket, not much information has been released about what to expect from motherboard partners. We know the platform's base memory speed is a smoking-fast DDR5-8000, up 800 MT/s over the Arrow Lake refresh chips. And if you don’t have a kit that can run those speeds and want to, it will cost at least <a href="https://www.newegg.com/p/pl?N=601410928%20600327642%20100007611&amp;d=DDR5-8000&amp;Order=1"><u>$700 from any reputable brand</u></a>. So the platform's cost of entry will be high unless you’re starting with compatible RAM. That said, you can run slower speeds if you already have DDR5, which would save money, but it could create a performance ceiling.</p><p>We’ll also see more abundant PCIe 5.0 connectivity. The upcoming Nova Lake CPUs are said to offer 36 PCIe Gen5 lanes. You get 16 lanes dedicated to graphics, eight lanes to storage, eight lanes to Thunderbolt 5 controllers (4x ea.), and four lanes for the DMI. That said, speeds are the same as the previous generation, since PCIe 4.0 x8 matches PCIe 5.0 x4. The chipset now has 12 dedicated PCIe 5.0 lanes, so motherboard manufacturers can include multiple PCIe 5.0 NVMe slots without sharing bandwidth from the primary PCIe slot or stealing USB4 bandwidth. In theory, you can have five PCIe 5.0 M.2 sockets between the CPU's native lanes and the chipset without sharing bandwidth among PCIe slots or M.2 sockets—a significant upgrade over Z890.</p><div ><table><tbody><tr><td class="firstcol " ><p><strong>Lane Category</strong></p></td><td  ><p><strong>Lane Count</strong></p></td><td  ><p><strong>Type</strong></p></td><td  ><p><strong>Functionality</strong></p></td></tr><tr><td class="firstcol " ><p><strong>Graphics</strong></p></td><td  ><p>16 Lanes</p></td><td  ><p>PCIe 5.0</p></td><td  ><p>Primary GPU Slot (x16)</p></td></tr><tr><td class="firstcol " ><p><strong>Storage</strong></p></td><td  ><p>8 Lanes</p></td><td  ><p>PCIe 5.0</p></td><td  ><p>Two dedicated M.2 NVMe sockets (4x ea.)</p></td></tr><tr><td class="firstcol " ><p><strong>Thunderbolt Reserved</strong></p></td><td  ><p>8 Lanes</p></td><td  ><p>PCIe 5.0</p></td><td  ><p>Dual Thunderbolt 5 Controllers (x4 ea.)</p></td></tr><tr><td class="firstcol " ><p><strong>DMI Reserved</strong></p></td><td  ><p>4 Lanes</p></td><td  ><p>PCIe 5.0</p></td><td  ><p>DMI 5.0 x4 Link to Chipset</p></td></tr><tr><td class="firstcol " ><p><strong>Total CPU Native</strong></p></td><td  ><p><strong>36 Lanes</strong></p></td><td  ><p><strong>Gen5</strong></p></td><td  ><p><strong>Total Lanes</strong></p></td></tr></tbody></table></div><p>Thunderbolt 5 (TB5) is another feature we’ll see more of on some of these new boards, offering baked-in support via a controller chip, typically the Intel JHL9580. We’ve seen at least one Z890 board (Asus ProArt) already using it, but it should be more prevalent in this generation. </p><p>TB5 speeds hit 80 Gbps (bidirectional) or up to a wicked-fast 120 Gbps in single-direction boost mode. TB5 also supports DisplayPort 2.1, USB4, and USB PD 3.1(100-130W depending). So for anyone who needs fast external storage, a board with TB5 plus an appropriately speedy M.2 external drive means transfers will be much faster.</p><figure role="gallery"><figure><img src="https://cdn.mos.cms.futurecdn.net/PvmBQTmWC6kcVqgQg3yB3F-1920-80.jpg" alt="Motherboard image" /><figcaption><small role="credit">Tom's Hardware</small></figcaption></figure><figure><img src="https://cdn.mos.cms.futurecdn.net/4qMsseDSFmgJJTsL2jce7F-1920-80.jpg" alt="Motherboard image" /><figcaption><small role="credit">Tom's Hardware</small></figcaption></figure></figure><p>On the networking side, expect another round of 2.5-5GbE on most boards, with the cheapest options opting for the less expensive 1GbE. Premium boards, on the other hand, will work in dual Ethernet connections and 10 GbE. We’ll see a lot of Realtek 8125/8126 (2.5/5GbE) chips, along with some Killer E3100G, and perhaps Killer’s E5000 mixed in, too. 10GbE will still rely on the Marvell AQtion, like the AQC113C.</p><p>Integrated Wi-Fi support will vary, with most SKUs opting to use Wi-Fi 7. You’ll see a smattering of the ‘slower’ 160 MHz Wi-Fi 7 which, in reality, is plenty fast and will not be a choke point for most users on less expensive boards. The faster 320 MHz widens the bandwidth highway from four lanes to eight, so you can move twice as much data simultaneously. We expect chips from MediaTek (MT7927), Intel (BE200/202), and Realtek (RTL9822AE) to dominate the Wi-Fi 7 scene for Nova Point boards. Just remember: to take advantage of Wi-Fi 7, you need a Wi-Fi 7-capable router in the first place.</p><h2 id="what-motherboards-can-we-expect-to-see">What motherboards can we expect to see?</h2><p>We’ve seen leaks from the Eurasian Economic Union database <a href="https://videocardz.com/newz/gigabyte-z990-motherboard-leak-reveals-24-models-for-intel-nova-lake-cpus" target="_blank">Gigabyte</a> (via Videocardz), listing a whopping 24 SKUs. We’ve also seen leaks about several ASRock motherboards via shipping records (<a href="https://www.club386.com/intel-z970-motherboards-shipping-nova-lake/" target="_blank">via Club386</a>). So far, there's not been a peep from MSI or Asus, but you can bet they’ll have a full complement of new boards of all shapes, sizes, and price points. But remember, just because something is listed in the EEU database doesn’t mean it will be released, or released in the U.S. So until we see an official list from the board partners, take this with a pinch of salt.</p><p>For Gigabyte, we see a lot of familiar faces in the list, including the flagship Aorus Xtreme, Aorus Master (now in three different SKUs), Aorus Pro and Elite (six total SKUs), and it even looks like they are planning multiple rear-connect Aorus Stealth options. We’ll also see several Aorus Elite and more affordable Eagle SKUs, along with more Duo X boards (this time from different families), all presumably using a 2-DIMM topology that supports high-density CQDIMMs. Micro ATX and Mini-ITX boards are also listed along with the wood-accent Aero models that made a splash this year.</p><p>ASRock is offering six different models, from the flagship-class Taichi (White) and Taichi Creator to the budget Pro RS Wifi and Z990 Challenger SKUs. In between, they have multiple variants of the Steel Legend Wi-Fi. What we don’t see from here are Micro ATX or Mini-ITX form factors, or any Z970 or B960 offerings out of the gate. We'll almost certainly be seeing those before long.</p><p>Asus has been tight-lipped, and as of this writing, we haven’t heard a peep about their boards. I’m confident we’ll see the usual ROG Maximus, ROG Strix, ProArt, TUF, and Prime lines (among others), with their typical Asus flair and value-add features like AI-assisted functionality (automating performance, optimizing cooling, managing network traffic, etc.). </p><p>MSI has also kept its cards close to the vest, and we haven’t seen much from them yet. But we do expect to see the company's lineup in full force. We expect to see the flagship Godlike board from the MEG series, high-end Ace, and mid-range MPG offerings like the Carbon and Edge, along with the MAG value series like the Tomahawk. I’d also imagine we’ll see some small-form-factor offerings, too, like a Z990 Edge Ti Wifi. In short, motherboard partners won’t miss the boat for a fresh revenue stream, and we’re sure to see a lot of options available come launch day.</p><p>Pricing for these future motherboards hasn't been listed. Still, you can imagine, with the increase in PCIe 5.0 bandwidth available and the copper needed to support it, hardware upgrades, and other external factors, it's safe to assume most SKUs will increase in price, or, with some luck, be the same as previous-gen versions. Either way, this is the start of a new platform, and those looking for a long-lasting socket should get in on the ground floor, assuming the performance is good, of course. </p><h2 id="final-thoughts">Final thoughts</h2><p>So who are these motherboards for, really? With the cost of video cards, storage, and, of course, RAM through the roof, upgrading to the new platform isn’t going to be cheap, especially if you want to get the most out of it. While RAM speeds tend not to make a significant difference in many cases, it can in others. So reaching the base memory speeds of this new platform will cost an arm and a leg. As mentioned earlier, it will run with existing RAM below that value. Still, performance losses will range from negligible to notable, depending on differences in speed and timings, as well as the work being executed and the application used. </p><p>Nova Point-based motherboards are shaping up as a meaningful step for Intel’s new platform: a new LGA 1954 socket, 12 PCIe 5.0 lanes attached to the chipset for Z990, Thunderbolt 5 support, and increased base memory speed to DDR5-8000. We even saw Intel add another enthusiast chipset to bridge the gap between enthusiast and mainstream in these trying times. Most of this information comes from leaks and rumors, and without a firm launch date from Intel, details and lineups could change before the Nova Lake CPUs arrive, supposedly in early 2027. We expect to see a lot of boards at CES 2027, with launches following soon after.</p><p>If the new CPUs are competitive with AMD's X3D offering, we can imagine those also costing more than the Arrow Lake refresh chips. If you already have a capable system, waiting to see the official lineups, performance, and pricing is the smarter move. But if you want the most PCIe 5.0 storage and the fastest external connectivity with TB5, this platform is built for you. Otherwise, you can likely make do with what you have for a little while longer, hopefully until inflated prices on at least some major components drop.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/pc-components/chipsets/intels-next-gen-nova-lake-needs-new-motherboards-what-we-expect-from-nova-point-and-z990-z970-and-b960-chipsets</link>
                                                                            <description>
                            <![CDATA[ New Intel CPUs mean a new LGA 1954 socket and new motherboards. Here’s what we know about Nova Point’s Z990, Z970, and B960 chipsets, including more PCIe 5.0, Thunderbolt 5, faster base RAM speeds, and active cooling ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">UCvzL2PCuGiKjsVqV7EyBF</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/3JguTaB6k8VyZTixTzb5tE-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Fri, 06 Nov 2026 12:00:00 +0000</pubDate>                                                                                                                                <updated>Fri, 06 Nov 2026 14:26:52 +0000</updated>
                                                                                                                                            <category><![CDATA[Chipsets]]></category>
                                                    <category><![CDATA[PC Components]]></category>
                                                    <category><![CDATA[Motherboards]]></category>
                                                                                                                    <dc:creator><![CDATA[ Joe Shields ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/tYLbbfsfgGWs5XBFcu3Dng-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Joe&#039;s tech journey began in the early 1980s with a Radio Shack Tandy TRS-80. After college, he built a custom PC, diving into modding, overclocking, and eventually extreme overclocking competitions at &lt;a href=&quot;http://Hwbot.org&quot; target=&quot;_blank&quot;&gt;Hwbot.org&lt;/a&gt;. Around 2010, he began writing news and reviews for &lt;a href=&quot;http://Overclockers.com&quot; target=&quot;_blank&quot;&gt;Overclockers.com&lt;/a&gt;, covering graphics cards, motherboards, storage, and processors. He went pro in 2018 at &lt;a href=&quot;http://AnandTech.com&quot; target=&quot;_blank&quot;&gt;AnandTech.com&lt;/a&gt; before joining Tom&#039;s Hardware as a Staff Writer, where he covers motherboards, cases, fans, chargers, deals, news, and more. When not benchmarking hardware or gathering data, Joe spends time with his wife, supports his two kids in high school athletics, loves to watch sports (Browns, Guardians, Cavaliers, and Buckeyes), or plays PUBG on PC after everyone else has gone to bed.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/3JguTaB6k8VyZTixTzb5tE-1920-80.jpg">
                                                            <media:credit><![CDATA[Tom&amp;#39;s Hardware]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Motherboard image]]></media:description>                                                            <media:text><![CDATA[Motherboard image]]></media:text>
                                <media:title type="plain"><![CDATA[Motherboard image]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/3JguTaB6k8VyZTixTzb5tE-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Intel’s new <a href="https://www.tomshardware.com/pc-components/cpus/intels-core-ultra-400-nova-lake-launch-schedule-leaks-out-mass-production-in-q4-first-nova-lake-cpus-in-q1-2027"><u>Nova Lake-S CPUs are rumored to be on track to start mass production</u></a> in Q4 2026, with a launch expected in Q1 2027. Unlike previous launches, we’re supposed to see mid-range SKU(s), supposedly an unlocked 28-core product(s), first in Q1, followed later in 2027 by flagship-class models that allegedly use two compute tiles and feature up to 52 cores. </p><p><a href="https://www.tomshardware.com/pc-components/cpus/details-about-intels-next-gen-nova-lake-cpus-keep-leaking-an-attempt-to-establish-a-timeline-based-on-what-we-know-so-far"><u>Lip-Bu Tan, Intel CEO, said earlier this year</u></a> that Nova Lake is “coming at the end of 2026.” Despite the growing number of rumors and statements, we still don’t have a concrete date, but we can expect to see the CPUs soon, and with them, a brand-new socket and motherboard chipsets.<br /><br />Regardless of how many cores they have or when they arrive, you will <em>need</em> a new motherboard for Nova Lake. The LGA 1851 socket is at the end of its life with the release of Core Ultra Series 2 processors like the <a href="https://www.techpowerup.com/review/intel-core-ultra-7-270k-plus/"><u>Core Ultra 7 270K Plus</u></a>, released earlier in 2026; that means no more new processors and a dead end. A new socket, LGA 1954, is necessary because Nova Lake CPUs have a significant increase in I/O and power needs. It’s also Intel’s first bLLC (big last-level cache) intended to take on AMD’s X3D CPUs. </p><p>We haven’t seen many details about the boards themselves, like pricing, official configurations, or product stacks, and we've only caught glimpses of one or two early models at Computex. Still, we have seen a few interesting platform details and rumors. To that end, we’ve put together a table of what we know so far about the specifications. Some things aren’t official yet and are noted by the tilde (~). We’ve also listed last-gen equivalents for reference.</p><div ><table><tbody><tr><td class="firstcol " ><p><strong>900 Series Chipset, “up to”</strong></p></td><td  ><p><strong>B960</strong><br /><strong>(Mainstream)</strong></p></td><td  ><p>B860<br />(Mainstream)</p></td><td  ><p><strong>Z990</strong><br /><strong>(Enthusiast)</strong></p></td><td  ><p>Z890<br />(Enthusiast)</p></td><td  ><p><strong>Z970</strong><br /><strong>(Enthusiast)</strong></p></td></tr><tr><td class="firstcol " ><p><strong>Total PCIe Lanes</strong></p></td><td  ><p>34</p></td><td  ><p>34</p></td><td  ><p>48</p></td><td  ><p>48</p></td><td  ><p>34</p></td></tr><tr><td class="firstcol " ><p><strong>Processor TB4/USB4</strong></p></td><td  ><p>1<br />(TB5, 80Gbps)</p></td><td  ><p>1</p></td><td  ><p>2 <br />(TB5, 80Gbps)</p></td><td  ><p>2</p></td><td  ><p>1 (TB5, 80Gbps)</p></td></tr><tr><td class="firstcol " ><p><strong>DMI Gen5 Lanes</strong></p></td><td  ><p><strong>2</strong></p></td><td  ><p>0 (4x G4)</p></td><td  ><p><strong>4</strong></p></td><td  ><p>0 (8x G4)</p></td><td  ><p><strong>2</strong></p></td></tr><tr><td class="firstcol " ><p><strong>Chipset PCIE 5.0 Lanes</strong></p></td><td  ><p>0</p></td><td  ><p>0</p></td><td  ><p><strong>12</strong></p></td><td  ><p>0</p></td><td  ><p>0</p></td></tr><tr><td class="firstcol " ><p><strong>Chipset PCIe 4.0 Lanes</strong></p></td><td  ><p>14</p></td><td  ><p>14</p></td><td  ><p>12</p></td><td  ><p>24</p></td><td  ><p>14</p></td></tr><tr><td class="firstcol " ><p><strong>SATA 3.0 (6G Lanes)</strong></p></td><td  ><p>4</p></td><td  ><p>4</p></td><td  ><p>8</p></td><td  ><p>8</p></td><td  ><p>4</p></td></tr><tr><td class="firstcol " ><p><strong>USB2 Ports</strong></p></td><td  ><p>12</p></td><td  ><p>12</p></td><td  ><p>14</p></td><td  ><p>14</p></td><td  ><p>12</p></td></tr><tr><td class="firstcol " ><p><strong>USB3.2 (20G) Ports</strong></p></td><td  ><p>~2</p></td><td  ><p>2</p></td><td  ><p>~5</p></td><td  ><p>5</p></td><td  ><p>~2</p></td></tr><tr><td class="firstcol " ><p><strong>USB3.2 (10G) Ports</strong></p></td><td  ><p>~2</p></td><td  ><p>4</p></td><td  ><p>~10</p></td><td  ><p>10</p></td><td  ><p>~4</p></td></tr><tr><td class="firstcol " ><p><strong>USB3.2 (5G) Ports </strong></p></td><td  ><p>~6</p></td><td  ><p>6</p></td><td  ><p>~10</p></td><td  ><p>10</p></td><td  ><p>~6</p></td></tr><tr><td class="firstcol " ><p><strong>IA OC</strong></p></td><td  ><p>No</p></td><td  ><p>No</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td></tr><tr><td class="firstcol " ><p><strong>BCLK OC </strong></p></td><td  ><p>No</p></td><td  ><p>No (board discretion)</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>No</p></td></tr><tr><td class="firstcol " ><p><strong>Memory OC</strong></p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td></tr><tr><td class="firstcol " ><p><strong>Processor PCIE 5.0 Slot Lane Config </strong></p></td><td  ><p>1x16</p></td><td  ><p>1x16 + 1x4</p></td><td  ><p>1x16 or 1x8 + 2x4 or 2x8 or 4x4</p></td><td  ><p>1x16 + 1x4 or 2x8 + 1x4 or 1x8 or 3x4</p></td><td  ><p>1x16</p></td></tr><tr><td class="firstcol " ><p><strong>Processor PCIE 5.0 Storage Lane Config</strong></p></td><td  ><p>1x4</p></td><td  ><p>1x4</p></td><td  ><p><strong>1x8 or 2x4</strong></p></td><td  ><p>1x4</p></td><td  ><p>1x4</p></td></tr><tr><td class="firstcol " ><p><strong>ECC</strong></p></td><td  ><p>No</p></td><td  ><p>No</p></td><td  ><p>No</p></td><td  ><p>No</p></td><td  ><p>No</p></td></tr><tr><td class="firstcol " ><p><strong>Simultaneous Displays Supported</strong></p></td><td  ><p>4</p></td><td  ><p>4</p></td><td  ><p>4</p></td><td  ><p>4</p></td><td  ><p>4</p></td></tr><tr><td class="firstcol " ><p><strong>PCIE RAID 0/1/5/10 Support</strong></p></td><td  ><p>No</p></td><td  ><p>No</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>No</p></td></tr><tr><td class="firstcol " ><p><strong>SATA RAID 0/1/5/10 Support</strong></p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td><td  ><p>Yes</p></td></tr><tr><td class="firstcol " ><p><strong>Intel vPro + Standard Manageability</strong></p></td><td  ><p>No</p></td><td  ><p>No</p></td><td  ><p><strong>No</strong></p></td><td  ><p>Yes</p></td><td  ><p><strong>No</strong></p></td></tr></tbody></table></div><p>The new lineup consists of five different chipsets. You get Z990 and Z970 for the enthusiast class, B960 for the Mainstream, W980 for workstations, and Q970. You may notice it’s missing the ‘H910’ platform for entry-level PCs, and now we see a second enthusiast-class platform, the Z970, wedged between the B960 and Z990. Z990 is still the flagship, offering more PCIe lanes, a DMI Gen5 x4 link, and 12 Gen5 lanes from the chipset. Z970 is the less expensive enthusiast option and is reportedly built on the same silicon as B960. You get CPU multiplier overclocking, but no BCLK adjustment, and of course fewer ports and lanes than the flagship.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1920px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="hq4daWHYbpMUL4jYgLBS5F" name="chipsets" alt="Motherboard image" src="https://cdn.mos.cms.futurecdn.net/hq4daWHYbpMUL4jYgLBS5F-1920-80.jpg" mos="" align="middle" fullscreen="" width="1920" height="1080" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p>The chipset package itself, internally codenamed “Nova Point,” will also differ slightly and is smaller than Z890 (600 mm² vs 658 mm²). The die-on-package shrinks to 72.5 mm² (11.15 x 6.5mm) compared to 92.9 mm² (11.15 x 8.33 mm), or roughly 22% smaller. Base power increases to 7.9 Watts, up from 6 W on the Z890, with a maximum of 14 W while working. </p><p>That power-density increase explains Intel’s rumored active-cooling guidelines for the chipset. When you start connecting components that use the chipset, think some PCIe slots and M.2 drives, the chipset raises its working voltage to maintain signal integrity for all of the bandwidth it now has to provide for the attached devices. Hopefully this is only needed in the high-end Z990/970, where it has a lot more lanes/bandwidth to support, as some of the fans on the chipset (remember early X570?) could get loud. </p><h2 id="what-to-expect-from-nova-point-motherboards">What to expect from Nova Point motherboards</h2><p>Beyond the new LGA 1954 socket, not much information has been released about what to expect from motherboard partners. We know the platform's base memory speed is a smoking-fast DDR5-8000, up 800 MT/s over the Arrow Lake refresh chips. And if you don’t have a kit that can run those speeds and want to, it will cost at least <a href="https://www.newegg.com/p/pl?N=601410928%20600327642%20100007611&amp;d=DDR5-8000&amp;Order=1"><u>$700 from any reputable brand</u></a>. So the platform's cost of entry will be high unless you’re starting with compatible RAM. That said, you can run slower speeds if you already have DDR5, which would save money, but it could create a performance ceiling.</p><p>We’ll also see more abundant PCIe 5.0 connectivity. The upcoming Nova Lake CPUs are said to offer 36 PCIe Gen5 lanes. You get 16 lanes dedicated to graphics, eight lanes to storage, eight lanes to Thunderbolt 5 controllers (4x ea.), and four lanes for the DMI. That said, speeds are the same as the previous generation, since PCIe 4.0 x8 matches PCIe 5.0 x4. The chipset now has 12 dedicated PCIe 5.0 lanes, so motherboard manufacturers can include multiple PCIe 5.0 NVMe slots without sharing bandwidth from the primary PCIe slot or stealing USB4 bandwidth. In theory, you can have five PCIe 5.0 M.2 sockets between the CPU's native lanes and the chipset without sharing bandwidth among PCIe slots or M.2 sockets—a significant upgrade over Z890.</p><div ><table><tbody><tr><td class="firstcol " ><p><strong>Lane Category</strong></p></td><td  ><p><strong>Lane Count</strong></p></td><td  ><p><strong>Type</strong></p></td><td  ><p><strong>Functionality</strong></p></td></tr><tr><td class="firstcol " ><p><strong>Graphics</strong></p></td><td  ><p>16 Lanes</p></td><td  ><p>PCIe 5.0</p></td><td  ><p>Primary GPU Slot (x16)</p></td></tr><tr><td class="firstcol " ><p><strong>Storage</strong></p></td><td  ><p>8 Lanes</p></td><td  ><p>PCIe 5.0</p></td><td  ><p>Two dedicated M.2 NVMe sockets (4x ea.)</p></td></tr><tr><td class="firstcol " ><p><strong>Thunderbolt Reserved</strong></p></td><td  ><p>8 Lanes</p></td><td  ><p>PCIe 5.0</p></td><td  ><p>Dual Thunderbolt 5 Controllers (x4 ea.)</p></td></tr><tr><td class="firstcol " ><p><strong>DMI Reserved</strong></p></td><td  ><p>4 Lanes</p></td><td  ><p>PCIe 5.0</p></td><td  ><p>DMI 5.0 x4 Link to Chipset</p></td></tr><tr><td class="firstcol " ><p><strong>Total CPU Native</strong></p></td><td  ><p><strong>36 Lanes</strong></p></td><td  ><p><strong>Gen5</strong></p></td><td  ><p><strong>Total Lanes</strong></p></td></tr></tbody></table></div><p>Thunderbolt 5 (TB5) is another feature we’ll see more of on some of these new boards, offering baked-in support via a controller chip, typically the Intel JHL9580. We’ve seen at least one Z890 board (Asus ProArt) already using it, but it should be more prevalent in this generation. </p><p>TB5 speeds hit 80 Gbps (bidirectional) or up to a wicked-fast 120 Gbps in single-direction boost mode. TB5 also supports DisplayPort 2.1, USB4, and USB PD 3.1(100-130W depending). So for anyone who needs fast external storage, a board with TB5 plus an appropriately speedy M.2 external drive means transfers will be much faster.</p><figure role="gallery"><figure><img src="https://cdn.mos.cms.futurecdn.net/PvmBQTmWC6kcVqgQg3yB3F-1920-80.jpg" alt="Motherboard image" /><figcaption><small role="credit">Tom's Hardware</small></figcaption></figure><figure><img src="https://cdn.mos.cms.futurecdn.net/4qMsseDSFmgJJTsL2jce7F-1920-80.jpg" alt="Motherboard image" /><figcaption><small role="credit">Tom's Hardware</small></figcaption></figure></figure><p>On the networking side, expect another round of 2.5-5GbE on most boards, with the cheapest options opting for the less expensive 1GbE. Premium boards, on the other hand, will work in dual Ethernet connections and 10 GbE. We’ll see a lot of Realtek 8125/8126 (2.5/5GbE) chips, along with some Killer E3100G, and perhaps Killer’s E5000 mixed in, too. 10GbE will still rely on the Marvell AQtion, like the AQC113C.</p><p>Integrated Wi-Fi support will vary, with most SKUs opting to use Wi-Fi 7. You’ll see a smattering of the ‘slower’ 160 MHz Wi-Fi 7 which, in reality, is plenty fast and will not be a choke point for most users on less expensive boards. The faster 320 MHz widens the bandwidth highway from four lanes to eight, so you can move twice as much data simultaneously. We expect chips from MediaTek (MT7927), Intel (BE200/202), and Realtek (RTL9822AE) to dominate the Wi-Fi 7 scene for Nova Point boards. Just remember: to take advantage of Wi-Fi 7, you need a Wi-Fi 7-capable router in the first place.</p><h2 id="what-motherboards-can-we-expect-to-see">What motherboards can we expect to see?</h2><p>We’ve seen leaks from the Eurasian Economic Union database <a href="https://videocardz.com/newz/gigabyte-z990-motherboard-leak-reveals-24-models-for-intel-nova-lake-cpus" target="_blank">Gigabyte</a> (via Videocardz), listing a whopping 24 SKUs. We’ve also seen leaks about several ASRock motherboards via shipping records (<a href="https://www.club386.com/intel-z970-motherboards-shipping-nova-lake/" target="_blank">via Club386</a>). So far, there's not been a peep from MSI or Asus, but you can bet they’ll have a full complement of new boards of all shapes, sizes, and price points. But remember, just because something is listed in the EEU database doesn’t mean it will be released, or released in the U.S. So until we see an official list from the board partners, take this with a pinch of salt.</p><p>For Gigabyte, we see a lot of familiar faces in the list, including the flagship Aorus Xtreme, Aorus Master (now in three different SKUs), Aorus Pro and Elite (six total SKUs), and it even looks like they are planning multiple rear-connect Aorus Stealth options. We’ll also see several Aorus Elite and more affordable Eagle SKUs, along with more Duo X boards (this time from different families), all presumably using a 2-DIMM topology that supports high-density CQDIMMs. Micro ATX and Mini-ITX boards are also listed along with the wood-accent Aero models that made a splash this year.</p><p>ASRock is offering six different models, from the flagship-class Taichi (White) and Taichi Creator to the budget Pro RS Wifi and Z990 Challenger SKUs. In between, they have multiple variants of the Steel Legend Wi-Fi. What we don’t see from here are Micro ATX or Mini-ITX form factors, or any Z970 or B960 offerings out of the gate. We'll almost certainly be seeing those before long.</p><p>Asus has been tight-lipped, and as of this writing, we haven’t heard a peep about their boards. I’m confident we’ll see the usual ROG Maximus, ROG Strix, ProArt, TUF, and Prime lines (among others), with their typical Asus flair and value-add features like AI-assisted functionality (automating performance, optimizing cooling, managing network traffic, etc.). </p><p>MSI has also kept its cards close to the vest, and we haven’t seen much from them yet. But we do expect to see the company's lineup in full force. We expect to see the flagship Godlike board from the MEG series, high-end Ace, and mid-range MPG offerings like the Carbon and Edge, along with the MAG value series like the Tomahawk. I’d also imagine we’ll see some small-form-factor offerings, too, like a Z990 Edge Ti Wifi. In short, motherboard partners won’t miss the boat for a fresh revenue stream, and we’re sure to see a lot of options available come launch day.</p><p>Pricing for these future motherboards hasn't been listed. Still, you can imagine, with the increase in PCIe 5.0 bandwidth available and the copper needed to support it, hardware upgrades, and other external factors, it's safe to assume most SKUs will increase in price, or, with some luck, be the same as previous-gen versions. Either way, this is the start of a new platform, and those looking for a long-lasting socket should get in on the ground floor, assuming the performance is good, of course. </p><h2 id="final-thoughts">Final thoughts</h2><p>So who are these motherboards for, really? With the cost of video cards, storage, and, of course, RAM through the roof, upgrading to the new platform isn’t going to be cheap, especially if you want to get the most out of it. While RAM speeds tend not to make a significant difference in many cases, it can in others. So reaching the base memory speeds of this new platform will cost an arm and a leg. As mentioned earlier, it will run with existing RAM below that value. Still, performance losses will range from negligible to notable, depending on differences in speed and timings, as well as the work being executed and the application used. </p><p>Nova Point-based motherboards are shaping up as a meaningful step for Intel’s new platform: a new LGA 1954 socket, 12 PCIe 5.0 lanes attached to the chipset for Z990, Thunderbolt 5 support, and increased base memory speed to DDR5-8000. We even saw Intel add another enthusiast chipset to bridge the gap between enthusiast and mainstream in these trying times. Most of this information comes from leaks and rumors, and without a firm launch date from Intel, details and lineups could change before the Nova Lake CPUs arrive, supposedly in early 2027. We expect to see a lot of boards at CES 2027, with launches following soon after.</p><p>If the new CPUs are competitive with AMD's X3D offering, we can imagine those also costing more than the Arrow Lake refresh chips. If you already have a capable system, waiting to see the official lineups, performance, and pricing is the smarter move. But if you want the most PCIe 5.0 storage and the fastest external connectivity with TB5, this platform is built for you. Otherwise, you can likely make do with what you have for a little while longer, hopefully until inflated prices on at least some major components drop.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Gelsinger-backed Nagare Labs’ Z2RAM promises SRAM speed at DRAM density ]]></title>
                                                                                                <dc:content><![CDATA[ <p>A Pat Gelsinger-backed Georgia Tech spinout, Nagare Labs, <a href="https://www.businesswire.com/news/home/20261026158062/en/Nagare-Labs-Emerges-from-Stealth-to-Break-AIs-Memory-Bottleneck">says its Z2RAM can put SRAM-speed memory</a>, at DRAM density, directly on top of AI chips. The founders’ own papers show a fast-improving cell that, tested up to 85 degrees Celsius, holds data for over 300 seconds. Gen 1 is projected to have 5x SRAM density with six stacked layers, 14x DRAM speed, and 3x less read energy than SRAM.</p><p>AI chips require a lot of bandwidth, and it has been challenging to get sufficient yields of high-bandwidth memory (HBM) to feed them. Even when manufacturers can get it, some reportedly plan to settle for fewer layers for yield reasons, and heat is another. Heat is also why <a href="https://www.tomshardware.com/tech-industry/semiconductors/hot-chips-2026-samsung-reveals-a-three-phase-hbm-roadmap-that-puts-logic-and-compute-inside-memory-zhbm-ultimately-stacks-dram-directly-on-top-of-the-processor">Samsung’s zHBM</a>, which stacks DRAM directly on the processor, is planned at only about four layers.</p><p>Alternatives include <a href="https://www.tomshardware.com/pc-components/ssds/sandisk-and-sk-hynix-unveil-hbf-spec-up-to-16-hi-nand-stacks-3-tb-s-bandwidth-ucie">high-bandwidth flash</a>, still in development, which stacks NAND like HBM and would sit beside the GPU, or SRAM stacked directly on the die, which is less dense and more expensive for the capacity, but also faster. DRAM, which is more capacious, can sit on top of the compute, but heat limits how hard the logic underneath can run.</p><p>That limit is already shaping chipmakers’ design path for stacked memory. <a href="https://www.tomshardware.com/pc-components/dram/micron-plans-to-use-wafer-bonding-and-high-na-euv-for-next-gen-dram-company-also-developing-vertical-high-bandwidth-memory-and-vertical-edge-inference-memory-integrated-directly-on-top-of-processors">Micron’s vHBM</a> plans, for example, include heat-spreading structures. With HBM stacked on a logic die, and DRAM needing to be kept at or below 95 degrees Celsius, the logic’s heat output is capped at 220 W/cm², or 300 W/cm² with better thermal vias, according to a 2025 simulation by co-founder Shimeng Yu’s group. Nagare Labs says peak compute has grown 70% a year over the past three decades, against 20% for memory bandwidth.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1318px;"><p class="vanilla-image-block" style="padding-top:52.50%;"><img id="bgFchqfCLSBwD7teMfkzE5" name="TH-nagare-fig4-jxcdc-thermal" alt="Simulated temperature maps of HBM stacked on a logic die, with logic power held to 220 and 300 watts per square centimeter to keep the DRAM at 95°C or below." src="https://cdn.mos.cms.futurecdn.net/bgFchqfCLSBwD7teMfkzE5-1920-80.jpg" mos="" align="middle" fullscreen="" width="1318" height="692" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: J. Sharda et al., IEEE JxCDC (CC BY 4.0), cropped)</span></figcaption></figure><p>The company describes its Z2RAM as “Two transistors and a shield. No capacitor,” where a write transistor places a charge on a storage node and the shield blocks noisy neighbors. The read transistor can sense without draining. In a gain cell like this, the read transistor’s gate capacitance holds the charge, so reads are non-destructive and the cell avoids DRAM’s write-back. Nagare aims for well under a picojoule per bit.</p><p>Nagare builds Z2RAM by depositing its layers on top of the accelerator and patterning them “together on one die,” much as 3D NAND flash scales by adding layers. There’s no interposer and no bonds between the memory layers, which is part of the secret sauce. One lithography exposure covers several layers at once, and standard 300 mm tools with DUV do the job, with no EUV required. The current roadmap is aiming at over 100 layers with “hundreds of thousands” of vertical connections.</p><p>Retaining data at higher temperatures is the challenge, and one the founders have probed repeatedly with their testing. A 2020 paper from co-founder Suman Datta’s group, on a tungsten-doped indium oxide two-transistor cell, had retention of 1 second at 25 degrees Celsius and just 300 milliseconds at 85°C. An update in 2025, which has Datta as a co-author, adds a third transistor and removes the gate overlap from its transistors, managing over 1,000 seconds at 25°C, up from 200 seconds with overlapped ones. It also simulated a GPU L2 at twice SRAM’s capacity but in the same area with 75–80% lower energy-delay-area product, a combined score that multiplies energy, delay and area.</p><p>Six months later, both founders showed a “folded asymmetric boosted” three-transistor cell that held data for over 300 seconds at 85°C, as compared to only 100 milliseconds for a conventional three-transistor design. That’s still below the roughly 95°C that Nagare says DRAM must stay under, so neither the founders’ published cells nor Z2RAM have shown retention above that line yet. The boosted cell is also “refresh-free,” which means data in the GPU’s last-level cache, profiled from AI workloads, doesn’t have time to decay before it is overwritten.</p><div ><table><caption>What the founders’ papers measured and simulated, and what Nagare projects</caption><thead><tr><th class="firstcol " ><p>Figure</p></th><th  ><p>Value</p></th><th  ><p>Measured, simulated or projected</p></th><th  ><p>Conditions</p></th><th  ><p>Source</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>2020 two-transistor cell, retention</p></td><td  ><p>1 s; 300 ms</p></td><td  ><p>Measured</p></td><td  ><p>25°C; 85°C</p></td><td  ><p>Datta’s group, IEDM 2020</p></td></tr><tr><td class="firstcol " ><p>2025 three-transistor cell (self-aligned), retention</p></td><td  ><p>More than 1,000 s</p></td><td  ><p>Measured</p></td><td  ><p>25°C</p></td><td  ><p>Georgia Tech, IEDM 2025</p></td></tr><tr><td class="firstcol " ><p>Conventional three-transistor cell, retention</p></td><td  ><p>0.1 s</p></td><td  ><p>Measured</p></td><td  ><p>85°C, to a 100 mV decay</p></td><td  ><p>Georgia Tech, VLSI 2026</p></td></tr><tr><td class="firstcol " ><p>Boosted three-transistor cell, retention</p></td><td  ><p>More than 300 s</p></td><td  ><p>Measured</p></td><td  ><p>85°C, to a 100 mV decay</p></td><td  ><p>Georgia Tech, VLSI 2026</p></td></tr><tr><td class="firstcol " ><p>Write pulse</p></td><td  ><p>10 ns</p></td><td  ><p>Measured</p></td><td  ><p>Two-transistor cell</p></td><td  ><p>Georgia Tech, IEDM 2025</p></td></tr><tr><td class="firstcol " ><p>Read delay</p></td><td  ><p>2.6 ns</p></td><td  ><p>Simulated</p></td><td  ><p>7nm design, 128×128 array</p></td><td  ><p>Georgia Tech, VLSI 2026</p></td></tr><tr><td class="firstcol " ><p>Cell area</p></td><td  ><p>0.0168 µm²</p></td><td  ><p>Projected</p></td><td  ><p>7nm design rules</p></td><td  ><p>Georgia Tech, VLSI 2026</p></td></tr><tr><td class="firstcol " ><p>Bit density</p></td><td  ><p>5.7× that of 7nm SRAM</p></td><td  ><p>Simulated</p></td><td  ><p>Four tiers, 256MB, same area</p></td><td  ><p>Yu’s group, IEEE Trans. Computers 2026</p></td></tr><tr><td class="firstcol " ><p>GPU L2 energy-delay-area product</p></td><td  ><p>75–80% lower than SRAM</p></td><td  ><p>Simulated</p></td><td  ><p>Twice the capacity in the same area</p></td><td  ><p>Georgia Tech, IEDM 2025</p></td></tr><tr><td class="firstcol " ><p>Logic power under stacked DRAM</p></td><td  ><p>220 W/cm² (300 W/cm² with better thermal vias)</p></td><td  ><p>Simulated</p></td><td  ><p>DRAM kept at 95°C or below</p></td><td  ><p>Yu’s group, JxCDC 2025</p></td></tr><tr><td class="firstcol " ><p>Nagare Gen 1 density</p></td><td  ><p>5× SRAM</p></td><td  ><p>Projected</p></td><td  ><p>Six layers, vs. 2nm HD SRAM</p></td><td  ><p>Nagare Labs</p></td></tr><tr><td class="firstcol " ><p>Nagare Gen 1 speed</p></td><td  ><p>14× faster than DRAM</p></td><td  ><p>Projected</p></td><td  ><p>Gen 1</p></td><td  ><p>Nagare Labs</p></td></tr><tr><td class="firstcol " ><p>Nagare Gen 1 read energy</p></td><td  ><p>3× less than SRAM</p></td><td  ><p>Projected</p></td><td  ><p>Gen 1</p></td><td  ><p>Nagare Labs</p></td></tr></tbody></table></div><p>Physical space is also limited for this application, and the founders project their boosted cell at 0.0168 µm² at 7nm, compared with 0.0199 µm² for <a href="https://www.tomshardware.com/news/no-sram-scaling-implies-on-more-expensive-cpus-and-gpus">a high-density SRAM cell on TSMC’s 3nm (N3) process</a>. A separate study from Yu’s group found that a four-tier, 256MB cache of two-transistor gain cells at 7nm could reach 5.7x the bit density of 7nm SRAM in the same area, in simulation. However, the return on investment falls off past seven or eight tiers because masks scale with layers. Nagare Labs’ claim of single lithography would potentially bypass this limitation.</p><p>That has not yet been proven, and nothing published yet shows it. “Where we are is literally lab prototypes,” Datta told <em>The Register</em>, adding that sampling of Z2RAM tiles is targeted within four years and production shipments, if they arrive, within the next decade. Yu told the outlet that no standard yet exists for integrating this kind of memory with an accelerator, which means Nagare might have to come up with the interface itself. Datta also mentioned reliability to the outlet, as single-electron effects can shift a cell’s threshold enough to upset the stored bit. It also remains to be seen if the tested three-transistor cells work sufficiently well as a two-transistor, single-shield product.</p><p>Even if Nagare clears those hurdles, a more important one lingers: It currently has no fab. The plan to scale depends on getting partners for fabrication as well as packaging. Nagare is reportedly looking for an R&amp;D fab partner so it can scale beyond the university clean room. Perhaps the most similar type of memory-on-compute shipping today, and the one Datta cited to the outlet as a model for working with chipmakers, is <a href="https://www.tomshardware.com/news/amd-unveils-more-ryzen-3d-packaging-and-v-cache-details-at-hot-chips">AMD’s 3D V-Cache</a>.</p><p>Z2RAM isn’t the first capacitorless memory to promise a 5x improvement over SRAM in density; it follows in the footsteps of <a href="https://www.tomshardware.com/news/zram-gen2,3909.html">Innovative Silicon’s Z-RAM</a> from 20 years ago, a technology AMD licensed, though it was not pursued. A Georgia Tech paper due at IEDM in December covers a different device, a one-transistor, one-ferroelectric-transistor cell, aimed at the same gap between SRAM latency and DRAM density.</p><p>Gelsinger, previously CEO of Intel, which started as a memory company, <a href="https://www.tomshardware.com/news/intel-kills-optane-memory-business-for-good">phased out Optane</a>, a low-latency non-volatile memory, in favor of CXL. Optane sought to bridge the gap between slower non-volatile storage, such as HDDs and standard SSDs, and DRAM. Now he sits on Nagare’s board, and Playground Global, where he is a general partner, led the startup’s $5 million round as Nagare searches for a solution to a different but related problem.</p><p>Nagare declined to share measured retention, latency, or energy figures. The founders’ boosted cell already holds data at 85°C longer than a cache needs, but until a fab-built, multi-layer Z2RAM stack is measured, retention above 95°C included, its headline numbers remain projections.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/semiconductors/gelsinger-backed-nagare-labs-capacitorless-z2ram-promises-sram-speed-with-dram-density-latest-published-cell-retains-data-for-over-300-seconds-at-85-c-but-further-figures-are-projected</link>
                                                                            <description>
                            <![CDATA[ Georgia Tech spinout Nagare Labs, backed by Playground Global, unveils Z2RAM, a capacitorless memory to stack on AI accelerators. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">bAUGjhovmVGeing2C3GjFY</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/wBsiTVdEXUvLhGduqG9Q2d-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Fri, 06 Nov 2026 11:00:00 +0000</pubDate>                                                                                                                                <updated>Fri, 06 Nov 2026 14:26:52 +0000</updated>
                                                                                                                                            <category><![CDATA[Semiconductors]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                    <category><![CDATA[Manufacturing]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/wBsiTVdEXUvLhGduqG9Q2d-1920-80.jpg">
                                                            <media:credit><![CDATA[Nagare Labs]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Shimeng Yu and Suman Datta, co-founders of Nagare Labs, stand beside a probe station and monitors in a lab.]]></media:description>                                                            <media:text><![CDATA[Shimeng Yu and Suman Datta, co-founders of Nagare Labs, stand beside a probe station and monitors in a lab.]]></media:text>
                                <media:title type="plain"><![CDATA[Shimeng Yu and Suman Datta, co-founders of Nagare Labs, stand beside a probe station and monitors in a lab.]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/wBsiTVdEXUvLhGduqG9Q2d-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>A Pat Gelsinger-backed Georgia Tech spinout, Nagare Labs, <a href="https://www.businesswire.com/news/home/20261026158062/en/Nagare-Labs-Emerges-from-Stealth-to-Break-AIs-Memory-Bottleneck">says its Z2RAM can put SRAM-speed memory</a>, at DRAM density, directly on top of AI chips. The founders’ own papers show a fast-improving cell that, tested up to 85 degrees Celsius, holds data for over 300 seconds. Gen 1 is projected to have 5x SRAM density with six stacked layers, 14x DRAM speed, and 3x less read energy than SRAM.</p><p>AI chips require a lot of bandwidth, and it has been challenging to get sufficient yields of high-bandwidth memory (HBM) to feed them. Even when manufacturers can get it, some reportedly plan to settle for fewer layers for yield reasons, and heat is another. Heat is also why <a href="https://www.tomshardware.com/tech-industry/semiconductors/hot-chips-2026-samsung-reveals-a-three-phase-hbm-roadmap-that-puts-logic-and-compute-inside-memory-zhbm-ultimately-stacks-dram-directly-on-top-of-the-processor">Samsung’s zHBM</a>, which stacks DRAM directly on the processor, is planned at only about four layers.</p><p>Alternatives include <a href="https://www.tomshardware.com/pc-components/ssds/sandisk-and-sk-hynix-unveil-hbf-spec-up-to-16-hi-nand-stacks-3-tb-s-bandwidth-ucie">high-bandwidth flash</a>, still in development, which stacks NAND like HBM and would sit beside the GPU, or SRAM stacked directly on the die, which is less dense and more expensive for the capacity, but also faster. DRAM, which is more capacious, can sit on top of the compute, but heat limits how hard the logic underneath can run.</p><p>That limit is already shaping chipmakers’ design path for stacked memory. <a href="https://www.tomshardware.com/pc-components/dram/micron-plans-to-use-wafer-bonding-and-high-na-euv-for-next-gen-dram-company-also-developing-vertical-high-bandwidth-memory-and-vertical-edge-inference-memory-integrated-directly-on-top-of-processors">Micron’s vHBM</a> plans, for example, include heat-spreading structures. With HBM stacked on a logic die, and DRAM needing to be kept at or below 95 degrees Celsius, the logic’s heat output is capped at 220 W/cm², or 300 W/cm² with better thermal vias, according to a 2025 simulation by co-founder Shimeng Yu’s group. Nagare Labs says peak compute has grown 70% a year over the past three decades, against 20% for memory bandwidth.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1318px;"><p class="vanilla-image-block" style="padding-top:52.50%;"><img id="bgFchqfCLSBwD7teMfkzE5" name="TH-nagare-fig4-jxcdc-thermal" alt="Simulated temperature maps of HBM stacked on a logic die, with logic power held to 220 and 300 watts per square centimeter to keep the DRAM at 95°C or below." src="https://cdn.mos.cms.futurecdn.net/bgFchqfCLSBwD7teMfkzE5-1920-80.jpg" mos="" align="middle" fullscreen="" width="1318" height="692" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: J. Sharda et al., IEEE JxCDC (CC BY 4.0), cropped)</span></figcaption></figure><p>The company describes its Z2RAM as “Two transistors and a shield. No capacitor,” where a write transistor places a charge on a storage node and the shield blocks noisy neighbors. The read transistor can sense without draining. In a gain cell like this, the read transistor’s gate capacitance holds the charge, so reads are non-destructive and the cell avoids DRAM’s write-back. Nagare aims for well under a picojoule per bit.</p><p>Nagare builds Z2RAM by depositing its layers on top of the accelerator and patterning them “together on one die,” much as 3D NAND flash scales by adding layers. There’s no interposer and no bonds between the memory layers, which is part of the secret sauce. One lithography exposure covers several layers at once, and standard 300 mm tools with DUV do the job, with no EUV required. The current roadmap is aiming at over 100 layers with “hundreds of thousands” of vertical connections.</p><p>Retaining data at higher temperatures is the challenge, and one the founders have probed repeatedly with their testing. A 2020 paper from co-founder Suman Datta’s group, on a tungsten-doped indium oxide two-transistor cell, had retention of 1 second at 25 degrees Celsius and just 300 milliseconds at 85°C. An update in 2025, which has Datta as a co-author, adds a third transistor and removes the gate overlap from its transistors, managing over 1,000 seconds at 25°C, up from 200 seconds with overlapped ones. It also simulated a GPU L2 at twice SRAM’s capacity but in the same area with 75–80% lower energy-delay-area product, a combined score that multiplies energy, delay and area.</p><p>Six months later, both founders showed a “folded asymmetric boosted” three-transistor cell that held data for over 300 seconds at 85°C, as compared to only 100 milliseconds for a conventional three-transistor design. That’s still below the roughly 95°C that Nagare says DRAM must stay under, so neither the founders’ published cells nor Z2RAM have shown retention above that line yet. The boosted cell is also “refresh-free,” which means data in the GPU’s last-level cache, profiled from AI workloads, doesn’t have time to decay before it is overwritten.</p><div ><table><caption>What the founders’ papers measured and simulated, and what Nagare projects</caption><thead><tr><th class="firstcol " ><p>Figure</p></th><th  ><p>Value</p></th><th  ><p>Measured, simulated or projected</p></th><th  ><p>Conditions</p></th><th  ><p>Source</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>2020 two-transistor cell, retention</p></td><td  ><p>1 s; 300 ms</p></td><td  ><p>Measured</p></td><td  ><p>25°C; 85°C</p></td><td  ><p>Datta’s group, IEDM 2020</p></td></tr><tr><td class="firstcol " ><p>2025 three-transistor cell (self-aligned), retention</p></td><td  ><p>More than 1,000 s</p></td><td  ><p>Measured</p></td><td  ><p>25°C</p></td><td  ><p>Georgia Tech, IEDM 2025</p></td></tr><tr><td class="firstcol " ><p>Conventional three-transistor cell, retention</p></td><td  ><p>0.1 s</p></td><td  ><p>Measured</p></td><td  ><p>85°C, to a 100 mV decay</p></td><td  ><p>Georgia Tech, VLSI 2026</p></td></tr><tr><td class="firstcol " ><p>Boosted three-transistor cell, retention</p></td><td  ><p>More than 300 s</p></td><td  ><p>Measured</p></td><td  ><p>85°C, to a 100 mV decay</p></td><td  ><p>Georgia Tech, VLSI 2026</p></td></tr><tr><td class="firstcol " ><p>Write pulse</p></td><td  ><p>10 ns</p></td><td  ><p>Measured</p></td><td  ><p>Two-transistor cell</p></td><td  ><p>Georgia Tech, IEDM 2025</p></td></tr><tr><td class="firstcol " ><p>Read delay</p></td><td  ><p>2.6 ns</p></td><td  ><p>Simulated</p></td><td  ><p>7nm design, 128×128 array</p></td><td  ><p>Georgia Tech, VLSI 2026</p></td></tr><tr><td class="firstcol " ><p>Cell area</p></td><td  ><p>0.0168 µm²</p></td><td  ><p>Projected</p></td><td  ><p>7nm design rules</p></td><td  ><p>Georgia Tech, VLSI 2026</p></td></tr><tr><td class="firstcol " ><p>Bit density</p></td><td  ><p>5.7× that of 7nm SRAM</p></td><td  ><p>Simulated</p></td><td  ><p>Four tiers, 256MB, same area</p></td><td  ><p>Yu’s group, IEEE Trans. Computers 2026</p></td></tr><tr><td class="firstcol " ><p>GPU L2 energy-delay-area product</p></td><td  ><p>75–80% lower than SRAM</p></td><td  ><p>Simulated</p></td><td  ><p>Twice the capacity in the same area</p></td><td  ><p>Georgia Tech, IEDM 2025</p></td></tr><tr><td class="firstcol " ><p>Logic power under stacked DRAM</p></td><td  ><p>220 W/cm² (300 W/cm² with better thermal vias)</p></td><td  ><p>Simulated</p></td><td  ><p>DRAM kept at 95°C or below</p></td><td  ><p>Yu’s group, JxCDC 2025</p></td></tr><tr><td class="firstcol " ><p>Nagare Gen 1 density</p></td><td  ><p>5× SRAM</p></td><td  ><p>Projected</p></td><td  ><p>Six layers, vs. 2nm HD SRAM</p></td><td  ><p>Nagare Labs</p></td></tr><tr><td class="firstcol " ><p>Nagare Gen 1 speed</p></td><td  ><p>14× faster than DRAM</p></td><td  ><p>Projected</p></td><td  ><p>Gen 1</p></td><td  ><p>Nagare Labs</p></td></tr><tr><td class="firstcol " ><p>Nagare Gen 1 read energy</p></td><td  ><p>3× less than SRAM</p></td><td  ><p>Projected</p></td><td  ><p>Gen 1</p></td><td  ><p>Nagare Labs</p></td></tr></tbody></table></div><p>Physical space is also limited for this application, and the founders project their boosted cell at 0.0168 µm² at 7nm, compared with 0.0199 µm² for <a href="https://www.tomshardware.com/news/no-sram-scaling-implies-on-more-expensive-cpus-and-gpus">a high-density SRAM cell on TSMC’s 3nm (N3) process</a>. A separate study from Yu’s group found that a four-tier, 256MB cache of two-transistor gain cells at 7nm could reach 5.7x the bit density of 7nm SRAM in the same area, in simulation. However, the return on investment falls off past seven or eight tiers because masks scale with layers. Nagare Labs’ claim of single lithography would potentially bypass this limitation.</p><p>That has not yet been proven, and nothing published yet shows it. “Where we are is literally lab prototypes,” Datta told <em>The Register</em>, adding that sampling of Z2RAM tiles is targeted within four years and production shipments, if they arrive, within the next decade. Yu told the outlet that no standard yet exists for integrating this kind of memory with an accelerator, which means Nagare might have to come up with the interface itself. Datta also mentioned reliability to the outlet, as single-electron effects can shift a cell’s threshold enough to upset the stored bit. It also remains to be seen if the tested three-transistor cells work sufficiently well as a two-transistor, single-shield product.</p><p>Even if Nagare clears those hurdles, a more important one lingers: It currently has no fab. The plan to scale depends on getting partners for fabrication as well as packaging. Nagare is reportedly looking for an R&amp;D fab partner so it can scale beyond the university clean room. Perhaps the most similar type of memory-on-compute shipping today, and the one Datta cited to the outlet as a model for working with chipmakers, is <a href="https://www.tomshardware.com/news/amd-unveils-more-ryzen-3d-packaging-and-v-cache-details-at-hot-chips">AMD’s 3D V-Cache</a>.</p><p>Z2RAM isn’t the first capacitorless memory to promise a 5x improvement over SRAM in density; it follows in the footsteps of <a href="https://www.tomshardware.com/news/zram-gen2,3909.html">Innovative Silicon’s Z-RAM</a> from 20 years ago, a technology AMD licensed, though it was not pursued. A Georgia Tech paper due at IEDM in December covers a different device, a one-transistor, one-ferroelectric-transistor cell, aimed at the same gap between SRAM latency and DRAM density.</p><p>Gelsinger, previously CEO of Intel, which started as a memory company, <a href="https://www.tomshardware.com/news/intel-kills-optane-memory-business-for-good">phased out Optane</a>, a low-latency non-volatile memory, in favor of CXL. Optane sought to bridge the gap between slower non-volatile storage, such as HDDs and standard SSDs, and DRAM. Now he sits on Nagare’s board, and Playground Global, where he is a general partner, led the startup’s $5 million round as Nagare searches for a solution to a different but related problem.</p><p>Nagare declined to share measured retention, latency, or energy figures. The founders’ boosted cell already holds data at 85°C longer than a cache needs, but until a fab-built, multi-layer Z2RAM stack is measured, retention above 95°C included, its headline numbers remain projections.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Marvell to buy Lumentum in $115 billion all-stock deal ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Marvell Technology Inc. has signed a definitive merger agreement to acquire Lumentum for an enterprise value of $115 billion, combining two of the biggest players in the data center optics market into one company. According to a joint <a href="https://investor.marvell.com/news-events/press-releases/detail/1040/marvell-and-lumentum-to-combine-creating-a-category-leader-in-ai-connectivity" target="_blank">press release</a> on November 2nd, Lumentum shareholders will receive 4.5 Marvell shares for each share they own, which the companies put at about $1,132 per share. This figure represents a 23.9% premium over the company's unaffected closing price on October 28, the last trading day before reports of the deal affected prices.</p><p>The agreement — which leaves Marvell's current shareholders with about 67% of the combined company and Lumentum's with about 33% — is reportedly the second-largest tech transaction deal ever after <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/spacex-acquires-xai-in-a-bid-to-make-orbiting-data-centers-a-reality-musk-plans-to-launch-a-million-tons-of-satellites-annually-targets-1tw-year-of-space-based-compute-capacity" target="_blank">the SpaceX-xAI merger</a>, according to Bloomberg. Both boards have approved the deal unanimously, and it is expected to close in the first half of 2028, pending regulatory and shareholder approval.</p><p>The acquisition represents another significant expansion of Marvell's optical infrastructure portfolio, following its $10 billion acquisition of Inphi in 2021 and its <a href="https://www.tomshardware.com/tech-industry/marvells-celestial-ai-acquisition-expands-its-role-in-ai-data-center-hardware" target="_blank">$3.5 billion upfront purchase of optical interconnect startup Celestial AI</a>, completed in February 2026. All three mergers and acquisitions involved companies handling different parts of one of the most critical AI infrastructures: optical connectivity.</p><p>As AI infrastructure continues to scale, optical connectivity is becoming increasingly important for moving vast amounts of data between processors, racks, data centers, and even on the accelerator board itself. With the acquisition of Lumentum, Marvell, which already supplies optical DSPs, silicon photonics, and high-speed SerDes, is positioning itself as one of the industry's major suppliers of optical interconnect technologies,<a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/marvell-details-vision-of-optically-interconnected-data-centers-spanning-across-thousands-of-kilometers-new-interconnects-sampling-later-this-year-would-allow-csps-to-pool-resources-based-on-workload" target="_blank"> spanning scale-across and scale-out networking to emerging scale-up and scale-in connections</a>.</p><h2 id="why-ai-infrastructure-is-moving-towards-optical-connectivity">Why AI infrastructure is moving towards optical connectivity</h2><p>Modern AI data centers house thousands of interconnected accelerators exchanging enormous volumes of data. As data volume, speed, and travel distance scale, copper interconnects, the conventional medium for inter-rack connections, hit physical limits, becoming impractical at around one to two meters. <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/tech-titans-team-up-to-form-optical-interconnect-alliance-to-solve-the-ai-buildouts-big-data-bottleneck-nvidia-amd-broadcom-and-more-set-sights-on-building-phy-to-break-through-the-limitations-of-copper" target="_blank">The industry's solution has been a transition to optics</a>, moving data as light instead of electrical signals.</p><p>First, optical transceivers and fiber optics replaced copper for inter-rack connections and beyond. Pluggable transceivers containing a laser chip, digital signal processor (DSP) chips, and several optical components convert the accelerator's electrical signals into light for transmission over fiber, then convert them back into electrical signals at the receiving end.</p><p>The industry is also looking to bring optics into the accelerator package itself with <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/co-packaged-optics-cpo-foundry-roadmaps-breaking-down-tsmc-intel-samsung-and-globalfoundries-approach-to-next-generation-scale-up-connectivity" target="_blank">co-packaged optics</a> and <a href="https://www.tomshardware.com/tech-industry/near-packaged-optics-gains-ground-aso-the-industry-hedges-against-co-packaged-optics-growing-pains" target="_blank">near-packaged optics</a>. In conventional packages, signals from the logic chip have to travel inches along copper traces across the board to reach the transceiver. Co-packaged optics takes the optical engine out of the pluggable module and places it as a photonic integrated circuit (PIC) chip directly on the switch or accelerator package, shrinking the electrical path. </p><p>These shifts to optics have made <a href="https://www.tomshardware.com/tech-industry/photonics/ai-data-center-optical-interconnect-market-to-hit-usd144-billion-by-2030-an-over-ten-fold-increase-from-2024-figures-according-to-new-projections-silicon-photonics-expected-to-account-for-nearly-two-thirds-of-revenue-driven-by-co-packaged-optics" target="_blank">optical interconnects one of the most critical AI infrastructure markets</a>, driving several multi-billion-dollar deals and major partnerships among industry players across the different technologies and processes involved.</p><h2 id="from-silicon-to-lasers-what-marvell-and-lumentum-bring-to-the-table">From silicon to lasers: What Marvell and Lumentum bring to the table</h2><p>The optical interconnects market encompasses several critical technologies. On the electronics side, semiconductor chips such as high-speed serializer / deserializers (SerDes) and optical DSPs prepare and process electrical data signals. Electronics also include analog components such as drivers and transimpedance amplifiers (TIAs) that interface with the optical hardware. Photonic integrated circuits (PICs) use microscopic optical components such as waveguides, modulators, and photodetectors to manipulate light, encode information onto it, and convert received light back into electrical signals, combining all of this into a single chip.</p><p>Whether an optical engine sits inside a pluggable transceiver or directly on the board, it still needs a laser, either integrated into the assembly or supplied externally, which brings us to another specialized branch of semiconductor manufacturing: laser technology.  Lastly, fiber optic cables and associated passive optical components transmit and distribute light while <a href="https://www.tomshardware.com/tech-industry/photonics/psiquantum-turns-its-photonics-know-how-into-an-optical-switch-for-supercharging-ai-data-center-gpu-networks-8x8-lab-prototype-loses-less-light-than-rivals-with-plans-to-develop-larger-devices" target="_blank">optical circuit switches</a> redirect optical connections through a network.</p><p>Marvell already covers most of the electrical and electronics side, with a portfolio that includes custom AI accelerators, Ethernet switches, high-speed SerDes interfaces, DSPs, optical drivers, TIAs, and silicon photonics. Lumentum, on the other hand, supplies many optical components, including high-performance lasers, modulators, photodetectors, optical transceivers, and optical circuit switches. Combining these portfolios and processes allows Marvell to develop more complete solutions, designing and optimizing electrical and optical components as part of the same system.</p><p>Perhaps the most strategically important part of Lumentum's business is its <a href="https://www.tomshardware.com/tech-industry/semiconductors/lumentum-ceo-says-the-indium-phosphide-shortage-will-become-worse-than-memory" target="_blank">expertise in compound semiconductors, particularly indium phosphide (InP)</a>, which is critical for laser and other optical chips. Unlike conventional silicon, which has an indirect bandgap and is inefficient at generating light, InP-based semiconductor structures can efficiently emit and detect light, making the material well suited for manufacturing high-performance lasers and other optical components used in high-speed communication systems.</p><p>Lumentum is also bringing its manufacturing experience to the merger. Unlike Marvell's predominantly fabless semiconductor business, Lumentum directly controls its front-end semiconductor fabrication processes, owning and operating specialized InP fabrication facilities, which its CEO Michael Hurlston described as the world's largest InP manufacturing footprint during the acquisition announcement. If the deal closes in 2028, Marvell will transition into a hybrid model, inheriting Lumentum's internal factories while keeping its chip side outsourced. The combined company would also gain access to gallium arsenide (GaAs) technologies.</p><h2 id="a-115-billion-bet-and-a-long-road-to-2028">A $115 billion bet and a long road to 2028</h2><p>Over the years, Marvell — already one of the <a href="https://www.tomshardware.com/tech-industry/semiconductors/custom-ai-asics-examined-from-broadcom-to-mtia" target="_blank">biggest custom ASIC suppliers, with clients such as Google, AWS, Meta, and Microsoft</a> — has also strategically positioned itself as a major player in the optical interconnects market, supplying the largest optical transceiver manufacturers. In March 2026, <a href="https://www.tomshardware.com/tech-industry/nvidia-invests-2-billion-in-marvell-whose-biggest-clients-are-trying-to-replace-nvidia-chips" target="_blank">Nvidia invested $2 billion directly in Marvell</a> in a partnership that brings Marvell’s custom XPUs and scale-up networking into Nvidia’s NVLink Fusion ecosystem. </p><p>Marvell's biggest optics competitor, Broadcom, <a href="https://www.tomshardware.com/openai-broadcom-to-co-develop-10gw-of-custom-ai-chips" target="_blank">also supplies custom AI accelerators</a>, networking silicon, and optical connectivity components, offering both the electronic chips and the optical components. Acquiring Lumentum would give Marvell a more comprehensive portfolio similar to Broadcom's. <a href="https://www.tomshardware.com/tech-industry/nvidia-invests-usd4-billion-into-photonics-firms-in-a-bid-to-bolster-data-center-interconnect-supply-chains-lumentum-and-coherent-investment-to-fund-u-s-r-and-d-and-manufacturing-facilities-supports-capacity-rights-and-future-access" target="_blank">Nvidia’s separate $2 billion stake in Lumentum</a> would also convert into Marvell shares, consolidating its two $2 billion bets into a single holding.</p><p>A key question is how vertical integration affects existing customers, as Marvell supplies optical module manufacturers, while Lumentum sells optical components to companies that also compete with its transceiver business. Merging operations could create concerns among customers that depend on the companies for critical components. Marvell CEO Matt Murphy addressed this issue during the company's November 2 investor conference call, stating that the company intends to maintain its existing component supply model. Lumentum CEO Michael Hurlston also emphasized that supplying competitors is already an established part of the optical industry's business model.</p><p>Financially, Marvell expects the combined company’s total addressable market to reach about $550 billion by 2030 — roughly $60 billion more than the two companies’ standalone estimates combined. The companies also expect more than $1 billion in annual run-rate cost synergies, primarily from cost of goods sold. Combining Marvell’s current-quarter guidance with Lumentum’s latest results, the companies put the merged business at about $17.7 billion in annualized revenue. </p><p>According to Murphy, pairing Lumentum’s next-quarter guidance with analyst consensus for Marvell’s next quarter lifts that figure to more than $21 billion. Lumentum's recent performance shows the growing demand behind these projections. The company reported $1.26 billion in revenue for its first fiscal quarter of 2027, a 136.5% year-over-year increase. It also announced a second major near-packaged and co-packaged optics design win, which Hurlston said he expects to generate more than $5 billion in revenue.</p><p>However, the merger's huge valuation raises questions about whether future growth will justify the acquisition price. Investors reacted negatively to the announcement, with Marvell shares closing down 12.7% at $171.41 on November 2 and Lumentum falling 18.6% to $684.40. For now, the agreement is subject to a lengthy review, with an outside date of November 1, 2027, and options to extend to February 1, May 1, and finally August 1, 2028, if regulatory approvals are still pending. If the deal then fails on regulatory grounds alone, Marvell owes Lumentum $3.2 billion. However, if either board backed out or took a rival offer, Marvell would pay $5.8 billion, and Lumentum $2.9 billion.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/semiconductors/marvell-to-buy-lumentum-in-usd115-billion-all-stock-deal-combined-company-to-supply-scale-up-and-scale-in-optics-for-ai-data-centers</link>
                                                                            <description>
                            <![CDATA[ Marvell has agreed to acquire Lumentum for $115 billion, combining its optical DSPs, SerDes, silicon photonics, and networking silicon with Lumentum’s lasers, transceivers, optical switches, and InP manufacturing footprint. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">2ko8x8eWR4qgHN8rdkxZJP</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/udFp2ca8Nu9dvCJweZPZEL-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Thu, 05 Nov 2026 12:00:00 +0000</pubDate>                                                                                                                                <updated>Thu, 05 Nov 2026 18:38:25 +0000</updated>
                                                                                                                                            <category><![CDATA[Semiconductors]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                    <category><![CDATA[Manufacturing]]></category>
                                                                                                                    <dc:creator><![CDATA[ Etiido Uko ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/BBrMt7jWtSo2Dc3iKoroyD-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Etiido Uko is a mechanical engineer and senior technical writer with over nine years of experience in documentation and reporting. He is deeply passionate about all things engineering and technology, and is an expert in gadgets, manufacturing, robotics, automotive, and aerospace. His work spans content creation for industry leaders across multiple sectors, including Autodesk, Siemens, Xometry, Telus, and Coca-Cola. When he is not writing or keeping up with the latest innovations, you can find him exploring lands unknown. Check out more of his work at etiidowrites.com.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/udFp2ca8Nu9dvCJweZPZEL-1920-80.jpg">
                                                            <media:credit><![CDATA[Marvell]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Marvell building]]></media:description>                                                            <media:text><![CDATA[Marvell building]]></media:text>
                                <media:title type="plain"><![CDATA[Marvell building]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/udFp2ca8Nu9dvCJweZPZEL-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Marvell Technology Inc. has signed a definitive merger agreement to acquire Lumentum for an enterprise value of $115 billion, combining two of the biggest players in the data center optics market into one company. According to a joint <a href="https://investor.marvell.com/news-events/press-releases/detail/1040/marvell-and-lumentum-to-combine-creating-a-category-leader-in-ai-connectivity" target="_blank">press release</a> on November 2nd, Lumentum shareholders will receive 4.5 Marvell shares for each share they own, which the companies put at about $1,132 per share. This figure represents a 23.9% premium over the company's unaffected closing price on October 28, the last trading day before reports of the deal affected prices.</p><p>The agreement — which leaves Marvell's current shareholders with about 67% of the combined company and Lumentum's with about 33% — is reportedly the second-largest tech transaction deal ever after <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/spacex-acquires-xai-in-a-bid-to-make-orbiting-data-centers-a-reality-musk-plans-to-launch-a-million-tons-of-satellites-annually-targets-1tw-year-of-space-based-compute-capacity" target="_blank">the SpaceX-xAI merger</a>, according to Bloomberg. Both boards have approved the deal unanimously, and it is expected to close in the first half of 2028, pending regulatory and shareholder approval.</p><p>The acquisition represents another significant expansion of Marvell's optical infrastructure portfolio, following its $10 billion acquisition of Inphi in 2021 and its <a href="https://www.tomshardware.com/tech-industry/marvells-celestial-ai-acquisition-expands-its-role-in-ai-data-center-hardware" target="_blank">$3.5 billion upfront purchase of optical interconnect startup Celestial AI</a>, completed in February 2026. All three mergers and acquisitions involved companies handling different parts of one of the most critical AI infrastructures: optical connectivity.</p><p>As AI infrastructure continues to scale, optical connectivity is becoming increasingly important for moving vast amounts of data between processors, racks, data centers, and even on the accelerator board itself. With the acquisition of Lumentum, Marvell, which already supplies optical DSPs, silicon photonics, and high-speed SerDes, is positioning itself as one of the industry's major suppliers of optical interconnect technologies,<a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/marvell-details-vision-of-optically-interconnected-data-centers-spanning-across-thousands-of-kilometers-new-interconnects-sampling-later-this-year-would-allow-csps-to-pool-resources-based-on-workload" target="_blank"> spanning scale-across and scale-out networking to emerging scale-up and scale-in connections</a>.</p><h2 id="why-ai-infrastructure-is-moving-towards-optical-connectivity">Why AI infrastructure is moving towards optical connectivity</h2><p>Modern AI data centers house thousands of interconnected accelerators exchanging enormous volumes of data. As data volume, speed, and travel distance scale, copper interconnects, the conventional medium for inter-rack connections, hit physical limits, becoming impractical at around one to two meters. <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/tech-titans-team-up-to-form-optical-interconnect-alliance-to-solve-the-ai-buildouts-big-data-bottleneck-nvidia-amd-broadcom-and-more-set-sights-on-building-phy-to-break-through-the-limitations-of-copper" target="_blank">The industry's solution has been a transition to optics</a>, moving data as light instead of electrical signals.</p><p>First, optical transceivers and fiber optics replaced copper for inter-rack connections and beyond. Pluggable transceivers containing a laser chip, digital signal processor (DSP) chips, and several optical components convert the accelerator's electrical signals into light for transmission over fiber, then convert them back into electrical signals at the receiving end.</p><p>The industry is also looking to bring optics into the accelerator package itself with <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/co-packaged-optics-cpo-foundry-roadmaps-breaking-down-tsmc-intel-samsung-and-globalfoundries-approach-to-next-generation-scale-up-connectivity" target="_blank">co-packaged optics</a> and <a href="https://www.tomshardware.com/tech-industry/near-packaged-optics-gains-ground-aso-the-industry-hedges-against-co-packaged-optics-growing-pains" target="_blank">near-packaged optics</a>. In conventional packages, signals from the logic chip have to travel inches along copper traces across the board to reach the transceiver. Co-packaged optics takes the optical engine out of the pluggable module and places it as a photonic integrated circuit (PIC) chip directly on the switch or accelerator package, shrinking the electrical path. </p><p>These shifts to optics have made <a href="https://www.tomshardware.com/tech-industry/photonics/ai-data-center-optical-interconnect-market-to-hit-usd144-billion-by-2030-an-over-ten-fold-increase-from-2024-figures-according-to-new-projections-silicon-photonics-expected-to-account-for-nearly-two-thirds-of-revenue-driven-by-co-packaged-optics" target="_blank">optical interconnects one of the most critical AI infrastructure markets</a>, driving several multi-billion-dollar deals and major partnerships among industry players across the different technologies and processes involved.</p><h2 id="from-silicon-to-lasers-what-marvell-and-lumentum-bring-to-the-table">From silicon to lasers: What Marvell and Lumentum bring to the table</h2><p>The optical interconnects market encompasses several critical technologies. On the electronics side, semiconductor chips such as high-speed serializer / deserializers (SerDes) and optical DSPs prepare and process electrical data signals. Electronics also include analog components such as drivers and transimpedance amplifiers (TIAs) that interface with the optical hardware. Photonic integrated circuits (PICs) use microscopic optical components such as waveguides, modulators, and photodetectors to manipulate light, encode information onto it, and convert received light back into electrical signals, combining all of this into a single chip.</p><p>Whether an optical engine sits inside a pluggable transceiver or directly on the board, it still needs a laser, either integrated into the assembly or supplied externally, which brings us to another specialized branch of semiconductor manufacturing: laser technology.  Lastly, fiber optic cables and associated passive optical components transmit and distribute light while <a href="https://www.tomshardware.com/tech-industry/photonics/psiquantum-turns-its-photonics-know-how-into-an-optical-switch-for-supercharging-ai-data-center-gpu-networks-8x8-lab-prototype-loses-less-light-than-rivals-with-plans-to-develop-larger-devices" target="_blank">optical circuit switches</a> redirect optical connections through a network.</p><p>Marvell already covers most of the electrical and electronics side, with a portfolio that includes custom AI accelerators, Ethernet switches, high-speed SerDes interfaces, DSPs, optical drivers, TIAs, and silicon photonics. Lumentum, on the other hand, supplies many optical components, including high-performance lasers, modulators, photodetectors, optical transceivers, and optical circuit switches. Combining these portfolios and processes allows Marvell to develop more complete solutions, designing and optimizing electrical and optical components as part of the same system.</p><p>Perhaps the most strategically important part of Lumentum's business is its <a href="https://www.tomshardware.com/tech-industry/semiconductors/lumentum-ceo-says-the-indium-phosphide-shortage-will-become-worse-than-memory" target="_blank">expertise in compound semiconductors, particularly indium phosphide (InP)</a>, which is critical for laser and other optical chips. Unlike conventional silicon, which has an indirect bandgap and is inefficient at generating light, InP-based semiconductor structures can efficiently emit and detect light, making the material well suited for manufacturing high-performance lasers and other optical components used in high-speed communication systems.</p><p>Lumentum is also bringing its manufacturing experience to the merger. Unlike Marvell's predominantly fabless semiconductor business, Lumentum directly controls its front-end semiconductor fabrication processes, owning and operating specialized InP fabrication facilities, which its CEO Michael Hurlston described as the world's largest InP manufacturing footprint during the acquisition announcement. If the deal closes in 2028, Marvell will transition into a hybrid model, inheriting Lumentum's internal factories while keeping its chip side outsourced. The combined company would also gain access to gallium arsenide (GaAs) technologies.</p><h2 id="a-115-billion-bet-and-a-long-road-to-2028">A $115 billion bet and a long road to 2028</h2><p>Over the years, Marvell — already one of the <a href="https://www.tomshardware.com/tech-industry/semiconductors/custom-ai-asics-examined-from-broadcom-to-mtia" target="_blank">biggest custom ASIC suppliers, with clients such as Google, AWS, Meta, and Microsoft</a> — has also strategically positioned itself as a major player in the optical interconnects market, supplying the largest optical transceiver manufacturers. In March 2026, <a href="https://www.tomshardware.com/tech-industry/nvidia-invests-2-billion-in-marvell-whose-biggest-clients-are-trying-to-replace-nvidia-chips" target="_blank">Nvidia invested $2 billion directly in Marvell</a> in a partnership that brings Marvell’s custom XPUs and scale-up networking into Nvidia’s NVLink Fusion ecosystem. </p><p>Marvell's biggest optics competitor, Broadcom, <a href="https://www.tomshardware.com/openai-broadcom-to-co-develop-10gw-of-custom-ai-chips" target="_blank">also supplies custom AI accelerators</a>, networking silicon, and optical connectivity components, offering both the electronic chips and the optical components. Acquiring Lumentum would give Marvell a more comprehensive portfolio similar to Broadcom's. <a href="https://www.tomshardware.com/tech-industry/nvidia-invests-usd4-billion-into-photonics-firms-in-a-bid-to-bolster-data-center-interconnect-supply-chains-lumentum-and-coherent-investment-to-fund-u-s-r-and-d-and-manufacturing-facilities-supports-capacity-rights-and-future-access" target="_blank">Nvidia’s separate $2 billion stake in Lumentum</a> would also convert into Marvell shares, consolidating its two $2 billion bets into a single holding.</p><p>A key question is how vertical integration affects existing customers, as Marvell supplies optical module manufacturers, while Lumentum sells optical components to companies that also compete with its transceiver business. Merging operations could create concerns among customers that depend on the companies for critical components. Marvell CEO Matt Murphy addressed this issue during the company's November 2 investor conference call, stating that the company intends to maintain its existing component supply model. Lumentum CEO Michael Hurlston also emphasized that supplying competitors is already an established part of the optical industry's business model.</p><p>Financially, Marvell expects the combined company’s total addressable market to reach about $550 billion by 2030 — roughly $60 billion more than the two companies’ standalone estimates combined. The companies also expect more than $1 billion in annual run-rate cost synergies, primarily from cost of goods sold. Combining Marvell’s current-quarter guidance with Lumentum’s latest results, the companies put the merged business at about $17.7 billion in annualized revenue. </p><p>According to Murphy, pairing Lumentum’s next-quarter guidance with analyst consensus for Marvell’s next quarter lifts that figure to more than $21 billion. Lumentum's recent performance shows the growing demand behind these projections. The company reported $1.26 billion in revenue for its first fiscal quarter of 2027, a 136.5% year-over-year increase. It also announced a second major near-packaged and co-packaged optics design win, which Hurlston said he expects to generate more than $5 billion in revenue.</p><p>However, the merger's huge valuation raises questions about whether future growth will justify the acquisition price. Investors reacted negatively to the announcement, with Marvell shares closing down 12.7% at $171.41 on November 2 and Lumentum falling 18.6% to $684.40. For now, the agreement is subject to a lengthy review, with an outside date of November 1, 2027, and options to extend to February 1, May 1, and finally August 1, 2028, if regulatory approvals are still pending. If the deal then fails on regulatory grounds alone, Marvell owes Lumentum $3.2 billion. However, if either board backed out or took a rival offer, Marvell would pay $5.8 billion, and Lumentum $2.9 billion.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ PHP vulnerability post-mortem reveals bug was known two years ago ]]></title>
                                                                                                <dc:content><![CDATA[ <p>The phrase "when it rains, it pours" applies both to weather and computer security, as proven the week of October 19, 2026, when OpenAI's GPT 5.6-Sol discovered <a href="https://www.tomshardware.com/tech-industry/cyber-security/openai-says-gpt-5-6-sol-helped-uncover-critical-php-flaws-that-could-affect-70-percent-of-websites-ai-assisted-researchers-turned-two-year-old-bug-that-stumped-human-developers-into-working-exploits-in-two-weeks">critical vulnerabilities</a> in the open-source programming language PHP. As published in a post-mortem, a small team of maintainers knew of the issue two years ago. Security experts had warned for decades about mounting technical debt, as agentic AI exponentially sped up vulnerability scanning, to the point that the usual 90-day disclosure window for security exploits <a href="https://www.tomshardware.com/tech-industry/cyber-security/standard-90-day-vulnerability-disclosure-policy-is-likely-dead-thanks-to-ai-leaving-worlds-systems-exposed-to-zero-day-attacks-security-expert-details-how-llm-assisted-bug-hunting-ushers-in-a-new-cyberworld-orders">became pointless</a>. PHP, one of the internet's ubiquitous software packages, <a href="https://www.tomshardware.com/tech-industry/cyber-security/lack-of-funding-likely-led-to-the-recent-php-security-snafu-but-thats-just-the-tip-of-the-iceberg-a-dreary-situation-caused-by-services-running-old-versions-with-no-support">was unlucky enough to in the crosshairs of bug-hunting AI agents.</a></p><p>PHP hosts around 70% of websites on the internet, and the vulnerability allows for remote code execution with enough attacker knowledge. That combination is enough to sound alarms, but the situation got worse once the collective internet realized that, at the time, 73% of PHP installations online were running end-of-life versions with no active support — eliciting face-palms from many techies, and an exasperated sigh from anyone familiar with the ecosystem, myself included. So ultimately the problem wasn't really the vulnerability, but the general complacency.</p><h2 id="how-the-attack-came-to-be">How the attack came to be</h2><p>The root cause of the vulnerability was an ancient architectural decision in the PHP engine's error handling. When a PHP process runs out of memory, it'll call the longjmp() C function, which means "immediately jump the execution to another place, no questions asked." Using longjmp() without a very good reason is generally unadvisable, though that by itself wasn't the problem.</p><p>The real issue is that if that jump happens while building a data structure (like adding an array entry, or initializing an object), that half-baked data remains in the memory heap. When the PHP engine moves execution to its cleanup-end-exit mechanism, that code will execute expecting good, complete data, but instead acts on inconsistent leftover bytes.</p><p>With some clever injection of just the right content in a web request, this mechanism can be exploited to run any code with PHP's permissions, as shown by <a href="https://www.bitsight.com/blog/php2shell-scanning-exploitation-patching" target="_blank">the ongoing</a> PHP2Shell exploit.</p><p>Gaining a remote shell does require an attacker to know the <a href="https://www.tenable.com/blog/php2shell-one-week-later-why-one-exploit-doesnt-work-everywhere" target="_blank">exact version</a> and build of PHP used, so it's not as trivial as just pressing the "Hack" button. Yet the wide blast radius, with tens of millions of vulnerable servers, many of them running the same exact version, is enough to give anyone pause. After the vulnerability was made public, many sysadmins disabled version number reporting on their PHP installation, so as not to aid prospective hackers.</p><p>The PHP Foundation's post-mortem indicates that the team <a href="https://www.tomshardware.com/tech-industry/cyber-security/lack-of-funding-likely-led-to-the-recent-php-security-snafu-but-thats-just-the-tip-of-the-iceberg-a-dreary-situation-caused-by-services-running-old-versions-with-no-support">had known about the issue</a> for two years, but didn't immediately see it as a cause for concern, and went back and forth several times on the best way to fix the problem while maintaining compatibility with third-party extensions.</p><p>The team admitted it had made management mistakes, but it's a small outfit with one full-time employee and 14 part-time contractors, tasked with maintaining one of the most widespread pieces of software on the internet. This is a common pattern: open-source software that's of <a href="https://www.tomshardware.com/tech-industry/cyber-security/modern-software-is-built-on-the-backs-of-low-paid-maintainers-opening-the-door-to-potential-threats-report-details-critically-underfunded-projects-that-remain-understaffed-or-unfunded">critical importance to modern computing</a>, yet undermanned, underfunded, or both.</p><h2 id="the-exploit-is-only-half-of-the-problem">The exploit is only half of the problem</h2><p>The standard patch process is reasonably trivial, needing only a point update and perhaps requiring a restart of some services. The currently maintained branches of PHP are 8.1 through the current 8.5. Sysadmins rarely install PHP directly, though, relying on their Linux distribution of choice or community repositories for updated packages. </p><p>Those 73% of sites still stuck with PHP 7 are now stuck between a rock and a hard place. When PHP moved from 7.4 to 8.0 in late 2020, there were a number of changes to the language, most of them minor, but still requiring revision work. Many folks likely expected their distro to keep backporting fixes, especially for LTS (Long-Term Support) variants still in their support period, with PHP 7.4 in tow. That excuse is arguably always a little thin, as even before the AI age there were already automated tools to upgrade existing PHP code with relatively little work.</p><p>Now, depending on their specific situation, those relying on ancient versions like 7.3 or older may be forced to bring their websites' code up to speed, given that not all distros are bound to have backported fixes. For reference, PHP 7.3 came out in 2018. Making matters worse, nearly 8% of PHP installations are of version 5.x, which came out in <em>2005</em>. Plus, a number of network appliances, NAS boxes, and similar hardware carries PHP within, and that too will need updating. Palo Alto Networks <a href="https://security.paloaltonetworks.com/CVE-2026-109618" target="_blank">issued an advisory</a> for products powered by its PanOS system, for example.</p><p>It's not all doom and gloom, however. Since the PHP Foundation published source code patches for 7.4 up to 8.1, most anyone can devise a patch for the older versions. Many entities did so, and there are packages available <a href="https://packages.sury.org/" target="_blank">for Debian/Ubuntu</a> flavors from Ondřej Surý (Debian developer) covering PHP 5.6 to 8.1. CloudLinux and AlmaLinux have <a href="https://blog.cloudlinux.com/php-cve-2026-109618-and-cve-2026-109619" target="_blank">likewise produced</a> backported fixes.</p><p>This is one occasion where paying for extended support comes in handy, as commercial vendors like Ubuntu Pro and Zend <a href="https://ubuntu.com/blog/php2shell-vulnerability-fixes-available" target="_blank">released packages</a> for positively ancient PHP versions, down to 5.5. Old website code is likely to have vulnerabilities other than the PHP engine, so a wholesale update-and-upgrade is much preferable, but very few CTOs will ever listen to <em>that</em> song.</p><h2 id="other-options">Other options</h2><p>Given how broad the problem is, and how it'll be a while before every machine is protected, there are a number of other means to mitigate the issue — though the proper fix is to update to a patched PHP version. Web application firewalls (WAFs) like Immunify360 may <a href="https://blog.imunify360.com/php2shell-3-million-knocks-on-the-imunify360-fleet/" target="_blank">block malicious requests</a> from even getting to the server in the first place.</p><p>Content delivery networks (CDNs) like Cloudflare or Fastly universally offer options for adding a WAF to their setup. Both firms immediately <a href="https://blog.cloudflare.com/php-vulnerabilities-waf-protection/" target="_blank">issued</a> their <a href="https://www.fastly.com/blog/fastlys-proactive-protection-critical-php-rce-cve-2026-109618" target="_blank">respective notices</a> and added detection of PHP2Shell to their respective rulesets. Finally, server-level security suites <a href="https://www.crowdstrike.com/en-us/blog/crowdstrike-protects-customers-from-php-rce-zero-days/" target="_blank">like CrowdStrike</a> can also detect and block malicious activity by the PHP binary itself, in case an attacker tries something funny. It's reasonable to guess that these protections have slowed exploitation efforts.</p><h2 id="so-where-are-we">So where are we?</h2><p>Right now, the PHP2Shell exploit is doing the rounds, with at least <a href="https://dreamgroup.com/blog/chained-php2shell-exploit-roots-25-servers" target="_blank">one high-profile case</a> involving a campaign that chained the PHP vulnerability together with Fragnesia (aka <a href="https://www.tomshardware.com/tech-industry/cyber-security/dirty-frag-exploit-gets-root-on-most-linux-machines-since-2017-no-patches-available-no-warning-given-copy-fail-like-vulnerability-had-its-embargo-broken">Dirty Frag</a>) to eventually get <em>kernel-level access</em> to 25 servers seemingly linked to the U.S. government. As of October 29, security firm BitSight had recorded copious attempts of PHP2Shell intrusion across its monitoring fleet, from many distinct origins.</p><p>BitSight's dark web forum sleuths indicate that there was no one-size-fits-all exploit package available at that date. This may indicate that hacking campaigns for now are likely to be specifically targeted, but a proverbial one-click red button is bound to appear at some point, likely with a deeper exploit in tow, as with the Fragnesia example above.</p><p>Meanwhile, several cybersecurity firms kept a close eye on patching progress since the October 19 reveal. Censys has a <a href="https://php2shell.censys.com/">real-time view</a>, updated hourly, and suffice it to say that things could be going a lot better. The initial count estimated the number of vulnerable PHP installs at around 45.7 million, out of 119 million total.</p><p>The number of unpatched services has mercifully dropped to around 29 million as of this writing, seemingly indicating a patch rate of around 1 million per day, depending on how you count. It's worth noting that fix backports may not change the version number, so those instances could still count as unpatched. </p><p>But remember how most everyone toggled off their server's PHP version number reporting. According to Censys, 62% of PHP servers it's consulting aren't reporting that datum at all, so there's no telling what the actual totals might be. If history is any indication, a significant percentage of those machines remain unpatched today.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/cyber-security/php-vulnerability-post-mortem-reveals-bug-was-known-two-years-ago-small-team-of-maintainers-knew-of-the-issue-before-it-was-made-public-as-new-php2shell-chained-exploit-ravages-servers</link>
                                                                            <description>
                            <![CDATA[ PHP has a harvest season to remember — AI-automated bug hunting reaps what years of technical debt have sown. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">omvYbgiGwJ2aDCywNMpL8W</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/X86spL5fTiVTArGvMUbuZ5-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Wed, 04 Nov 2026 12:30:00 +0000</pubDate>                                                                                                                                <updated>Thu, 05 Nov 2026 14:01:56 +0000</updated>
                                                                                                                                            <category><![CDATA[Cybersecurity]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                <author><![CDATA[ editors@tomshardware.com (Bruno Ferreira) ]]></author>                    <dc:creator><![CDATA[ Bruno Ferreira ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/ZQiPPaXaAuQ4VrVEYnnR7G-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Bruno Ferreira&#039;s journey kicked off with the venerable ZX Spectrum, a cassette player, and his hopes and dreams. He quickly realized he had more fun figuring out how computers work than he did actually using the things. Kicking off a developer career with C and Assembly before moving to scripting languages, he&#039;s worn many hats, including both database architect and systems administration. As a teen, Bruno co-founded a web development outfit where he was for 17 years before moving on to spend nearly a decade at The Tech Report as a writer, editor, and (of course) developer. In this decade, he&#039;s been at Asus, MLCommons, and HotHardware, among others. When not fiddling with computers and games, his love for music and production sends him off to live shows and festivals. Occasionally, he pretends he can play the guitar and bass.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/X86spL5fTiVTArGvMUbuZ5-1920-80.jpg">
                                                            <media:credit><![CDATA[Getty Images]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[PHP website on a laptop ]]></media:description>                                                            <media:text><![CDATA[PHP website on a laptop ]]></media:text>
                                <media:title type="plain"><![CDATA[PHP website on a laptop ]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/X86spL5fTiVTArGvMUbuZ5-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>The phrase "when it rains, it pours" applies both to weather and computer security, as proven the week of October 19, 2026, when OpenAI's GPT 5.6-Sol discovered <a href="https://www.tomshardware.com/tech-industry/cyber-security/openai-says-gpt-5-6-sol-helped-uncover-critical-php-flaws-that-could-affect-70-percent-of-websites-ai-assisted-researchers-turned-two-year-old-bug-that-stumped-human-developers-into-working-exploits-in-two-weeks">critical vulnerabilities</a> in the open-source programming language PHP. As published in a post-mortem, a small team of maintainers knew of the issue two years ago. Security experts had warned for decades about mounting technical debt, as agentic AI exponentially sped up vulnerability scanning, to the point that the usual 90-day disclosure window for security exploits <a href="https://www.tomshardware.com/tech-industry/cyber-security/standard-90-day-vulnerability-disclosure-policy-is-likely-dead-thanks-to-ai-leaving-worlds-systems-exposed-to-zero-day-attacks-security-expert-details-how-llm-assisted-bug-hunting-ushers-in-a-new-cyberworld-orders">became pointless</a>. PHP, one of the internet's ubiquitous software packages, <a href="https://www.tomshardware.com/tech-industry/cyber-security/lack-of-funding-likely-led-to-the-recent-php-security-snafu-but-thats-just-the-tip-of-the-iceberg-a-dreary-situation-caused-by-services-running-old-versions-with-no-support">was unlucky enough to in the crosshairs of bug-hunting AI agents.</a></p><p>PHP hosts around 70% of websites on the internet, and the vulnerability allows for remote code execution with enough attacker knowledge. That combination is enough to sound alarms, but the situation got worse once the collective internet realized that, at the time, 73% of PHP installations online were running end-of-life versions with no active support — eliciting face-palms from many techies, and an exasperated sigh from anyone familiar with the ecosystem, myself included. So ultimately the problem wasn't really the vulnerability, but the general complacency.</p><h2 id="how-the-attack-came-to-be">How the attack came to be</h2><p>The root cause of the vulnerability was an ancient architectural decision in the PHP engine's error handling. When a PHP process runs out of memory, it'll call the longjmp() C function, which means "immediately jump the execution to another place, no questions asked." Using longjmp() without a very good reason is generally unadvisable, though that by itself wasn't the problem.</p><p>The real issue is that if that jump happens while building a data structure (like adding an array entry, or initializing an object), that half-baked data remains in the memory heap. When the PHP engine moves execution to its cleanup-end-exit mechanism, that code will execute expecting good, complete data, but instead acts on inconsistent leftover bytes.</p><p>With some clever injection of just the right content in a web request, this mechanism can be exploited to run any code with PHP's permissions, as shown by <a href="https://www.bitsight.com/blog/php2shell-scanning-exploitation-patching" target="_blank">the ongoing</a> PHP2Shell exploit.</p><p>Gaining a remote shell does require an attacker to know the <a href="https://www.tenable.com/blog/php2shell-one-week-later-why-one-exploit-doesnt-work-everywhere" target="_blank">exact version</a> and build of PHP used, so it's not as trivial as just pressing the "Hack" button. Yet the wide blast radius, with tens of millions of vulnerable servers, many of them running the same exact version, is enough to give anyone pause. After the vulnerability was made public, many sysadmins disabled version number reporting on their PHP installation, so as not to aid prospective hackers.</p><p>The PHP Foundation's post-mortem indicates that the team <a href="https://www.tomshardware.com/tech-industry/cyber-security/lack-of-funding-likely-led-to-the-recent-php-security-snafu-but-thats-just-the-tip-of-the-iceberg-a-dreary-situation-caused-by-services-running-old-versions-with-no-support">had known about the issue</a> for two years, but didn't immediately see it as a cause for concern, and went back and forth several times on the best way to fix the problem while maintaining compatibility with third-party extensions.</p><p>The team admitted it had made management mistakes, but it's a small outfit with one full-time employee and 14 part-time contractors, tasked with maintaining one of the most widespread pieces of software on the internet. This is a common pattern: open-source software that's of <a href="https://www.tomshardware.com/tech-industry/cyber-security/modern-software-is-built-on-the-backs-of-low-paid-maintainers-opening-the-door-to-potential-threats-report-details-critically-underfunded-projects-that-remain-understaffed-or-unfunded">critical importance to modern computing</a>, yet undermanned, underfunded, or both.</p><h2 id="the-exploit-is-only-half-of-the-problem">The exploit is only half of the problem</h2><p>The standard patch process is reasonably trivial, needing only a point update and perhaps requiring a restart of some services. The currently maintained branches of PHP are 8.1 through the current 8.5. Sysadmins rarely install PHP directly, though, relying on their Linux distribution of choice or community repositories for updated packages. </p><p>Those 73% of sites still stuck with PHP 7 are now stuck between a rock and a hard place. When PHP moved from 7.4 to 8.0 in late 2020, there were a number of changes to the language, most of them minor, but still requiring revision work. Many folks likely expected their distro to keep backporting fixes, especially for LTS (Long-Term Support) variants still in their support period, with PHP 7.4 in tow. That excuse is arguably always a little thin, as even before the AI age there were already automated tools to upgrade existing PHP code with relatively little work.</p><p>Now, depending on their specific situation, those relying on ancient versions like 7.3 or older may be forced to bring their websites' code up to speed, given that not all distros are bound to have backported fixes. For reference, PHP 7.3 came out in 2018. Making matters worse, nearly 8% of PHP installations are of version 5.x, which came out in <em>2005</em>. Plus, a number of network appliances, NAS boxes, and similar hardware carries PHP within, and that too will need updating. Palo Alto Networks <a href="https://security.paloaltonetworks.com/CVE-2026-109618" target="_blank">issued an advisory</a> for products powered by its PanOS system, for example.</p><p>It's not all doom and gloom, however. Since the PHP Foundation published source code patches for 7.4 up to 8.1, most anyone can devise a patch for the older versions. Many entities did so, and there are packages available <a href="https://packages.sury.org/" target="_blank">for Debian/Ubuntu</a> flavors from Ondřej Surý (Debian developer) covering PHP 5.6 to 8.1. CloudLinux and AlmaLinux have <a href="https://blog.cloudlinux.com/php-cve-2026-109618-and-cve-2026-109619" target="_blank">likewise produced</a> backported fixes.</p><p>This is one occasion where paying for extended support comes in handy, as commercial vendors like Ubuntu Pro and Zend <a href="https://ubuntu.com/blog/php2shell-vulnerability-fixes-available" target="_blank">released packages</a> for positively ancient PHP versions, down to 5.5. Old website code is likely to have vulnerabilities other than the PHP engine, so a wholesale update-and-upgrade is much preferable, but very few CTOs will ever listen to <em>that</em> song.</p><h2 id="other-options">Other options</h2><p>Given how broad the problem is, and how it'll be a while before every machine is protected, there are a number of other means to mitigate the issue — though the proper fix is to update to a patched PHP version. Web application firewalls (WAFs) like Immunify360 may <a href="https://blog.imunify360.com/php2shell-3-million-knocks-on-the-imunify360-fleet/" target="_blank">block malicious requests</a> from even getting to the server in the first place.</p><p>Content delivery networks (CDNs) like Cloudflare or Fastly universally offer options for adding a WAF to their setup. Both firms immediately <a href="https://blog.cloudflare.com/php-vulnerabilities-waf-protection/" target="_blank">issued</a> their <a href="https://www.fastly.com/blog/fastlys-proactive-protection-critical-php-rce-cve-2026-109618" target="_blank">respective notices</a> and added detection of PHP2Shell to their respective rulesets. Finally, server-level security suites <a href="https://www.crowdstrike.com/en-us/blog/crowdstrike-protects-customers-from-php-rce-zero-days/" target="_blank">like CrowdStrike</a> can also detect and block malicious activity by the PHP binary itself, in case an attacker tries something funny. It's reasonable to guess that these protections have slowed exploitation efforts.</p><h2 id="so-where-are-we">So where are we?</h2><p>Right now, the PHP2Shell exploit is doing the rounds, with at least <a href="https://dreamgroup.com/blog/chained-php2shell-exploit-roots-25-servers" target="_blank">one high-profile case</a> involving a campaign that chained the PHP vulnerability together with Fragnesia (aka <a href="https://www.tomshardware.com/tech-industry/cyber-security/dirty-frag-exploit-gets-root-on-most-linux-machines-since-2017-no-patches-available-no-warning-given-copy-fail-like-vulnerability-had-its-embargo-broken">Dirty Frag</a>) to eventually get <em>kernel-level access</em> to 25 servers seemingly linked to the U.S. government. As of October 29, security firm BitSight had recorded copious attempts of PHP2Shell intrusion across its monitoring fleet, from many distinct origins.</p><p>BitSight's dark web forum sleuths indicate that there was no one-size-fits-all exploit package available at that date. This may indicate that hacking campaigns for now are likely to be specifically targeted, but a proverbial one-click red button is bound to appear at some point, likely with a deeper exploit in tow, as with the Fragnesia example above.</p><p>Meanwhile, several cybersecurity firms kept a close eye on patching progress since the October 19 reveal. Censys has a <a href="https://php2shell.censys.com/">real-time view</a>, updated hourly, and suffice it to say that things could be going a lot better. The initial count estimated the number of vulnerable PHP installs at around 45.7 million, out of 119 million total.</p><p>The number of unpatched services has mercifully dropped to around 29 million as of this writing, seemingly indicating a patch rate of around 1 million per day, depending on how you count. It's worth noting that fix backports may not change the version number, so those instances could still count as unpatched. </p><p>But remember how most everyone toggled off their server's PHP version number reporting. According to Censys, 62% of PHP servers it's consulting aren't reporting that datum at all, so there's no telling what the actual totals might be. If history is any indication, a significant percentage of those machines remain unpatched today.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Google reveals new Gemini 4 Carbon AI model, narrowly beating Claude Opus 5.5 ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Last week, Google released its latest AI model, Gemini 4 Carbon, topping intelligence rankings charts. The model has landed with a grander splash than most. As <a href="https://x.com/ArtificialAnlys/status/2115494553780367384" target="_blank">Artificial Analysis highlights</a>, Gemini 4 Carbon is now the highest-ranked AI model on its latest Intelligence Index V5 benchmark suite, beating Claude Opus 5.5 by one point. It achieves the number one score across four of the nine evaluations, and does so at a lower cost per task than other high-ranking models, like Claude Opus 5.5. Google is offering access to the model for cyber defenders first through its Fairwind program, with wider availability coming sometime after.</p><p>Perhaps more crucially, Gemini 4 Carbon stands apart by topping Artificial Analysis' new AA-Omniscience v 1.1 test that tracks hallucinations and knowledgeable recall of information. As governments and institutions look to maximize value and security from their use of AI, being able to claim the lowest hallucination score could prove a significant selling point.</p><p>Then again, even the best scored only 54 on the -100 to +100 scale, suggesting even the best frontier models in 2026 are a far cry from being consistently reliable when it comes to accurately recalling information.</p><h2 id="the-king-is-dead-long-live-the-king">The king is dead, long live the king</h2><p>AI model rankings change fast and regularly in 2026, and it's no surprise that the latest release from a frontier lab has pushed the envelope, but Google's Gemini 4 Carbon can be seen as a comeback after languishing with irregular releases for much of the year. </p><p>Gemini 4 Carbon (high) scores a 59 on the Artificial Analysis Intelligence Index V5 benchmark. That puts it a single point ahead of Claude Opus 5.5 (max with fallback), and two points ahead of OpenAI's GPT-6 Astra (max). It's also nine points ahead of Meta's Muse Wonder Full (Max), and 10 points ahead of Claude Fable 5.1 (max with fallback).</p><p>But raw, generalized intelligence isn't the whole picture with frontier models in 2026. As developers <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/frontier-ai-faces-pricing-reckoning-as-token-volume-explodes-25-fold-mid-tier-models-deliver-90-percent-of-flagship-capability-at-one-sixth-the-cost" target="_blank">chase the Pareto frontier</a>— the best-possible confluence of cost and intelligence — Gemini 4 Carbon is also impressively affordable. Its cost per task on the Intelligence Index test is actually slightly cheaper than Gemini 3.8 Flash (high), priced at $7.76 each. That's several dollars cheaper than either Claude Opus 5.5 or Sonnet 5.5, both max with fallback, and not far off half the price of Claude Fable 5.1, which costs $12.99 per task.</p><p>Gemini 4 Carbon is still more expensive than OpenAI's GPT 5 Astra and about six times more expensive than GPT 6.1 Sol (max). It's about 13 times as expensive per task as the open-weight <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/deepseeks-latest-ai-model-is-smarter-than-ever-but-its-low-running-costs-come-at-a-price-open-weight-model-is-fit-for-agentic-workloads-but-is-more-prone-to-hallucination-than-proprietary-competitors" target="_blank">DeepSeek V4.1 Flash 1017, too, although that model measures a mere 32 on the Intelligence Index rankings. </a></p><p>Gemini 4 Carbon isn't the best at everything, but it may be the best at most of the high price points. At least for now. <a href="https://arena.ai/leaderboard/chat/text/pareto" target="_blank">Arena.ai has it holding the top spot</a> on its Pareto Frontier chart at the more expensive end of the spectrum, but at lower price points, the Chinese open-weight models like GLM 5.3-Flash and DeepSeek V4 Pro take over.</p><p>Artificial Analysis also highlights that Gemini 4 Carbon isn't as good as Claude Code with Claude Opus 5.5 at programming, and it falls behind Grok 4.7, Muse Wonder, and Kimi K3.1 Pro on knowledge work. </p><p>But perhaps the most impressive score is Gemini 4's 54 on the AA-Omniscience Index, which tracks knowledge and hallucinations, beating out the competition by often significant margins.</p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2115494556548599953"><p lang="en" dir="ltr">Gemini 4 Carbon costs $7.76 per Intelligence Index task, ~23% less than Claude Opus 5.5 (max, $10.04) but ~1.2x GPT-6 Astra (max, $6.25). It uses 123k output tokens per task, 29% fewer than Claude Opus 5.5 (max, 173k). At the same $4/$20 standard pricing, its Cost per Task is in line with Gemini 4 Argon ($7.68), while scoring 9 points higher on the Intelligence Index.<a href="https://twitter.com/cantworkitout/status/2115494556548599953">October 28, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><h2 id="careful-and-compliant">Careful and Compliant</h2><p>In <a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/" target="_blank">Google's announcement of the new Gemini 4 model</a>, it talked up its collaboration with the U.S. government for model screening and pre-release testing. In a similar manner to the way Anthropic handled Mythos' release in 2025, Google is rolling out access to Gemini 4 Carbon at a slower pace. </p><p>"We're taking a deliberate approach to rollout and are actively engaged in the U.S. government’s voluntary process for pre-release model testing," <a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/" target="_blank">Google's launch statement reads</a>.</p><p>To that end, it's opening up Carbon to a subset of "cyber defenders," and will gradually make the model more widely available through trusted partners through <a href="https://deepmind.google/fairwind-program/" target="_blank">its Fairwind Program</a>. Before it makes Carbon available to developers, it wants to iterate on its guardrails to make the AI safer and more effective for cybersecurity purposes, making it harder for nefarious actors to use it for exploits and hacking.</p><p>Where the idea of slowing the industry and introducing greater oversight for model development has risen and fallen in the estimations of the major tech companies, it does appear to be cementing itself as more of a foundational idea for future model releases. Google talks up its release on the improved safety and security of its model, highlighting that it thwarted all attempts from the Gray Swan Indirect Prompt Injection benchmark to have it do something it shouldn't. </p><p>Google claims Carbon cuts unsupported claims in its responses twice as much as the Gemini 4 Argon model it debuted just two weeks ago. If it can't complete an action, it will just not do so, rather than fabricating an answer or hedging around missing evidence, Google claims.</p><p>It also suggests Carbon will present sources and suggested actions to the user more readily, rather than pursuing its own intentions when there are discrepancies or uncertainties, allegedly cutting model overreach by 26%.</p><p>As governments and companies look to make realistic use of AI, tracking usage, managing token budgets, and preventing workers from creating new exploit paths through overuse, a model's safety, security, and reliability will be paramount. </p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2115494564081570069"><p lang="en" dir="ltr">Gemini 4 Carbon scores 54 on AA-Omniscience, the highest of any model, 6 points ahead of Claude Opus 5.5 (max, 48). It answers more questions correctly than Gemini 4 Argon (58% vs 48% accuracy), at the cost of a higher hallucination rate (11% vs 4%). This is still far below Claude Opus 5.5 (max) and GPT-6 Astra (max), both at 50%<a href="https://twitter.com/cantworkitout/status/2115494564081570069">October 28, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><h2 id="a-new-pareto-frontier">A new Pareto Frontier?</h2><p>For the past year, a clear focus from frontier and open-weight model developers has been on efficiency and cost. While raising intelligence and improving capabilities, particularly through agentic workflows, has seen great development alongside pomp and circumstance, many of the major announcements and developments have also <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/the-price-of-ai-is-crashing-faster-than-the-rate-of-moores-law-report-suggests-intelligence-costs-are-in-freefall-outpacing-comparative-technologies-like-compute-dna-sequencing-and-lithium-batteries" target="_blank">considered cost a clear component.</a></p><p>That's not likely to change, with the latest research from Gartner showing that <a href="https://www.gartner.com/en/newsroom/press-releases/2026-10-26-gartner-hr-research-finds-77-percent-of-board-members-expect-ai-to-deliver-productivity-and-cost-savings-yet-just-3-percent-of-organizations-exceed-cost-savings-goals-via-productivity" target="_blank">just 3% of companies are actually achieving cost and productivity improvements</a> from AI use. But as organizations push to find the real value in AI, maximize token budgets, and try to empower workers who have found improvements from using AI, they'll want consistency.</p><p>Tracking return on investment with AI is notoriously difficult, with hallucinations making that far harder. If Google can sell its latest model(s) as the safest and least likely to hallucinate, and do so with high intelligence at a relatively low cost, that might be the new winning formula. </p><p>At least until the next model drops.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/artificial-intelligence/google-reveals-new-gemini-4-carbon-ai-model-narrowly-beating-claude-opus-5-5-frontier-ai-model-is-also-the-least-likely-to-hallucinate-benchmarks-claim</link>
                                                                            <description>
                            <![CDATA[ Last week, Google released its latest AI model, Gemini 4 Carbon, topping intelligence rankings charts. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">GM823wq7NFMUw4fHUZnFFe</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/bCKPeU3stBtYmdyp38NMsd-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Tue, 03 Nov 2026 13:00:00 +0000</pubDate>                                                                                                                                <updated>Tue, 03 Nov 2026 14:55:57 +0000</updated>
                                                                                                                                            <category><![CDATA[Artificial Intelligence]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Jon Martindale ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/YeutDv8zJmhi7xH35MSt8Z-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;After building his first computers in his teens, Jon Martindale has spent the past two decades covering the latest advances in technology. From displays to PC components, blockchain to AI, and tablets to standing desk accessories, Jon has covered just about every facet of the tech space in his varied career. He has bylines at Forbes, USNews, Lifewire, DigitalTrends, PCWorld, and a range of other sites. He brings that same level of expertise and professional insight to Toms Hardware.Away from writing, Jon is an avid reader, board gamer, and fitness enthusiast. He lives in rural Gloucestershire with his wife, two children, and French Bulldog cross.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/bCKPeU3stBtYmdyp38NMsd-1920-80.jpg">
                                                            <media:credit><![CDATA[VCG/VCG via Getty Images]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Gemini 4 Carbon name on phone.]]></media:description>                                                            <media:text><![CDATA[Gemini 4 Carbon name on phone.]]></media:text>
                                <media:title type="plain"><![CDATA[Gemini 4 Carbon name on phone.]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/bCKPeU3stBtYmdyp38NMsd-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Last week, Google released its latest AI model, Gemini 4 Carbon, topping intelligence rankings charts. The model has landed with a grander splash than most. As <a href="https://x.com/ArtificialAnlys/status/2115494553780367384" target="_blank">Artificial Analysis highlights</a>, Gemini 4 Carbon is now the highest-ranked AI model on its latest Intelligence Index V5 benchmark suite, beating Claude Opus 5.5 by one point. It achieves the number one score across four of the nine evaluations, and does so at a lower cost per task than other high-ranking models, like Claude Opus 5.5. Google is offering access to the model for cyber defenders first through its Fairwind program, with wider availability coming sometime after.</p><p>Perhaps more crucially, Gemini 4 Carbon stands apart by topping Artificial Analysis' new AA-Omniscience v 1.1 test that tracks hallucinations and knowledgeable recall of information. As governments and institutions look to maximize value and security from their use of AI, being able to claim the lowest hallucination score could prove a significant selling point.</p><p>Then again, even the best scored only 54 on the -100 to +100 scale, suggesting even the best frontier models in 2026 are a far cry from being consistently reliable when it comes to accurately recalling information.</p><h2 id="the-king-is-dead-long-live-the-king">The king is dead, long live the king</h2><p>AI model rankings change fast and regularly in 2026, and it's no surprise that the latest release from a frontier lab has pushed the envelope, but Google's Gemini 4 Carbon can be seen as a comeback after languishing with irregular releases for much of the year. </p><p>Gemini 4 Carbon (high) scores a 59 on the Artificial Analysis Intelligence Index V5 benchmark. That puts it a single point ahead of Claude Opus 5.5 (max with fallback), and two points ahead of OpenAI's GPT-6 Astra (max). It's also nine points ahead of Meta's Muse Wonder Full (Max), and 10 points ahead of Claude Fable 5.1 (max with fallback).</p><p>But raw, generalized intelligence isn't the whole picture with frontier models in 2026. As developers <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/frontier-ai-faces-pricing-reckoning-as-token-volume-explodes-25-fold-mid-tier-models-deliver-90-percent-of-flagship-capability-at-one-sixth-the-cost" target="_blank">chase the Pareto frontier</a>— the best-possible confluence of cost and intelligence — Gemini 4 Carbon is also impressively affordable. Its cost per task on the Intelligence Index test is actually slightly cheaper than Gemini 3.8 Flash (high), priced at $7.76 each. That's several dollars cheaper than either Claude Opus 5.5 or Sonnet 5.5, both max with fallback, and not far off half the price of Claude Fable 5.1, which costs $12.99 per task.</p><p>Gemini 4 Carbon is still more expensive than OpenAI's GPT 5 Astra and about six times more expensive than GPT 6.1 Sol (max). It's about 13 times as expensive per task as the open-weight <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/deepseeks-latest-ai-model-is-smarter-than-ever-but-its-low-running-costs-come-at-a-price-open-weight-model-is-fit-for-agentic-workloads-but-is-more-prone-to-hallucination-than-proprietary-competitors" target="_blank">DeepSeek V4.1 Flash 1017, too, although that model measures a mere 32 on the Intelligence Index rankings. </a></p><p>Gemini 4 Carbon isn't the best at everything, but it may be the best at most of the high price points. At least for now. <a href="https://arena.ai/leaderboard/chat/text/pareto" target="_blank">Arena.ai has it holding the top spot</a> on its Pareto Frontier chart at the more expensive end of the spectrum, but at lower price points, the Chinese open-weight models like GLM 5.3-Flash and DeepSeek V4 Pro take over.</p><p>Artificial Analysis also highlights that Gemini 4 Carbon isn't as good as Claude Code with Claude Opus 5.5 at programming, and it falls behind Grok 4.7, Muse Wonder, and Kimi K3.1 Pro on knowledge work. </p><p>But perhaps the most impressive score is Gemini 4's 54 on the AA-Omniscience Index, which tracks knowledge and hallucinations, beating out the competition by often significant margins.</p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2115494556548599953"><p lang="en" dir="ltr">Gemini 4 Carbon costs $7.76 per Intelligence Index task, ~23% less than Claude Opus 5.5 (max, $10.04) but ~1.2x GPT-6 Astra (max, $6.25). It uses 123k output tokens per task, 29% fewer than Claude Opus 5.5 (max, 173k). At the same $4/$20 standard pricing, its Cost per Task is in line with Gemini 4 Argon ($7.68), while scoring 9 points higher on the Intelligence Index.<a href="https://twitter.com/cantworkitout/status/2115494556548599953">October 28, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><h2 id="careful-and-compliant">Careful and Compliant</h2><p>In <a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/" target="_blank">Google's announcement of the new Gemini 4 model</a>, it talked up its collaboration with the U.S. government for model screening and pre-release testing. In a similar manner to the way Anthropic handled Mythos' release in 2025, Google is rolling out access to Gemini 4 Carbon at a slower pace. </p><p>"We're taking a deliberate approach to rollout and are actively engaged in the U.S. government’s voluntary process for pre-release model testing," <a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/" target="_blank">Google's launch statement reads</a>.</p><p>To that end, it's opening up Carbon to a subset of "cyber defenders," and will gradually make the model more widely available through trusted partners through <a href="https://deepmind.google/fairwind-program/" target="_blank">its Fairwind Program</a>. Before it makes Carbon available to developers, it wants to iterate on its guardrails to make the AI safer and more effective for cybersecurity purposes, making it harder for nefarious actors to use it for exploits and hacking.</p><p>Where the idea of slowing the industry and introducing greater oversight for model development has risen and fallen in the estimations of the major tech companies, it does appear to be cementing itself as more of a foundational idea for future model releases. Google talks up its release on the improved safety and security of its model, highlighting that it thwarted all attempts from the Gray Swan Indirect Prompt Injection benchmark to have it do something it shouldn't. </p><p>Google claims Carbon cuts unsupported claims in its responses twice as much as the Gemini 4 Argon model it debuted just two weeks ago. If it can't complete an action, it will just not do so, rather than fabricating an answer or hedging around missing evidence, Google claims.</p><p>It also suggests Carbon will present sources and suggested actions to the user more readily, rather than pursuing its own intentions when there are discrepancies or uncertainties, allegedly cutting model overreach by 26%.</p><p>As governments and companies look to make realistic use of AI, tracking usage, managing token budgets, and preventing workers from creating new exploit paths through overuse, a model's safety, security, and reliability will be paramount. </p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2115494564081570069"><p lang="en" dir="ltr">Gemini 4 Carbon scores 54 on AA-Omniscience, the highest of any model, 6 points ahead of Claude Opus 5.5 (max, 48). It answers more questions correctly than Gemini 4 Argon (58% vs 48% accuracy), at the cost of a higher hallucination rate (11% vs 4%). This is still far below Claude Opus 5.5 (max) and GPT-6 Astra (max), both at 50%<a href="https://twitter.com/cantworkitout/status/2115494564081570069">October 28, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><h2 id="a-new-pareto-frontier">A new Pareto Frontier?</h2><p>For the past year, a clear focus from frontier and open-weight model developers has been on efficiency and cost. While raising intelligence and improving capabilities, particularly through agentic workflows, has seen great development alongside pomp and circumstance, many of the major announcements and developments have also <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/the-price-of-ai-is-crashing-faster-than-the-rate-of-moores-law-report-suggests-intelligence-costs-are-in-freefall-outpacing-comparative-technologies-like-compute-dna-sequencing-and-lithium-batteries" target="_blank">considered cost a clear component.</a></p><p>That's not likely to change, with the latest research from Gartner showing that <a href="https://www.gartner.com/en/newsroom/press-releases/2026-10-26-gartner-hr-research-finds-77-percent-of-board-members-expect-ai-to-deliver-productivity-and-cost-savings-yet-just-3-percent-of-organizations-exceed-cost-savings-goals-via-productivity" target="_blank">just 3% of companies are actually achieving cost and productivity improvements</a> from AI use. But as organizations push to find the real value in AI, maximize token budgets, and try to empower workers who have found improvements from using AI, they'll want consistency.</p><p>Tracking return on investment with AI is notoriously difficult, with hallucinations making that far harder. If Google can sell its latest model(s) as the safest and least likely to hallucinate, and do so with high intelligence at a relatively low cost, that might be the new winning formula. </p><p>At least until the next model drops.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ PsiQuantum turns its photonics know-how into an optical switch for supercharging AI data center GPU networks ]]></title>
                                                                                                <dc:content><![CDATA[ <p>PsiQuantum, a quantum computing company, announced in <a href="https://www.psiquantum.com/news-import/psiquantum-unveils-optical-circuit-switch-technology-built-on-advanced-low-loss-silicon-photonics-platform">a release</a> earlier this week that it has turned its <a href="https://www.tomshardware.com/tech-industry/quantum-computing/quantum-photonics-roadmap-how-xanadu-and-psiquantum-are-looking-to-transfer-qubits-through-beams-of-light">low-loss silicon photonics</a> into an optical switch for use in AI data centers. Every optical switch typically comes with loss of light: the more light a switch receives, the less reaches the other side, with the amount of light lost measured in decibels. Typical data center optics can tolerate half the light (3 dB) being lost. PsiQuantum's switch debuted with a demo that typically loses only 1.2 dB of signal, or about a quarter of the light in lab tests, and the all-optical switch can achieve it without an amplifier. The prototype is, however, comparatively small, with eight inputs and outputs (8x8) against competitors’ 32 to 300, but the company says larger devices are in development.</p><p>The optical circuit switch (OCS) passes light straight through, from fiber to fiber, without converting it to or from electrical signals. Regular data center switches make that conversion at every hop, which takes power and adds latency. An OCS, instead, uses a fixed light path that it maintains until a change is ordered. Its main penalty comes down to how much light is lost when passing through the switch. PsiQuantum’s switch, which uses circuits etched on a silicon chip with no moving parts, unlike Google’s and Lumentum’s tiny mirrors, or microelectromechanical systems (MEMS), would be used alongside regular packet switches.</p><p>Skipping those conversions matters more as AI clusters grow, because moving data between GPUs is quickly becoming a bottleneck. Optics are a high-bandwidth improvement over copper, which has limitations in reach, although <a href="https://www.tomshardware.com/tech-industry/semiconductors/lumentum-ceo-says-the-indium-phosphide-shortage-will-become-worse-than-memory">a shortage of lasers</a> is a problem for the alternative. Improving communication bandwidth and distance is the goal, so that different racks of GPUs can talk at speed and act as one.</p><p>Switches similar to PsiQuantum’s <a href="https://www.tomshardware.com/tech-industry/photonics-and-high-speed-data-movement-is-the-next-big-ai-bottleneck-following-copper-power-dram-and-nand">made by Salience Labs</a> and iPronics use built-in amplification and, at least in the former case, are also rated at under 2 dB, or under 37% of the light lost. <a href="https://www.tomshardware.com/tech-industry/semiconductors/google-splits-its-tpu-into-two-chips-for-the-first-time-with-training-and-inference-variants">Google’s mirror-based switch</a>, Palomar, can also typically stay under 2 dB and has 136 ports. Salience’s switch sits at 32 and Lumentum’s at 300, which leaves the eight-port prototype a ways behind on how many fibers it can handle.</p><div ><table><caption>Optical circuit switches by loss and size</caption><thead><tr><th class="firstcol " ><p>Switch</p></th><th  ><p>Ports (in × out)</p></th><th  ><p>Insertion loss, and what the figure is</p></th><th  ><p>Reconfiguration</p></th><th  ><p>How it switches</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>PsiQuantum 8×8 demo</p></td><td  ><p>8×8</p></td><td  ><p>1.2 dB (about 24% of the light lost), typical in the company’s lab tests</p></td><td  ><p>Under 3 ms</p></td><td  ><p>Silicon photonics chip, no amplification, no moving parts</p></td></tr><tr><td class="firstcol " ><p>Google Palomar</p></td><td  ><p>136×136</p></td><td  ><p>Typically under 2 dB (under 37% lost) across all port pairings; 2 dB worst case</p></td><td  ><p>Milliseconds</p></td><td  ><p>MEMS mirrors (moving parts)</p></td></tr><tr><td class="firstcol " ><p>Lumentum R300</p></td><td  ><p>300×300</p></td><td  ><p>Close to 1.5 dB (about 29% lost), typical at start of life, Lumentum’s switching GM told <em>Gazettabyte</em></p></td><td  ><p>Not stated</p></td><td  ><p>MEMS mirrors (moving parts)</p></td></tr><tr><td class="firstcol " ><p>Salience Labs OC-32M</p></td><td  ><p>32×32</p></td><td  ><p>Under 2 dB (under 37% lost) port to port, a ceiling, after built-in amplification</p></td><td  ><p>Under 300 µs</p></td><td  ><p>Silicon photonics chip with amplification</p></td></tr><tr><td class="firstcol " ><p>iPronics ONE-32</p></td><td  ><p>32 (layout not stated)</p></td><td  ><p>Not published</p></td><td  ><p>Under 300 µs</p></td><td  ><p>Silicon photonics chip with optical amplifiers</p></td></tr></tbody></table></div><p>PsiQuantum's switch is an 8x8, meaning eight in and eight out, and “every input can reach every output.” It’s also “strictly non-blocking,” so a new connection never disturbs ones already in place. In the lab, it measured a typical 1.2 dB insertion loss, 40 dB of crosstalk (about 0.01% of the light leaking to other ports), and under 3 ms of reconfiguration time, all of which exceed “initial targets set for these devices.” The loss figure is measured fiber to fiber, which includes getting light onto and off the chip but not connector losses, PsiQuantum told <em>Tom’s Hardware Premium</em>.</p><p>The platform loses very little light at each component, which matters because a quantum computer can tolerate only so much lost light. PsiQuantum’s 2025 <a href="https://www.nature.com/articles/s41586-025-08820-7"><em>Nature</em> paper</a> reports silicon nitride waveguides losing as little as 0.5 dB per meter (about 11% of the light per meter), splitters 0.5 millidecibels, and crossings 1.2 millidecibels, the last two a few hundredths of a percent of the light or less each.</p><p>The company also says chip-to-fiber coupling has reached about 60 millidecibels, roughly 1.4% of the light lost, in its best published lab results, although without a full fiber array attached. Those figures set a best case, not a worst case, for a 64-port switch, since light crosses more components as ports are added and the losses add up.</p><p>How much loss is acceptable varies with the optics at the end of each link. <a href="https://www.tomshardware.com/tech-industry/nvidia-invests-usd4-billion-into-photonics-firms-in-a-bid-to-bolster-data-center-interconnect-supply-chains-lumentum-and-coherent-investment-to-fund-u-s-r-and-d-and-manufacturing-facilities-supports-capacity-rights-and-future-access">Coherent</a>, an optics maker, says that using an OCS can add “up to 3 dB additional insertion loss,” but its transceivers are able to work with that. Lumentum told <a href="https://gazettabyte.com/lumentums-optical-circuit-switch-for-ai-data-centres/" target="_blank"><em>Gazettabyte</em> </a>that 3 dB is good enough to work with off-the-shelf FR4, DR4, and DR8 optical modules using the switch; those are the standard plug-in options used to link GPU servers and switches, at distances of 500 meters to 2 kilometers. PsiQuantum’s 8x8 is built for the O-band, the wavelength range those modules use. PsiQuantum told <a href="https://www.semafor.com/article/10/28/2026/quantum-computing-companys-first-breakthrough-is-in-ai-data-centers" target="_blank"><em>Semafor</em></a> that, to be workable, its switch needed insertion loss under 2 to 3 dB.</p><p>The 8x8 prototype was designed for conventional data center traffic and comes in well below the 3 dB loss target in a request from Nvidia for proposals, the company says. The port count marks it as a demonstration vehicle, but PsiQuantum has a 64x64 version taped out with a 72x72 underway and is “actively developing larger devices.” With crossings that lose so little light, the scale-up might hold. PsiQuantum declined to share details on future products, though it says the loss, which may be slightly higher, will stay very low.</p><p>Speed remains less important than insertion loss for how OCSes are currently used. Google, which has the most deployments, numbering in the “tens of thousands,” uses its switches to resize TPU pods or to route around failed chips, a job its “millisecond-scale” MEMS switches handle. PsiQuantum’s under 3 ms is within that class, but Salience and iPronics claim less than 300 microseconds for their amplified chip switches, which iPronics highlights for mid-workload network rewiring to hide the switch time while chips compute.</p><p>There is quite a lot of room for speed to improve, and PsiQuantum already has a material for that, used in an existing 64-port switch for its quantum systems, which need to switch light at gigahertz rates. The material, barium titanate (BTO), changes how it bends light when a voltage is applied, and the company chose it over lithium niobate, a more common selection, because BTO had the strongest effect it could find. This keeps switching elements short and losses low at speeds roughly a million times faster than the demo. BTO is “not required for the demonstration vehicle or initial products,” which use a thermally switched silicon nitride platform, but it is “an additional path to faster, more advanced future products.” </p><p>An argument for starting now is that building thousands of such switches for conventional applications would reduce the risk for the packaging, assembly, and supply chain of parts PsiQuantum builds for quantum computing. AI and quantum computing are “complementary technologies,” it says, and this is the chance to put its technology into action. When PsiQuantum hired Rob Soderbery as an executive vice president in July, it said he would advance its silicon photonics platform, “accelerating its application across AI networking and other next-generation computing infrastructure,” hinting at the roadmap. The company’s <a href="https://www.tomshardware.com/tech-industry/quantum-computing/ibm-spins-off-americas-first-quantum-chip-foundry-with-2-billion-in-federal-and-private-funding">CHIPS award</a> was finalized in September for “high-performance optical switches,” among other components.</p><p>Co-founder Pete Shadbolt praised Google’s alternative MEMS switch but would prefer to have similar-quality switching on a monolithic chip that needs no moving parts. Nvidia has a stake through its venture arm, NVentures, and collaborates with PsiQuantum on “GPU-QPU integration,” that is, graphics and quantum processing units, as well as silicon photonics. Nvidia also joined rival iPronics’ $125 million round in September to develop its second-generation optical switch.</p><p>Right now, inside Google, OCSes rewire AI clusters, and research firm Cignal AI expects most deployments to stay there through the end of the decade; outside of Google, OCS is still at “proof-of-concept and early trial.” The firm raised its market estimates earlier this year, though, roughly threefold for 2026 and by more than 40% for 2029. An opening exists for PsiQuantum in GPU networks, which ties into Nvidia’s interest, though adoption there is expected to be limited until then. The company says samples and testing kits have already gone to several computing companies, including GPU vendors and hyperscalers.</p><p>PsiQuantum should have commercial products “announced in the coming months,” along with updates on larger devices, while competitor Salience has already demonstrated 64 ports at this year’s European Conference on Optical Communication. Salience’s roadmap already has 128- and 256-port products. Scaling is always an issue, but if PsiQuantum’s figures, which haven’t been independently verified, hold up at 64 ports without amplifiers, then larger monolithic OCSes are on the way.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/photonics/psiquantum-turns-its-photonics-know-how-into-an-optical-switch-for-supercharging-ai-data-center-gpu-networks-8x8-lab-prototype-loses-less-light-than-rivals-with-plans-to-develop-larger-devices</link>
                                                                            <description>
                            <![CDATA[ PsiQuantum’s eight-port optical switch typically loses 1.2 dB, under a quarter of its light, where data centers allow up to 3 dB. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">o86s6NBaUfobiVYp4hW5qj</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/N7owNVHhCfsFWp2DAu3woc-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Fri, 30 Oct 2026 14:30:00 +0000</pubDate>                                                                                                                                <updated>Sat, 31 Oct 2026 03:12:45 +0000</updated>
                                                                                                                                            <category><![CDATA[Photonics]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/N7owNVHhCfsFWp2DAu3woc-1920-80.jpg">
                                                            <media:credit><![CDATA[PsiQuantum]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[A green PsiQuantum circuit board module with a photonic chip and a ribbon of optical fibers attached, on a blue background.]]></media:description>                                                            <media:text><![CDATA[A green PsiQuantum circuit board module with a photonic chip and a ribbon of optical fibers attached, on a blue background.]]></media:text>
                                <media:title type="plain"><![CDATA[A green PsiQuantum circuit board module with a photonic chip and a ribbon of optical fibers attached, on a blue background.]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/N7owNVHhCfsFWp2DAu3woc-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>PsiQuantum, a quantum computing company, announced in <a href="https://www.psiquantum.com/news-import/psiquantum-unveils-optical-circuit-switch-technology-built-on-advanced-low-loss-silicon-photonics-platform">a release</a> earlier this week that it has turned its <a href="https://www.tomshardware.com/tech-industry/quantum-computing/quantum-photonics-roadmap-how-xanadu-and-psiquantum-are-looking-to-transfer-qubits-through-beams-of-light">low-loss silicon photonics</a> into an optical switch for use in AI data centers. Every optical switch typically comes with loss of light: the more light a switch receives, the less reaches the other side, with the amount of light lost measured in decibels. Typical data center optics can tolerate half the light (3 dB) being lost. PsiQuantum's switch debuted with a demo that typically loses only 1.2 dB of signal, or about a quarter of the light in lab tests, and the all-optical switch can achieve it without an amplifier. The prototype is, however, comparatively small, with eight inputs and outputs (8x8) against competitors’ 32 to 300, but the company says larger devices are in development.</p><p>The optical circuit switch (OCS) passes light straight through, from fiber to fiber, without converting it to or from electrical signals. Regular data center switches make that conversion at every hop, which takes power and adds latency. An OCS, instead, uses a fixed light path that it maintains until a change is ordered. Its main penalty comes down to how much light is lost when passing through the switch. PsiQuantum’s switch, which uses circuits etched on a silicon chip with no moving parts, unlike Google’s and Lumentum’s tiny mirrors, or microelectromechanical systems (MEMS), would be used alongside regular packet switches.</p><p>Skipping those conversions matters more as AI clusters grow, because moving data between GPUs is quickly becoming a bottleneck. Optics are a high-bandwidth improvement over copper, which has limitations in reach, although <a href="https://www.tomshardware.com/tech-industry/semiconductors/lumentum-ceo-says-the-indium-phosphide-shortage-will-become-worse-than-memory">a shortage of lasers</a> is a problem for the alternative. Improving communication bandwidth and distance is the goal, so that different racks of GPUs can talk at speed and act as one.</p><p>Switches similar to PsiQuantum’s <a href="https://www.tomshardware.com/tech-industry/photonics-and-high-speed-data-movement-is-the-next-big-ai-bottleneck-following-copper-power-dram-and-nand">made by Salience Labs</a> and iPronics use built-in amplification and, at least in the former case, are also rated at under 2 dB, or under 37% of the light lost. <a href="https://www.tomshardware.com/tech-industry/semiconductors/google-splits-its-tpu-into-two-chips-for-the-first-time-with-training-and-inference-variants">Google’s mirror-based switch</a>, Palomar, can also typically stay under 2 dB and has 136 ports. Salience’s switch sits at 32 and Lumentum’s at 300, which leaves the eight-port prototype a ways behind on how many fibers it can handle.</p><div ><table><caption>Optical circuit switches by loss and size</caption><thead><tr><th class="firstcol " ><p>Switch</p></th><th  ><p>Ports (in × out)</p></th><th  ><p>Insertion loss, and what the figure is</p></th><th  ><p>Reconfiguration</p></th><th  ><p>How it switches</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>PsiQuantum 8×8 demo</p></td><td  ><p>8×8</p></td><td  ><p>1.2 dB (about 24% of the light lost), typical in the company’s lab tests</p></td><td  ><p>Under 3 ms</p></td><td  ><p>Silicon photonics chip, no amplification, no moving parts</p></td></tr><tr><td class="firstcol " ><p>Google Palomar</p></td><td  ><p>136×136</p></td><td  ><p>Typically under 2 dB (under 37% lost) across all port pairings; 2 dB worst case</p></td><td  ><p>Milliseconds</p></td><td  ><p>MEMS mirrors (moving parts)</p></td></tr><tr><td class="firstcol " ><p>Lumentum R300</p></td><td  ><p>300×300</p></td><td  ><p>Close to 1.5 dB (about 29% lost), typical at start of life, Lumentum’s switching GM told <em>Gazettabyte</em></p></td><td  ><p>Not stated</p></td><td  ><p>MEMS mirrors (moving parts)</p></td></tr><tr><td class="firstcol " ><p>Salience Labs OC-32M</p></td><td  ><p>32×32</p></td><td  ><p>Under 2 dB (under 37% lost) port to port, a ceiling, after built-in amplification</p></td><td  ><p>Under 300 µs</p></td><td  ><p>Silicon photonics chip with amplification</p></td></tr><tr><td class="firstcol " ><p>iPronics ONE-32</p></td><td  ><p>32 (layout not stated)</p></td><td  ><p>Not published</p></td><td  ><p>Under 300 µs</p></td><td  ><p>Silicon photonics chip with optical amplifiers</p></td></tr></tbody></table></div><p>PsiQuantum's switch is an 8x8, meaning eight in and eight out, and “every input can reach every output.” It’s also “strictly non-blocking,” so a new connection never disturbs ones already in place. In the lab, it measured a typical 1.2 dB insertion loss, 40 dB of crosstalk (about 0.01% of the light leaking to other ports), and under 3 ms of reconfiguration time, all of which exceed “initial targets set for these devices.” The loss figure is measured fiber to fiber, which includes getting light onto and off the chip but not connector losses, PsiQuantum told <em>Tom’s Hardware Premium</em>.</p><p>The platform loses very little light at each component, which matters because a quantum computer can tolerate only so much lost light. PsiQuantum’s 2025 <a href="https://www.nature.com/articles/s41586-025-08820-7"><em>Nature</em> paper</a> reports silicon nitride waveguides losing as little as 0.5 dB per meter (about 11% of the light per meter), splitters 0.5 millidecibels, and crossings 1.2 millidecibels, the last two a few hundredths of a percent of the light or less each.</p><p>The company also says chip-to-fiber coupling has reached about 60 millidecibels, roughly 1.4% of the light lost, in its best published lab results, although without a full fiber array attached. Those figures set a best case, not a worst case, for a 64-port switch, since light crosses more components as ports are added and the losses add up.</p><p>How much loss is acceptable varies with the optics at the end of each link. <a href="https://www.tomshardware.com/tech-industry/nvidia-invests-usd4-billion-into-photonics-firms-in-a-bid-to-bolster-data-center-interconnect-supply-chains-lumentum-and-coherent-investment-to-fund-u-s-r-and-d-and-manufacturing-facilities-supports-capacity-rights-and-future-access">Coherent</a>, an optics maker, says that using an OCS can add “up to 3 dB additional insertion loss,” but its transceivers are able to work with that. Lumentum told <a href="https://gazettabyte.com/lumentums-optical-circuit-switch-for-ai-data-centres/" target="_blank"><em>Gazettabyte</em> </a>that 3 dB is good enough to work with off-the-shelf FR4, DR4, and DR8 optical modules using the switch; those are the standard plug-in options used to link GPU servers and switches, at distances of 500 meters to 2 kilometers. PsiQuantum’s 8x8 is built for the O-band, the wavelength range those modules use. PsiQuantum told <a href="https://www.semafor.com/article/10/28/2026/quantum-computing-companys-first-breakthrough-is-in-ai-data-centers" target="_blank"><em>Semafor</em></a> that, to be workable, its switch needed insertion loss under 2 to 3 dB.</p><p>The 8x8 prototype was designed for conventional data center traffic and comes in well below the 3 dB loss target in a request from Nvidia for proposals, the company says. The port count marks it as a demonstration vehicle, but PsiQuantum has a 64x64 version taped out with a 72x72 underway and is “actively developing larger devices.” With crossings that lose so little light, the scale-up might hold. PsiQuantum declined to share details on future products, though it says the loss, which may be slightly higher, will stay very low.</p><p>Speed remains less important than insertion loss for how OCSes are currently used. Google, which has the most deployments, numbering in the “tens of thousands,” uses its switches to resize TPU pods or to route around failed chips, a job its “millisecond-scale” MEMS switches handle. PsiQuantum’s under 3 ms is within that class, but Salience and iPronics claim less than 300 microseconds for their amplified chip switches, which iPronics highlights for mid-workload network rewiring to hide the switch time while chips compute.</p><p>There is quite a lot of room for speed to improve, and PsiQuantum already has a material for that, used in an existing 64-port switch for its quantum systems, which need to switch light at gigahertz rates. The material, barium titanate (BTO), changes how it bends light when a voltage is applied, and the company chose it over lithium niobate, a more common selection, because BTO had the strongest effect it could find. This keeps switching elements short and losses low at speeds roughly a million times faster than the demo. BTO is “not required for the demonstration vehicle or initial products,” which use a thermally switched silicon nitride platform, but it is “an additional path to faster, more advanced future products.” </p><p>An argument for starting now is that building thousands of such switches for conventional applications would reduce the risk for the packaging, assembly, and supply chain of parts PsiQuantum builds for quantum computing. AI and quantum computing are “complementary technologies,” it says, and this is the chance to put its technology into action. When PsiQuantum hired Rob Soderbery as an executive vice president in July, it said he would advance its silicon photonics platform, “accelerating its application across AI networking and other next-generation computing infrastructure,” hinting at the roadmap. The company’s <a href="https://www.tomshardware.com/tech-industry/quantum-computing/ibm-spins-off-americas-first-quantum-chip-foundry-with-2-billion-in-federal-and-private-funding">CHIPS award</a> was finalized in September for “high-performance optical switches,” among other components.</p><p>Co-founder Pete Shadbolt praised Google’s alternative MEMS switch but would prefer to have similar-quality switching on a monolithic chip that needs no moving parts. Nvidia has a stake through its venture arm, NVentures, and collaborates with PsiQuantum on “GPU-QPU integration,” that is, graphics and quantum processing units, as well as silicon photonics. Nvidia also joined rival iPronics’ $125 million round in September to develop its second-generation optical switch.</p><p>Right now, inside Google, OCSes rewire AI clusters, and research firm Cignal AI expects most deployments to stay there through the end of the decade; outside of Google, OCS is still at “proof-of-concept and early trial.” The firm raised its market estimates earlier this year, though, roughly threefold for 2026 and by more than 40% for 2029. An opening exists for PsiQuantum in GPU networks, which ties into Nvidia’s interest, though adoption there is expected to be limited until then. The company says samples and testing kits have already gone to several computing companies, including GPU vendors and hyperscalers.</p><p>PsiQuantum should have commercial products “announced in the coming months,” along with updates on larger devices, while competitor Salience has already demonstrated 64 ports at this year’s European Conference on Optical Communication. Salience’s roadmap already has 128- and 256-port products. Scaling is always an issue, but if PsiQuantum’s figures, which haven’t been independently verified, hold up at 64 ports without amplifiers, then larger monolithic OCSes are on the way.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Anthropic’s IPO filing shows most of its compute goes to R&D, not serving customers ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Anthropic spent less than a third of its $8.9 billion first-half compute bill on serving customers, with most of the rest going into training, research, and running Claude internally, <a href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm">its S-1 filing</a> shows. It has meanwhile signed up for hundreds of billions of dollars more in TPU and GPU capacity. By our math, it has meanwhile signed up for roughly $800 billion more in TPU and GPU capacity, and revenue per compute dollar jumped from $0.40 in Q1 2025 to $2.08 in Q2 2026, which the company links to capacity purchased 12–18 months ahead.</p><p><a href="https://www.tomshardware.com/tech-industry/anthropic-files-to-go-public-claude-maker-races-openai-and-spacex-to-ipo">Anthropic’s IPO</a> is likely to beat that of rival OpenAI, which <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-hit-with-sweeping-probe-from-massive-coalition-of-42-us-state-attorneys-general-just-days-after-reported-ipo-filing-subpoena-targets-chatgpt-makers-ads-data-practices-handling-of-minors-model-sycophancy-and-safety-policies">filed confidentially</a> before CEO Sam Altman said its own won’t come this year, amid widespread doubts about an AI bubble. Whether AI labs can cover what it costs to serve their models is an open question.</p><p>Anthropic appears to do so easily: it spent $2.4 billion serving customers in the first half of the year, versus $16.3 billion in revenue. Paying for what it has on order is more difficult. Its five compute deals with the biggest stated amounts would cost an estimated $103 billion a year on average once delivered, more than twice its revenue at Q2’s pace of about $46 billion a year. The company also warns that margins may compress as new hardware lands ahead of the revenue it will support.</p><p>The company has three main costs: the cost of revenue, which is the inference cost of running Claude for paying customers; R&amp;D, which is mostly compute for training and research into new models plus any internal use; and sales and marketing, which covers the cut taken by channel partners and cloud platform resellers.</p><p>Comparing the first half of last year to this year’s, cost of revenue is up about 2.3x, but with revenue up 14x, that means Anthropic keeps an estimated 85 cents for every dollar after serving costs, up from just 12 cents. R&amp;D is up about 5 times, with 66% or more of the compute going to it, and that’s the main source of its operating loss. Sales and marketing is up 16x, and doesn’t show up in cost of revenue, but its partner fees and marketing would lower what it keeps to 66 cents a dollar. The trend here is that serving Claude now more than covers its costs.</p><div ><table><caption>Where Anthropic’s Compute Goes</caption><thead><tr><th class="firstcol empty" ></th><th  ><p>First half of 2025</p></th><th  ><p>First half of 2026</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Revenue</p></td><td  ><p>$1.16 billion</p></td><td  ><p>$16.29 billion</p></td></tr><tr><td class="firstcol " ><p>Compute and infrastructure, all workloads</p></td><td  ><p>$2.33 billion</p></td><td  ><p>$8.95 billion</p></td></tr><tr><td class="firstcol " ><p>Serving customers (cost of revenue)</p></td><td  ><p>$1.02 billion</p></td><td  ><p>$2.38 billion</p></td></tr><tr><td class="firstcol " ><p>Free access to Claude, at most (est.)</p></td><td  ><p>$119 million</p></td><td  ><p>$640 million</p></td></tr><tr><td class="firstcol " ><p>Training, research and internal use, at least (est.)</p></td><td  ><p>$1.20 billion</p></td><td  ><p>$5.93 billion</p></td></tr><tr><td class="firstcol " ><p>Serving share of compute, at most</p></td><td  ><p>44%</p></td><td  ><p>27%</p></td></tr><tr><td class="firstcol " ><p>Revenue left after serving costs</p></td><td  ><p>12%</p></td><td  ><p>85%</p></td></tr><tr><td class="firstcol " ><p>Revenue left after serving costs, partner fees and marketing</p></td><td  ><p>2%</p></td><td  ><p>66%</p></td></tr></tbody></table></div><p>Serving has kept getting cheaper for customers since June: in September, Anthropic’s Opus 5.5 model cut costs per task by about 40% versus Opus 5, says the company, and Fable 5.1 cut the cost of typical workloads by about 25% against Fable 5; each came from “a combination of a lower price and improved efficiency,” made “while preserving our margins.”</p><p>As for compute, which runs through all three costs, Anthropic runs it as “a single, shared resource across all workloads” and doesn’t disclose how it decides the split between training, research, and serving. Serving was limited, as the company has “consistently operated in a compute-constrained environment where customer demand exceeds available compute,” which suggests serving costs were muted by a lack of supply, not just efficiency.</p><p>Anthropic runs a “multi-chip compute architecture across AWS Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs,” without giving any percentage splits or GW measurements. For TPUs, it has five-year equipment <a href="https://www.tomshardware.com/tech-industry/broadcom-expands-anthropic-deal-to-3-5gw-of-google-tpu-capacity-from-2027">leases with Broadcom</a> for “TPU-based compute capacity,” with rent of “approximately $161.2 billion excluding interest”; some of that capacity came online this year, with more from 2027. Another $114.5 billion or more is <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropic-signs-deal-with-google-cloud-to-expand-tpu-chip-capacity-ai-company-expects-to-have-over-1gw-of-processing-power-in-2026">with Google</a> “between April 2026 and July 2033,” with more from Google and other TPU providers from 2028 to 2031, estimated to be in the $125–$150 billion range.</p><p>Trainium comes through Amazon, with a deal for “at least $110.0 billion between May 2026 and April 2036,” though the filing doesn’t split AWS spend by chip. Nvidia GPUs come through a series of deals, including “up to $120.5 billion … extending into 2029” <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/musks-spacex-has-rented-out-access-to-its-supercomputers-220-000-nvidia-gpus-and-300-megawatts-of-ai-compute-power-to-rival-anthropic-musk-says-no-one-set-off-my-evil-detector-antrhropic-also-interested-in-orbital-data-centers">with SpaceXAI</a>, “at least $44.9 billion” extending into 2033 with Microsoft, “up to approximately $44.6 billion” also into 2033 with Nscale, and about $41.3 billion into 2034 with Lambda. There is also an <a href="https://www.tomshardware.com/tech-industry/amd-to-supply-anthropic-with-2-gigawatts-of-instinct-mi450-gpus">AMD hardware supply agreement</a> which, in aggregate, may “exceed $20.0 billion” if Anthropic takes the full initial capacity. These deals each have their own conditions for termination.</p><div ><table><caption>Anthropic’s Compute Contracts</caption><thead><tr><th class="firstcol " ><p>Partner</p></th><th  ><p>Hardware</p></th><th  ><p>Amount</p></th><th  ><p>Term</p></th><th  ><p>Can Anthropic walk away?</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Broadcom</p></td><td  ><p>Google TPUs (equipment leases)</p></td><td  ><p>About $161.2 billion in rent</p></td><td  ><p>Five-year leases; some capacity online in 2026, more from 2027</p></td><td  ><p>No, except on default</p></td></tr><tr><td class="firstcol " ><p>Google</p></td><td  ><p>Google TPUs</p></td><td  ><p>At least $114.5 billion</p></td><td  ><p>April 2026–July 2033</p></td><td  ><p>No, except on Google’s uncured material breach; minimum owed regardless of use</p></td></tr><tr><td class="firstcol " ><p>TPU providers, including Google</p></td><td  ><p>Google TPUs</p></td><td  ><p>$125–150 billion (estimated)</p></td><td  ><p>Delivered 2028–2031</p></td><td  ><p>Not stated; payments depend on capacity made available and deployment conditions</p></td></tr><tr><td class="firstcol " ><p>Amazon</p></td><td  ><p>AWS (Trainium among its chips)</p></td><td  ><p>At least $110 billion</p></td><td  ><p>May 2026–April 2036</p></td><td  ><p>No, except on Amazon’s uncured material breach; minimum owed regardless of use</p></td></tr><tr><td class="firstcol " ><p>SpaceXAI</p></td><td  ><p>Nvidia GPUs</p></td><td  ><p>Up to $120.5 billion; about $3.1 billion a month at full delivery</p></td><td  ><p>Into 2029; full capacity expected December 2026</p></td><td  ><p>Yes, on 90 days’ notice once the initial periods end in January and March 2027</p></td></tr><tr><td class="firstcol " ><p>Microsoft</p></td><td  ><p>Azure (Nvidia GPUs)</p></td><td  ><p>At least $44.9 billion</p></td><td  ><p>November 2026–May 2033</p></td><td  ><p>No, except on Microsoft’s uncured material breach or for undelivered tranches</p></td></tr><tr><td class="firstcol " ><p>Nscale</p></td><td  ><p>Nvidia Vera Rubin NVL72</p></td><td  ><p>Up to $44.6 billion</p></td><td  ><p>Into 2033</p></td><td  ><p>No, except in limited cases (missed delivery or availability targets, uncured material breach); billed only as capacity is delivered and accepted</p></td></tr><tr><td class="firstcol " ><p>Lambda</p></td><td  ><p>Nvidia GPUs</p></td><td  ><p>About $41.3 billion</p></td><td  ><p>Into 2034</p></td><td  ><p>Partly: either side can end $34.9 billion of it if Lambda doesn’t secure the facility lease in time; otherwise only on Lambda’s uncured material breach or missed delivery milestones</p></td></tr><tr><td class="firstcol " ><p>AMD</p></td><td  ><p>Instinct MI450 (Helios)</p></td><td  ><p>Expected to exceed $20 billion; no fixed dollar commitment</p></td><td  ><p>First capacity in 2027</p></td><td  ><p>No, except on AMD’s uncured material breach</p></td></tr><tr><td class="firstcol " ><p>Total, deals named in the S-1 (our sum)</p></td><td  ><p>All of the above</p></td><td  ><p>About $782–807 billion</p></td><td  ><p>2026–2036</p></td><td  ><p>Mix of minimums, caps and estimates; the S-1 leaves its own total blank and mentions other deals it doesn’t describe</p></td></tr></tbody></table></div><p>Anthropic’s minimum purchase commitments, which exclude equipment leases such as Broadcom’s, have gone from about $55 billion at the end of last year to about $265.6 billion by the end of Q2 this year. Since then, it has signed about $245 billion in new minimums, leases included, and TPUs are the biggest line overall. The Nvidia capacity, Microsoft’s aside, is also the easiest for Anthropic to walk away from.</p><p>With all these multi-year commitments, bought well ahead of delivery, expenses are paid and recognized “as compute becomes available to us,” the company says. That compute base “did not fluctuate with in-period demand,” so revenue per compute dollar increased as demand surged. When demand outruns that capacity, Anthropic buys short-term compute, which “typically carries a higher cost per unit.”</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1600px;"><p class="vanilla-image-block" style="padding-top:62.50%;"><img id="Buowiok5uwNMfzbyNPiWWA" name="anthropic-s1-revenue-per-compute-dollar" alt="Bar chart of Anthropic’s quarterly revenue and compute expenses from Q1 2025 to Q2 2026: revenue per compute dollar rose from $0.40 to $2.08, and revenue passed compute in Q1 2026." src="https://cdn.mos.cms.futurecdn.net/Buowiok5uwNMfzbyNPiWWA-1920-80.png" mos="" align="middle" fullscreen="" width="1600" height="1000" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Future)</span></figcaption></figure><p>Our $103 billion estimate is straight-line, spreading each deal evenly over its term except SpaceXAI, which is at its full monthly rate, while payments ramp up as capacity is delivered. Also not included in the estimate is interest on the Broadcom leases. Compute also does not necessarily have to earn revenue, especially in R&amp;D.</p><p>The filing has many blanks, such as the total commitment figure, the contractual obligations table, Q3 revenue, and Run-Rate Revenue, which an amended S-1 would normally fill in before pricing. No date for that has been given. Microsoft’s minimum starts in November, and SpaceXAI’s full delivery is expected in December, with non-cancelable periods ending in January and March 2027, after which either side can cancel on 90 days’ notice. More Broadcom TPU capacity and AMD’s first tranche follow next year, by which point most of these deals will be billing.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-ipo-filing-shows-most-of-its-compute-goes-to-r-and-d-not-serving-customers-company-details-massive-usd800-billion-in-tpu-and-gpu-contracts</link>
                                                                            <description>
                            <![CDATA[ Anthropic’s S-1 splits its $8.9 billion H1 compute bill: most went to R&D, with less than a third estimated to serve Claude customers. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">UdB7bAYAHHxn7gbo5pcdf</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/XL3SdCXJKQ5duBipYp56rG-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Thu, 29 Oct 2026 11:30:00 +0000</pubDate>                                                                                                                                <updated>Thu, 29 Oct 2026 13:10:36 +0000</updated>
                                                                                                                                            <category><![CDATA[Artificial Intelligence]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/XL3SdCXJKQ5duBipYp56rG-1920-80.jpg">
                                                            <media:credit><![CDATA[Google]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Rows of Google Ironwood TPU racks in a data center, with blue, green and yellow cabling under orange overhead cable trays.]]></media:description>                                                            <media:text><![CDATA[Rows of Google Ironwood TPU racks in a data center, with blue, green and yellow cabling under orange overhead cable trays.]]></media:text>
                                <media:title type="plain"><![CDATA[Rows of Google Ironwood TPU racks in a data center, with blue, green and yellow cabling under orange overhead cable trays.]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/XL3SdCXJKQ5duBipYp56rG-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Anthropic spent less than a third of its $8.9 billion first-half compute bill on serving customers, with most of the rest going into training, research, and running Claude internally, <a href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm">its S-1 filing</a> shows. It has meanwhile signed up for hundreds of billions of dollars more in TPU and GPU capacity. By our math, it has meanwhile signed up for roughly $800 billion more in TPU and GPU capacity, and revenue per compute dollar jumped from $0.40 in Q1 2025 to $2.08 in Q2 2026, which the company links to capacity purchased 12–18 months ahead.</p><p><a href="https://www.tomshardware.com/tech-industry/anthropic-files-to-go-public-claude-maker-races-openai-and-spacex-to-ipo">Anthropic’s IPO</a> is likely to beat that of rival OpenAI, which <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-hit-with-sweeping-probe-from-massive-coalition-of-42-us-state-attorneys-general-just-days-after-reported-ipo-filing-subpoena-targets-chatgpt-makers-ads-data-practices-handling-of-minors-model-sycophancy-and-safety-policies">filed confidentially</a> before CEO Sam Altman said its own won’t come this year, amid widespread doubts about an AI bubble. Whether AI labs can cover what it costs to serve their models is an open question.</p><p>Anthropic appears to do so easily: it spent $2.4 billion serving customers in the first half of the year, versus $16.3 billion in revenue. Paying for what it has on order is more difficult. Its five compute deals with the biggest stated amounts would cost an estimated $103 billion a year on average once delivered, more than twice its revenue at Q2’s pace of about $46 billion a year. The company also warns that margins may compress as new hardware lands ahead of the revenue it will support.</p><p>The company has three main costs: the cost of revenue, which is the inference cost of running Claude for paying customers; R&amp;D, which is mostly compute for training and research into new models plus any internal use; and sales and marketing, which covers the cut taken by channel partners and cloud platform resellers.</p><p>Comparing the first half of last year to this year’s, cost of revenue is up about 2.3x, but with revenue up 14x, that means Anthropic keeps an estimated 85 cents for every dollar after serving costs, up from just 12 cents. R&amp;D is up about 5 times, with 66% or more of the compute going to it, and that’s the main source of its operating loss. Sales and marketing is up 16x, and doesn’t show up in cost of revenue, but its partner fees and marketing would lower what it keeps to 66 cents a dollar. The trend here is that serving Claude now more than covers its costs.</p><div ><table><caption>Where Anthropic’s Compute Goes</caption><thead><tr><th class="firstcol empty" ></th><th  ><p>First half of 2025</p></th><th  ><p>First half of 2026</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Revenue</p></td><td  ><p>$1.16 billion</p></td><td  ><p>$16.29 billion</p></td></tr><tr><td class="firstcol " ><p>Compute and infrastructure, all workloads</p></td><td  ><p>$2.33 billion</p></td><td  ><p>$8.95 billion</p></td></tr><tr><td class="firstcol " ><p>Serving customers (cost of revenue)</p></td><td  ><p>$1.02 billion</p></td><td  ><p>$2.38 billion</p></td></tr><tr><td class="firstcol " ><p>Free access to Claude, at most (est.)</p></td><td  ><p>$119 million</p></td><td  ><p>$640 million</p></td></tr><tr><td class="firstcol " ><p>Training, research and internal use, at least (est.)</p></td><td  ><p>$1.20 billion</p></td><td  ><p>$5.93 billion</p></td></tr><tr><td class="firstcol " ><p>Serving share of compute, at most</p></td><td  ><p>44%</p></td><td  ><p>27%</p></td></tr><tr><td class="firstcol " ><p>Revenue left after serving costs</p></td><td  ><p>12%</p></td><td  ><p>85%</p></td></tr><tr><td class="firstcol " ><p>Revenue left after serving costs, partner fees and marketing</p></td><td  ><p>2%</p></td><td  ><p>66%</p></td></tr></tbody></table></div><p>Serving has kept getting cheaper for customers since June: in September, Anthropic’s Opus 5.5 model cut costs per task by about 40% versus Opus 5, says the company, and Fable 5.1 cut the cost of typical workloads by about 25% against Fable 5; each came from “a combination of a lower price and improved efficiency,” made “while preserving our margins.”</p><p>As for compute, which runs through all three costs, Anthropic runs it as “a single, shared resource across all workloads” and doesn’t disclose how it decides the split between training, research, and serving. Serving was limited, as the company has “consistently operated in a compute-constrained environment where customer demand exceeds available compute,” which suggests serving costs were muted by a lack of supply, not just efficiency.</p><p>Anthropic runs a “multi-chip compute architecture across AWS Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs,” without giving any percentage splits or GW measurements. For TPUs, it has five-year equipment <a href="https://www.tomshardware.com/tech-industry/broadcom-expands-anthropic-deal-to-3-5gw-of-google-tpu-capacity-from-2027">leases with Broadcom</a> for “TPU-based compute capacity,” with rent of “approximately $161.2 billion excluding interest”; some of that capacity came online this year, with more from 2027. Another $114.5 billion or more is <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropic-signs-deal-with-google-cloud-to-expand-tpu-chip-capacity-ai-company-expects-to-have-over-1gw-of-processing-power-in-2026">with Google</a> “between April 2026 and July 2033,” with more from Google and other TPU providers from 2028 to 2031, estimated to be in the $125–$150 billion range.</p><p>Trainium comes through Amazon, with a deal for “at least $110.0 billion between May 2026 and April 2036,” though the filing doesn’t split AWS spend by chip. Nvidia GPUs come through a series of deals, including “up to $120.5 billion … extending into 2029” <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/musks-spacex-has-rented-out-access-to-its-supercomputers-220-000-nvidia-gpus-and-300-megawatts-of-ai-compute-power-to-rival-anthropic-musk-says-no-one-set-off-my-evil-detector-antrhropic-also-interested-in-orbital-data-centers">with SpaceXAI</a>, “at least $44.9 billion” extending into 2033 with Microsoft, “up to approximately $44.6 billion” also into 2033 with Nscale, and about $41.3 billion into 2034 with Lambda. There is also an <a href="https://www.tomshardware.com/tech-industry/amd-to-supply-anthropic-with-2-gigawatts-of-instinct-mi450-gpus">AMD hardware supply agreement</a> which, in aggregate, may “exceed $20.0 billion” if Anthropic takes the full initial capacity. These deals each have their own conditions for termination.</p><div ><table><caption>Anthropic’s Compute Contracts</caption><thead><tr><th class="firstcol " ><p>Partner</p></th><th  ><p>Hardware</p></th><th  ><p>Amount</p></th><th  ><p>Term</p></th><th  ><p>Can Anthropic walk away?</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Broadcom</p></td><td  ><p>Google TPUs (equipment leases)</p></td><td  ><p>About $161.2 billion in rent</p></td><td  ><p>Five-year leases; some capacity online in 2026, more from 2027</p></td><td  ><p>No, except on default</p></td></tr><tr><td class="firstcol " ><p>Google</p></td><td  ><p>Google TPUs</p></td><td  ><p>At least $114.5 billion</p></td><td  ><p>April 2026–July 2033</p></td><td  ><p>No, except on Google’s uncured material breach; minimum owed regardless of use</p></td></tr><tr><td class="firstcol " ><p>TPU providers, including Google</p></td><td  ><p>Google TPUs</p></td><td  ><p>$125–150 billion (estimated)</p></td><td  ><p>Delivered 2028–2031</p></td><td  ><p>Not stated; payments depend on capacity made available and deployment conditions</p></td></tr><tr><td class="firstcol " ><p>Amazon</p></td><td  ><p>AWS (Trainium among its chips)</p></td><td  ><p>At least $110 billion</p></td><td  ><p>May 2026–April 2036</p></td><td  ><p>No, except on Amazon’s uncured material breach; minimum owed regardless of use</p></td></tr><tr><td class="firstcol " ><p>SpaceXAI</p></td><td  ><p>Nvidia GPUs</p></td><td  ><p>Up to $120.5 billion; about $3.1 billion a month at full delivery</p></td><td  ><p>Into 2029; full capacity expected December 2026</p></td><td  ><p>Yes, on 90 days’ notice once the initial periods end in January and March 2027</p></td></tr><tr><td class="firstcol " ><p>Microsoft</p></td><td  ><p>Azure (Nvidia GPUs)</p></td><td  ><p>At least $44.9 billion</p></td><td  ><p>November 2026–May 2033</p></td><td  ><p>No, except on Microsoft’s uncured material breach or for undelivered tranches</p></td></tr><tr><td class="firstcol " ><p>Nscale</p></td><td  ><p>Nvidia Vera Rubin NVL72</p></td><td  ><p>Up to $44.6 billion</p></td><td  ><p>Into 2033</p></td><td  ><p>No, except in limited cases (missed delivery or availability targets, uncured material breach); billed only as capacity is delivered and accepted</p></td></tr><tr><td class="firstcol " ><p>Lambda</p></td><td  ><p>Nvidia GPUs</p></td><td  ><p>About $41.3 billion</p></td><td  ><p>Into 2034</p></td><td  ><p>Partly: either side can end $34.9 billion of it if Lambda doesn’t secure the facility lease in time; otherwise only on Lambda’s uncured material breach or missed delivery milestones</p></td></tr><tr><td class="firstcol " ><p>AMD</p></td><td  ><p>Instinct MI450 (Helios)</p></td><td  ><p>Expected to exceed $20 billion; no fixed dollar commitment</p></td><td  ><p>First capacity in 2027</p></td><td  ><p>No, except on AMD’s uncured material breach</p></td></tr><tr><td class="firstcol " ><p>Total, deals named in the S-1 (our sum)</p></td><td  ><p>All of the above</p></td><td  ><p>About $782–807 billion</p></td><td  ><p>2026–2036</p></td><td  ><p>Mix of minimums, caps and estimates; the S-1 leaves its own total blank and mentions other deals it doesn’t describe</p></td></tr></tbody></table></div><p>Anthropic’s minimum purchase commitments, which exclude equipment leases such as Broadcom’s, have gone from about $55 billion at the end of last year to about $265.6 billion by the end of Q2 this year. Since then, it has signed about $245 billion in new minimums, leases included, and TPUs are the biggest line overall. The Nvidia capacity, Microsoft’s aside, is also the easiest for Anthropic to walk away from.</p><p>With all these multi-year commitments, bought well ahead of delivery, expenses are paid and recognized “as compute becomes available to us,” the company says. That compute base “did not fluctuate with in-period demand,” so revenue per compute dollar increased as demand surged. When demand outruns that capacity, Anthropic buys short-term compute, which “typically carries a higher cost per unit.”</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1600px;"><p class="vanilla-image-block" style="padding-top:62.50%;"><img id="Buowiok5uwNMfzbyNPiWWA" name="anthropic-s1-revenue-per-compute-dollar" alt="Bar chart of Anthropic’s quarterly revenue and compute expenses from Q1 2025 to Q2 2026: revenue per compute dollar rose from $0.40 to $2.08, and revenue passed compute in Q1 2026." src="https://cdn.mos.cms.futurecdn.net/Buowiok5uwNMfzbyNPiWWA-1920-80.png" mos="" align="middle" fullscreen="" width="1600" height="1000" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Future)</span></figcaption></figure><p>Our $103 billion estimate is straight-line, spreading each deal evenly over its term except SpaceXAI, which is at its full monthly rate, while payments ramp up as capacity is delivered. Also not included in the estimate is interest on the Broadcom leases. Compute also does not necessarily have to earn revenue, especially in R&amp;D.</p><p>The filing has many blanks, such as the total commitment figure, the contractual obligations table, Q3 revenue, and Run-Rate Revenue, which an amended S-1 would normally fill in before pricing. No date for that has been given. Microsoft’s minimum starts in November, and SpaceXAI’s full delivery is expected in December, with non-cancelable periods ending in January and March 2027, after which either side can cancel on 90 days’ notice. More Broadcom TPU capacity and AMD’s first tranche follow next year, by which point most of these deals will be billing.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ FCC rebuffs TP-Link again as it attempts to sell Wi-Fi 8 gear in the US ]]></title>
                                                                                                <dc:content><![CDATA[ <p>TP-Link has been trying everything in its power to gain conditional approval from the Federal Communications Commission (FCC). According to a new <a href="https://www.bloomberg.com/news/articles/2026-10-27/fcc-warns-tp-link-its-routers-breach-power-limits-in-latest-snub"><u><em>Bloomberg</em></u><u> report</u></a>, the FCC is citing discrepancies in power levels in TP-Link routers as the latest roadblock on its path to <a href="https://www.tomshardware.com/networking/routers/tp-link-opens-global-preorders-for-its-first-wi-fi-8-router-but-ban-keeps-it-out-of-us-market-us-left-off-launch-list-as-fcc-freeze-refuses-to-thaw"><u>selling new Wi-Fi 8 products stateside</u></a>. </p><p>In letters dated October 15, 2026, the FCC Office of Engineering and Technology made some damning claims about TP-Link's reporting practices. Specifically, the FCC alleged that TP-Link "included false or incorrect information in its applications” related to equipment authorizations for four individual router models (including <a href="https://www.tomshardware.com/networking/routers/tp-link-announces-its-first-consumer-wi-fi-8-routers-archer-8-ultra-preorder-commences-september-30-in-select-regions"><u>high-end Archer models</u></a>). </p><p>For a company that truly needs to dot every "i" and cross every "t" when it comes to a Trump administration that is increasingly skeptical of TP-Link's past ties and motives, any perceived missteps are even more magnified. However, in addition to the alleged wonky product information provided to the FCC, perhaps more troubling is that three of the four routers under scrutiny also had higher-than-allowed power usage levels.</p><p>In addition to power limits, the FCC's testing regimen for routers includes evaluations of electromagnetic compatibility (EMC), radio frequency (RF) emissions, and <a href="https://www.tomshardware.com/networking/routers/tp-link-routers-face-potential-u-s-ban-over-alleged-china-related-national-security-concerns-company-vigorously-disputes-department-of-commerces-findings"><u>national security compliance</u></a>. The latter has been a big sticking point with the FCC and the U.S. Department of Defense given the proliferation of TP-Link networking gear in civilian and governmental capacities.</p><p>We reached out to TP-Link about the FCC’s findings and were provided with the following statement:</p><p><em>"Routine spot checks have been a standard FCC practice across the consumer electronics industry for many years, and individual product test results can occasionally vary. </em></p><p><em>TP-Link Systems fully supports the FCC’s Office of Engineering and Technology and its role of ensuring that all manufacturers adhere to prescribed specifications for their certified products. We remain fully committed to working collaboratively with the FCC to resolve any technical questions that arise — and if necessary, address any confirmed technical issues as swiftly as possible."</em> </p><p>It will likely take a lot more convincing to appease the FCC, which has so far given TP-Link the cold shoulder. And it’s not just the FCC; <a href="https://www.tomshardware.com/networking/routers/tp-link-routers-face-potential-u-s-ban-over-alleged-china-related-national-security-concerns-company-vigorously-disputes-department-of-commerces-findings"><u>members of Congress</u></a> and multiple state attorneys general have filed their grievances with TP-Link over alleged consumer privacy violations and Chinese ties.</p><p>For its part, TP-Link has made some significant moves to get in the good graces of the U.S. government, even while competitors like <a href="https://www.tomshardware.com/networking/routers/amazon-eero-and-leo-routers-gain-fcc-conditional-approval-for-us-sales-eero-products-can-skirt-router-ban-for-the-next-18-months-firm-joins-netgear-on-approval-list"><u>Amazon</u></a>, <a href="https://www.tomshardware.com/networking/routers/asus-routers-gain-fccs-conditional-approval-for-sale-in-the-us-as-tp-link-remains-locked-out-asuss-wi-fi-8-ambitions-remain-intact#xenforo-comments-3900280"><u>Asus</u></a>, and <a href="https://www.tomshardware.com/networking/routers/netgear-secures-conditional-approval-from-the-fcc-following-router-ban-company-can-continue-importing-foreign-made-routers-through-october-2027"><u>Netgear</u></a> have sailed through compliance checks to gain conditional approval to sell wireless products in the U.S. Besides moving its U.S. headquarters to California, the company earlier this month announced a <a href="https://www.tomshardware.com/networking/routers/tp-link-proposes-usd1-billion-us-wi-fi-router-manufacturing-alliance-with-key-tronic-potential-deal-seeks-to-end-ongoing-saga-blocking-its-wi-fi-8-products-from-american-homes"><u>$1 billion joint venture with Key Tronic</u></a> to manufacture Wi-Fi 8 routers in the U.S. </p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/networking/routers/fcc-rebuffs-tp-link-again-as-it-attempts-to-sell-wi-fi-8-gear-in-the-us-fcc-alleges-inaccurate-documentation-and-out-of-spec-power-levels-for-routers</link>
                                                                            <description>
                            <![CDATA[ TP-Link just can’t seem to catch a break with the FCC, as it’s not under fire for exceeding power limits on its routers ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">CoJM7p5hMYcpiptS3toG7X</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/MbZxghsMGayruAPEDs86j7-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Wed, 28 Oct 2026 15:07:35 +0000</pubDate>                                                                                                                                <updated>Wed, 28 Oct 2026 15:43:47 +0000</updated>
                                                                                                                                            <category><![CDATA[Wi-Fi Routers]]></category>
                                                    <category><![CDATA[Networking]]></category>
                                                                                                <author><![CDATA[ brandon.hill@futurenet.com (Brandon Hill) ]]></author>                    <dc:creator><![CDATA[ Brandon Hill ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/yHeufe7JcvuJBhYPkSexNf-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Brandon has been tinkering with PCs since childhood and received his first &quot;real&quot; PC, an IBM Aptiva 310, in the mid-1990s. He next went on to build his first custom PC with an Intel Celeron 300A processor overclocked to 450MHz on an Abit BH6 motherboard. Brandon has written about PC and Mac tech since the late 1990s, first at AnandTech before moving to DailyTech and later to Hot Hardware. When Brandon is not consuming copious amounts of tech news, he can be found enjoying the NC mountains or the beach with his wife and two sons.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/MbZxghsMGayruAPEDs86j7-1920-80.jpg">
                                                            <media:credit><![CDATA[TP-Link]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[TP-Link Wifi Router Render]]></media:description>                                                            <media:text><![CDATA[TP-Link Wifi Router Render]]></media:text>
                                <media:title type="plain"><![CDATA[TP-Link Wifi Router Render]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/MbZxghsMGayruAPEDs86j7-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>TP-Link has been trying everything in its power to gain conditional approval from the Federal Communications Commission (FCC). According to a new <a href="https://www.bloomberg.com/news/articles/2026-10-27/fcc-warns-tp-link-its-routers-breach-power-limits-in-latest-snub"><u><em>Bloomberg</em></u><u> report</u></a>, the FCC is citing discrepancies in power levels in TP-Link routers as the latest roadblock on its path to <a href="https://www.tomshardware.com/networking/routers/tp-link-opens-global-preorders-for-its-first-wi-fi-8-router-but-ban-keeps-it-out-of-us-market-us-left-off-launch-list-as-fcc-freeze-refuses-to-thaw"><u>selling new Wi-Fi 8 products stateside</u></a>. </p><p>In letters dated October 15, 2026, the FCC Office of Engineering and Technology made some damning claims about TP-Link's reporting practices. Specifically, the FCC alleged that TP-Link "included false or incorrect information in its applications” related to equipment authorizations for four individual router models (including <a href="https://www.tomshardware.com/networking/routers/tp-link-announces-its-first-consumer-wi-fi-8-routers-archer-8-ultra-preorder-commences-september-30-in-select-regions"><u>high-end Archer models</u></a>). </p><p>For a company that truly needs to dot every "i" and cross every "t" when it comes to a Trump administration that is increasingly skeptical of TP-Link's past ties and motives, any perceived missteps are even more magnified. However, in addition to the alleged wonky product information provided to the FCC, perhaps more troubling is that three of the four routers under scrutiny also had higher-than-allowed power usage levels.</p><p>In addition to power limits, the FCC's testing regimen for routers includes evaluations of electromagnetic compatibility (EMC), radio frequency (RF) emissions, and <a href="https://www.tomshardware.com/networking/routers/tp-link-routers-face-potential-u-s-ban-over-alleged-china-related-national-security-concerns-company-vigorously-disputes-department-of-commerces-findings"><u>national security compliance</u></a>. The latter has been a big sticking point with the FCC and the U.S. Department of Defense given the proliferation of TP-Link networking gear in civilian and governmental capacities.</p><p>We reached out to TP-Link about the FCC’s findings and were provided with the following statement:</p><p><em>"Routine spot checks have been a standard FCC practice across the consumer electronics industry for many years, and individual product test results can occasionally vary. </em></p><p><em>TP-Link Systems fully supports the FCC’s Office of Engineering and Technology and its role of ensuring that all manufacturers adhere to prescribed specifications for their certified products. We remain fully committed to working collaboratively with the FCC to resolve any technical questions that arise — and if necessary, address any confirmed technical issues as swiftly as possible."</em> </p><p>It will likely take a lot more convincing to appease the FCC, which has so far given TP-Link the cold shoulder. And it’s not just the FCC; <a href="https://www.tomshardware.com/networking/routers/tp-link-routers-face-potential-u-s-ban-over-alleged-china-related-national-security-concerns-company-vigorously-disputes-department-of-commerces-findings"><u>members of Congress</u></a> and multiple state attorneys general have filed their grievances with TP-Link over alleged consumer privacy violations and Chinese ties.</p><p>For its part, TP-Link has made some significant moves to get in the good graces of the U.S. government, even while competitors like <a href="https://www.tomshardware.com/networking/routers/amazon-eero-and-leo-routers-gain-fcc-conditional-approval-for-us-sales-eero-products-can-skirt-router-ban-for-the-next-18-months-firm-joins-netgear-on-approval-list"><u>Amazon</u></a>, <a href="https://www.tomshardware.com/networking/routers/asus-routers-gain-fccs-conditional-approval-for-sale-in-the-us-as-tp-link-remains-locked-out-asuss-wi-fi-8-ambitions-remain-intact#xenforo-comments-3900280"><u>Asus</u></a>, and <a href="https://www.tomshardware.com/networking/routers/netgear-secures-conditional-approval-from-the-fcc-following-router-ban-company-can-continue-importing-foreign-made-routers-through-october-2027"><u>Netgear</u></a> have sailed through compliance checks to gain conditional approval to sell wireless products in the U.S. Besides moving its U.S. headquarters to California, the company earlier this month announced a <a href="https://www.tomshardware.com/networking/routers/tp-link-proposes-usd1-billion-us-wi-fi-router-manufacturing-alliance-with-key-tronic-potential-deal-seeks-to-end-ongoing-saga-blocking-its-wi-fi-8-products-from-american-homes"><u>$1 billion joint venture with Key Tronic</u></a> to manufacture Wi-Fi 8 routers in the U.S. </p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Why native PC ports of console games are suddenly everywhere ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Over the past weeks and months, it's increasingly commonplace to encounter native PC ports of previously console-exclusive video games, developed <a href="https://www.tomshardware.com/video-games/retro-gaming/mario-kart-wii-recompiled-for-pc-using-ai-with-4k-potential-and-uncapped-frame-rates-first-static-recompilation-of-a-wii-game-supports-over-200-tracks-thanks-to-retro-rewind-compatibility">with the help of frontier AI models</a> and a new tool named Reverse Engineer Anything (REA). With <a href="https://www.tomshardware.com/video-games/retro-gaming/goldeneye-007-for-n64-has-been-100-percent-decompiled-success-of-half-decade-project-opens-up-possibilities-for-complex-mods-and-ports">dozens of titles</a> ported to PC over the past few months, the big questions are: why is this now possible, and what changed? </p><p>The better-known conversions like <em>Super Mario Galaxy</em> and <em>Modern Warfare 2</em> use AI agents, but their handlers mostly still had to interpret results and direct each step of the process. That's now much easier with <a href="https://github.com/morluto/rea">Reverse Engineer Anything</a>. Although it shows up in many conversations, REA doesn't do the heavy lifting by itself. It's actually a mix between a toolkit and agent harness, wiring together previously disparate tools under one ceiling.</p><p>REA analyzes and manages collected information tools: binary analyzers, decompilers, metadata tools, firmware images, database readers, and so on end up linked via agents. Those tools are made available to your AI agent of choice, along with skill files with additional instructions. </p><p>Simplistic as it may sound, it's the glue that once again significantly cuts down the time spent on "chores" that a developer would otherwise have to do while reverse-engineering code, which started with emulating consoles themselves.</p><p>Emulation of video game consoles on PC is nothing new, having emerged in the 90s in the 2D era, and gradually transitioning to 3D consoles such as the PlayStation 1, and Nintendo 64. Throughout the early 2020s, emulators and toolkits evolved at a rapid pace, covering more systems than ever — even <a href="https://en.wikipedia.org/wiki/PlayStation_3">the PlayStation 3</a> and its notorious Cell processor. Physical hardware is taken apart and inspected more easily and at deeper levels than ever, as tools like the Raspberry Pi become cheaper and more powerful. These elements set the stage for the ultimate task: to fully decompile and recompile a game and convert it to run on native PC instructions, rather than relying on software emulation. Only one barrier remained: dedicated development time.</p><p>To reverse-engineer a game, you need dozens of different tools, analyze the game's executable, run it, inspect which files it reads, which calls it makes to the operating system, what memory it allocates, and so on and so forth. This is exceedingly tedious, and yet can't be done half-heartedly: one errant byte is a bug. It's like undoing a machine with a million cogs and keeping track exactly where each one was, its size, and orientation. Modern games grew, too, and can have millions of lines of code. </p><p>But that kind of nitpicky, repetitive work is just what computers excel at — even non-deterministic ones like LLMs. 2026's AI models like <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-claims-gpt-6-astra-is-an-ethereal-alien-mind-with-agi-like-qualities-company-warns-of-alignment-challenges-as-new-frontier-leader-emerges">GPT-6 Astra</a> and Claude 5.5 families are "smarter" in the sense that they follow complicated instructions more closely and can better keep track of their tasks, granting them a higher degree of autonomy. The high token efficiency of <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/frontier-ai-faces-pricing-reckoning-as-token-volume-explodes-25-fold-mid-tier-models-deliver-90-percent-of-flagship-capability-at-one-sixth-the-cost">lower-grade variants like Claude Sonnet</a> also makes them somewhat affordable, if not cheap, to use on a humongous project like taking apart an entire game.</p><p>Given how AI bots can hallucinate and make up strange replies and code, you may wonder how a project of this scale ever gets accomplished. It's simple: your agent hands you decompiled source code, and it compiles it again to see if it matches the original byte-for-byte. Even if they hallucinate, the bots can try over and over until they get it right, and developers can correct them if needed.</p><p>Reverse-engineering needs multiple tools, all with their own technical terms and idiosyncrasies. Whereas before developers had to wrangle them across terminal windows, now you can get your agent to twiddle bits and talk between programs on your behalf, so long as you issue the right orders. They don't have to stop for breaks, either.</p><p>With relatively simple prompts, you can poke around and break down a game or application — provided you have one of the pricier agent subscriptions, or can absorb high token bills, as for example the <em>Modern Warfare 2</em> decomp <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/dev-used-ai-agents-and-more-than-500-billion-tokens-to-rebuild-call-of-duty-modern-warfare-2s-source-code-83-percent-of-the-2009-shooters-functions-now-match-the-original-byte-for-byte">took three months</a> and up to 700 <em>billion</em> tokens to accomplish. Had its developer used REA, he may have cut down on time and tokens both.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/video-games/why-native-pc-ports-of-console-games-are-suddenly-everywhere-reverse-engineer-anything-gives-ai-agents-a-unified-toolkit-and-byte-for-byte-checks-keep-hallucinations-in-line</link>
                                                                            <description>
                            <![CDATA[ With dozens of titles unofficially ported to PC over the past few months, the big questions are: why is this now possible, and what changed? ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">o3HytLfWPTTrNTfRk5KZR</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/ngAU8FJ5HzYQKpYhy4eVBM-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Wed, 28 Oct 2026 12:00:00 +0000</pubDate>                                                                                                                                <updated>Wed, 28 Oct 2026 17:22:00 +0000</updated>
                                                                                                                                            <category><![CDATA[Video Games]]></category>
                                                                                                <author><![CDATA[ editors@tomshardware.com (Bruno Ferreira) ]]></author>                    <dc:creator><![CDATA[ Bruno Ferreira ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/ZQiPPaXaAuQ4VrVEYnnR7G-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Bruno Ferreira&#039;s journey kicked off with the venerable ZX Spectrum, a cassette player, and his hopes and dreams. He quickly realized he had more fun figuring out how computers work than he did actually using the things. Kicking off a developer career with C and Assembly before moving to scripting languages, he&#039;s worn many hats, including both database architect and systems administration. As a teen, Bruno co-founded a web development outfit where he was for 17 years before moving on to spend nearly a decade at The Tech Report as a writer, editor, and (of course) developer. In this decade, he&#039;s been at Asus, MLCommons, and HotHardware, among others. When not fiddling with computers and games, his love for music and production sends him off to live shows and festivals. Occasionally, he pretends he can play the guitar and bass.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/ngAU8FJ5HzYQKpYhy4eVBM-1920-80.jpg">
                                                            <media:credit><![CDATA[Getty Images]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Robot Render in an abstract computerized landscape]]></media:description>                                                            <media:text><![CDATA[Robot Render in an abstract computerized landscape]]></media:text>
                                <media:title type="plain"><![CDATA[Robot Render in an abstract computerized landscape]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/ngAU8FJ5HzYQKpYhy4eVBM-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Over the past weeks and months, it's increasingly commonplace to encounter native PC ports of previously console-exclusive video games, developed <a href="https://www.tomshardware.com/video-games/retro-gaming/mario-kart-wii-recompiled-for-pc-using-ai-with-4k-potential-and-uncapped-frame-rates-first-static-recompilation-of-a-wii-game-supports-over-200-tracks-thanks-to-retro-rewind-compatibility">with the help of frontier AI models</a> and a new tool named Reverse Engineer Anything (REA). With <a href="https://www.tomshardware.com/video-games/retro-gaming/goldeneye-007-for-n64-has-been-100-percent-decompiled-success-of-half-decade-project-opens-up-possibilities-for-complex-mods-and-ports">dozens of titles</a> ported to PC over the past few months, the big questions are: why is this now possible, and what changed? </p><p>The better-known conversions like <em>Super Mario Galaxy</em> and <em>Modern Warfare 2</em> use AI agents, but their handlers mostly still had to interpret results and direct each step of the process. That's now much easier with <a href="https://github.com/morluto/rea">Reverse Engineer Anything</a>. Although it shows up in many conversations, REA doesn't do the heavy lifting by itself. It's actually a mix between a toolkit and agent harness, wiring together previously disparate tools under one ceiling.</p><p>REA analyzes and manages collected information tools: binary analyzers, decompilers, metadata tools, firmware images, database readers, and so on end up linked via agents. Those tools are made available to your AI agent of choice, along with skill files with additional instructions. </p><p>Simplistic as it may sound, it's the glue that once again significantly cuts down the time spent on "chores" that a developer would otherwise have to do while reverse-engineering code, which started with emulating consoles themselves.</p><p>Emulation of video game consoles on PC is nothing new, having emerged in the 90s in the 2D era, and gradually transitioning to 3D consoles such as the PlayStation 1, and Nintendo 64. Throughout the early 2020s, emulators and toolkits evolved at a rapid pace, covering more systems than ever — even <a href="https://en.wikipedia.org/wiki/PlayStation_3">the PlayStation 3</a> and its notorious Cell processor. Physical hardware is taken apart and inspected more easily and at deeper levels than ever, as tools like the Raspberry Pi become cheaper and more powerful. These elements set the stage for the ultimate task: to fully decompile and recompile a game and convert it to run on native PC instructions, rather than relying on software emulation. Only one barrier remained: dedicated development time.</p><p>To reverse-engineer a game, you need dozens of different tools, analyze the game's executable, run it, inspect which files it reads, which calls it makes to the operating system, what memory it allocates, and so on and so forth. This is exceedingly tedious, and yet can't be done half-heartedly: one errant byte is a bug. It's like undoing a machine with a million cogs and keeping track exactly where each one was, its size, and orientation. Modern games grew, too, and can have millions of lines of code. </p><p>But that kind of nitpicky, repetitive work is just what computers excel at — even non-deterministic ones like LLMs. 2026's AI models like <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-claims-gpt-6-astra-is-an-ethereal-alien-mind-with-agi-like-qualities-company-warns-of-alignment-challenges-as-new-frontier-leader-emerges">GPT-6 Astra</a> and Claude 5.5 families are "smarter" in the sense that they follow complicated instructions more closely and can better keep track of their tasks, granting them a higher degree of autonomy. The high token efficiency of <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/frontier-ai-faces-pricing-reckoning-as-token-volume-explodes-25-fold-mid-tier-models-deliver-90-percent-of-flagship-capability-at-one-sixth-the-cost">lower-grade variants like Claude Sonnet</a> also makes them somewhat affordable, if not cheap, to use on a humongous project like taking apart an entire game.</p><p>Given how AI bots can hallucinate and make up strange replies and code, you may wonder how a project of this scale ever gets accomplished. It's simple: your agent hands you decompiled source code, and it compiles it again to see if it matches the original byte-for-byte. Even if they hallucinate, the bots can try over and over until they get it right, and developers can correct them if needed.</p><p>Reverse-engineering needs multiple tools, all with their own technical terms and idiosyncrasies. Whereas before developers had to wrangle them across terminal windows, now you can get your agent to twiddle bits and talk between programs on your behalf, so long as you issue the right orders. They don't have to stop for breaks, either.</p><p>With relatively simple prompts, you can poke around and break down a game or application — provided you have one of the pricier agent subscriptions, or can absorb high token bills, as for example the <em>Modern Warfare 2</em> decomp <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/dev-used-ai-agents-and-more-than-500-billion-tokens-to-rebuild-call-of-duty-modern-warfare-2s-source-code-83-percent-of-the-2009-shooters-functions-now-match-the-original-byte-for-byte">took three months</a> and up to 700 <em>billion</em> tokens to accomplish. Had its developer used REA, he may have cut down on time and tokens both.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Standalone Nvidia Vera CPUs land in Oracle Cloud's multi-gigawatt deployment ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Nvidia and Oracle have announced general availability of <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-launches-vera-rubin-nvl72-ai-supercomputer-at-ces-promises-up-to-5x-greater-inference-performance-and-10x-lower-cost-per-token-than-blackwell-coming-2h-2026">Nvidia’s Vera Rubin NVL72 </a>rack and <a href="https://www.tomshardware.com/pc-components/cpus/nvidia-spills-the-beans-on-vera-cpu-spec-benchmarks-revealed-olympus-architecture-detailed-and-more">Vera CPUs </a>on Oracle Cloud Infrastructure (OCI) Superclusters at Oracle AI World on Tuesday. The announcement comes hot on the heels of Google Cloud announcing general availability for Vera Rubin last week, putting two major hyperscaler feathers in Nvidia’s cap. However, Oracle’s announcement is interesting, specifically calling out standalone Vera instances on OCI Superclusters. We've included a few benchmarks we've performed on a Vera server for a forthcoming review to highlight some of the chip's performance characteristics, but do note that these tests were not conducted on OCI infrastructure. </p><p>Neoclouds have been banging the drum of Vera for the past several weeks, with instances under way at Nscale, CoreWeave, and Nebius, among others, so it makes sense to see Vera in the GPU-dominant and AI-focused neoclouds. A standalone deployment at a hyperscaler like Oracle sends a different signal, however. </p><p>Although hyperscalers maintain large fleets of hardware, sometimes even renting from the neoclouds, the CPU focus is primarily on x86, with OCI Superclusters otherwise offering the Ampere Altra for an Arm-based instance. It’s Nvidia’s first standalone CPU push into the hyperscale data center, placing Vera alongside the x86 competition regardless of the GPU compute. </p><p>Vera is not meant to upend the x86 guard, at least in isolation. Ian Buck, Nvidia’s mind behind CUDA and the face of Nvidia’s AI infrastructure buildout, said in March — when the chip was revealed — that the world couldn’t be served by a single, 88-core SKU. However, Vera is opening the door at hyperscalers like Oracle to expand in future CPU generations. </p><p>AMD and Intel maintain the lead when it comes to single-socket density, which is only set to expand in the next generation, as AMD’s upcoming <a href="https://www.tomshardware.com/pc-components/cpus/amds-256-core-epyc-9996-venice-claims-up-to-a-3-4x-jump-over-intel-xeon-competition-20-percent-over-nvidia-vera-zen-6-comes-with-up-to-1024mb-of-l3-16-channel-memory-and-5ghz-clock-speeds">EPYC Venice CPUs</a> can pack up to 512 threads in a single socket. However, Vera offers some distinct advantages due to its topology depending on the workload, many of which are becoming increasingly relevant in the agentic AI era. </p><p>Memory bandwidth, for instance, is one of Vera’s strong suits, and that’s not just a marketing claim. Nvidia has said the memory system is capable of 1.2 TB/s of bandwidth, but we can see Nvidia’s lead outside of theoretical numbers in our testing below. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:77.49%;"><img id="HbfrtT4iSpeag52RE4z8Ac" name="Vera Benchmarks - Triad" alt="Vera OMP Spread Triad Stream benchmark" src="https://cdn.mos.cms.futurecdn.net/HbfrtT4iSpeag52RE4z8Ac-1920-80.png" mos="" align="middle" fullscreen="" width="1999" height="1549" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p>Take the industry-standard Stream benchmark as an example. Nvidia’s use of LPDDR5X-9600, along with a monolithic die and distributed mesh for caching, allows it to more than double the bandwidth available on a dual Intel Xeon 6980P system, even when it’s paired with MRDIMMs sporting speeds of 8,800 MT/s. We’re showing triad here, as it’s the most “demanding” of Stream’s operations, but the leads for Vera are similar across add, copy, and scale as well. </p><p>This bandwidth is critical for some types of agentic loops, particularly those dealing with large database queries or extensive documentation. You can see the OpenMP “spread” policy offers a massive boost to Vera’s memory bandwidth, as well. Both AMD and Intel maintain centralized memory control and system-level cache, while Vera distributes it in a mesh across the die. The result, at least in OpenMP set properly, is much higher memory bandwidth by binding processes as far away from each other as possible.</p><p>Another critical performance vector for agentic AI is Python. The agentic world is expanding, but much of the foundation of the AI software ecosystem is built on Python. Common agentic frameworks like CrewAI and LangGraph are written in Python, and although we’re seeing more frameworks built in languages like Rust and TypeScript, Python performance is still critical to agentic loops. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:77.54%;"><img id="gFfVmJSkrcQ8AqrWReeLKc" name="Vera Benchmarks - PyBench" alt="Vera Pybench Benchmarks" src="https://cdn.mos.cms.futurecdn.net/gFfVmJSkrcQ8AqrWReeLKc-1920-80.png" mos="" align="middle" fullscreen="" width="1999" height="1550" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p>And, unsurprisingly, it’s an area where Vera is also strong, beating out CPUs that are much more powerful on paper. Popular agentic harnesses are built on Python, but so are many of the applications that also run on Python, including foundational AI libraries like PyTorch and TensorFlow. </p><p>Nvidia’s focus on agentic AI makes sense given the times we’re living in, but it can miss the forest for the trees. Oracle bringing Vera standalone instances online shifts the focus onto the broader implications of Vera’s entrance into the data center. There are some aspects of performance that characterize agentic AI workloads, but many agent workloads are regular data center workloads in the first place, just under high concurrency. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:77.14%;"><img id="SFBUqg66rMWFMDAdpLWWFc" name="Vera Benchmarks - SQLite" alt="Vera SQLite benchmarks" src="https://cdn.mos.cms.futurecdn.net/SFBUqg66rMWFMDAdpLWWFc-1920-80.png" mos="" align="middle" fullscreen="" width="1999" height="1542" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p>For instance, SQLite performance is important for any server querying a large database, regardless of its written by logic, a person, or an AI agent. Nvidia’s second-gen Scalable Coherency Fabric (SCF), which we saw at play in Stream, shows up in a real workload here, with these database queries largely bound by data movement across the system. </p><p>And that is what makes Vera so interesting in the context of Oracle’s announcement. It’s a CPU built for agentic AI, says Nvidia, but the crosstalk with general data center workloads is strong. It’s not just a CPU to use because you need access to Nvidia’s accelerators. Depending on the workload, it may even be a better choice than the x86 guard. </p><p>Ultimately, we have to see how customers of these hyperscalers use (or don’t use) Vera in production. It’s demonstrated its usefulness in agentic AI loops, but it’s also showing early promise in certain general-purpose workloads. Oracle’s standalone deployment is a sign that there’s some demand for Vera that isn't exclusively tied at the hip of a Rubin GPU. </p><p>Oracle is approaching cautiously, however, as any hyperscaler should be when bringing up a multi-gigawatt deployment. The company said it will also offer Nvidia’s HGX NVL8, bringing eight Rubin GPUs together in a cluster, connected to a dual, 64-core Intel Xeon 6 host with 4TB of DDR5 memory. </p><p>Our testing of Vera shared in this article is part of a larger benchmarking session we had with the chip, and we'll share the technical details of the test setup in our forthcoming review. We’ll be publishing our full results shortly on <em>Tom’s Hardware. </em></p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-and-oracle-bring-vera-cpu-instances-online-in-multi-gigawatt-deployment-oci-superclusters-also-set-to-expand-with-rubin-only-clusters-attached-to-dual-xeon-6-hosts</link>
                                                                            <description>
                            <![CDATA[ Nvidia’s Vera CPU is now generally available on Oracle’s OCI Superclusters, a win for Nvidia as it moves past NVL72 to a standalone Vera deployment at a major hyperscaler. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">fMQHmE4jowDMWMraSyt5VA</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/VjsMJHUMPcH4dpNwMJen5P-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Wed, 28 Oct 2026 11:00:00 +0000</pubDate>                                                                                                                                <updated>Wed, 28 Oct 2026 13:15:01 +0000</updated>
                                                                                                                                            <category><![CDATA[Artificial Intelligence]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Jake Roach ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/h6PRM8bTimCTnNfoAYfjAi-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Jake Roach has been bending pins and busting solder joints since the mid-2000s. From trying to run scratched CDs of &lt;em&gt;Delta Force &lt;/em&gt;and &lt;em&gt;Unreal Tournament &lt;/em&gt;to spitting out virtual machines on a Threadripper, Jake has been on the hunt for the latest hardware and highest performance for decades. That eventually spun up a career, with Jake serving as Lead Reporter at Digital Trends, as well as contributing to outlets like XDA, PC Invasion, Business Insider, and WIRED. At Tom’s Hardware, Jake is focused on consumer and workstation CPUs. Outside working hours, you’ll find him knee-deep in the latest roguelite taking over Steam, spending way too much money on &lt;em&gt;Magic: The Gathering, &lt;/em&gt;or forcing his lazy corgi onto walks.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/VjsMJHUMPcH4dpNwMJen5P-1920-80.jpg">
                                                            <media:credit><![CDATA[Getty Images / Bloomberg]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[The Nvidia Vera CPU and PCB ]]></media:description>                                                            <media:text><![CDATA[The Nvidia Vera CPU and PCB ]]></media:text>
                                <media:title type="plain"><![CDATA[The Nvidia Vera CPU and PCB ]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/VjsMJHUMPcH4dpNwMJen5P-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Nvidia and Oracle have announced general availability of <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-launches-vera-rubin-nvl72-ai-supercomputer-at-ces-promises-up-to-5x-greater-inference-performance-and-10x-lower-cost-per-token-than-blackwell-coming-2h-2026">Nvidia’s Vera Rubin NVL72 </a>rack and <a href="https://www.tomshardware.com/pc-components/cpus/nvidia-spills-the-beans-on-vera-cpu-spec-benchmarks-revealed-olympus-architecture-detailed-and-more">Vera CPUs </a>on Oracle Cloud Infrastructure (OCI) Superclusters at Oracle AI World on Tuesday. The announcement comes hot on the heels of Google Cloud announcing general availability for Vera Rubin last week, putting two major hyperscaler feathers in Nvidia’s cap. However, Oracle’s announcement is interesting, specifically calling out standalone Vera instances on OCI Superclusters. We've included a few benchmarks we've performed on a Vera server for a forthcoming review to highlight some of the chip's performance characteristics, but do note that these tests were not conducted on OCI infrastructure. </p><p>Neoclouds have been banging the drum of Vera for the past several weeks, with instances under way at Nscale, CoreWeave, and Nebius, among others, so it makes sense to see Vera in the GPU-dominant and AI-focused neoclouds. A standalone deployment at a hyperscaler like Oracle sends a different signal, however. </p><p>Although hyperscalers maintain large fleets of hardware, sometimes even renting from the neoclouds, the CPU focus is primarily on x86, with OCI Superclusters otherwise offering the Ampere Altra for an Arm-based instance. It’s Nvidia’s first standalone CPU push into the hyperscale data center, placing Vera alongside the x86 competition regardless of the GPU compute. </p><p>Vera is not meant to upend the x86 guard, at least in isolation. Ian Buck, Nvidia’s mind behind CUDA and the face of Nvidia’s AI infrastructure buildout, said in March — when the chip was revealed — that the world couldn’t be served by a single, 88-core SKU. However, Vera is opening the door at hyperscalers like Oracle to expand in future CPU generations. </p><p>AMD and Intel maintain the lead when it comes to single-socket density, which is only set to expand in the next generation, as AMD’s upcoming <a href="https://www.tomshardware.com/pc-components/cpus/amds-256-core-epyc-9996-venice-claims-up-to-a-3-4x-jump-over-intel-xeon-competition-20-percent-over-nvidia-vera-zen-6-comes-with-up-to-1024mb-of-l3-16-channel-memory-and-5ghz-clock-speeds">EPYC Venice CPUs</a> can pack up to 512 threads in a single socket. However, Vera offers some distinct advantages due to its topology depending on the workload, many of which are becoming increasingly relevant in the agentic AI era. </p><p>Memory bandwidth, for instance, is one of Vera’s strong suits, and that’s not just a marketing claim. Nvidia has said the memory system is capable of 1.2 TB/s of bandwidth, but we can see Nvidia’s lead outside of theoretical numbers in our testing below. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:77.49%;"><img id="HbfrtT4iSpeag52RE4z8Ac" name="Vera Benchmarks - Triad" alt="Vera OMP Spread Triad Stream benchmark" src="https://cdn.mos.cms.futurecdn.net/HbfrtT4iSpeag52RE4z8Ac-1920-80.png" mos="" align="middle" fullscreen="" width="1999" height="1549" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p>Take the industry-standard Stream benchmark as an example. Nvidia’s use of LPDDR5X-9600, along with a monolithic die and distributed mesh for caching, allows it to more than double the bandwidth available on a dual Intel Xeon 6980P system, even when it’s paired with MRDIMMs sporting speeds of 8,800 MT/s. We’re showing triad here, as it’s the most “demanding” of Stream’s operations, but the leads for Vera are similar across add, copy, and scale as well. </p><p>This bandwidth is critical for some types of agentic loops, particularly those dealing with large database queries or extensive documentation. You can see the OpenMP “spread” policy offers a massive boost to Vera’s memory bandwidth, as well. Both AMD and Intel maintain centralized memory control and system-level cache, while Vera distributes it in a mesh across the die. The result, at least in OpenMP set properly, is much higher memory bandwidth by binding processes as far away from each other as possible.</p><p>Another critical performance vector for agentic AI is Python. The agentic world is expanding, but much of the foundation of the AI software ecosystem is built on Python. Common agentic frameworks like CrewAI and LangGraph are written in Python, and although we’re seeing more frameworks built in languages like Rust and TypeScript, Python performance is still critical to agentic loops. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:77.54%;"><img id="gFfVmJSkrcQ8AqrWReeLKc" name="Vera Benchmarks - PyBench" alt="Vera Pybench Benchmarks" src="https://cdn.mos.cms.futurecdn.net/gFfVmJSkrcQ8AqrWReeLKc-1920-80.png" mos="" align="middle" fullscreen="" width="1999" height="1550" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p>And, unsurprisingly, it’s an area where Vera is also strong, beating out CPUs that are much more powerful on paper. Popular agentic harnesses are built on Python, but so are many of the applications that also run on Python, including foundational AI libraries like PyTorch and TensorFlow. </p><p>Nvidia’s focus on agentic AI makes sense given the times we’re living in, but it can miss the forest for the trees. Oracle bringing Vera standalone instances online shifts the focus onto the broader implications of Vera’s entrance into the data center. There are some aspects of performance that characterize agentic AI workloads, but many agent workloads are regular data center workloads in the first place, just under high concurrency. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1999px;"><p class="vanilla-image-block" style="padding-top:77.14%;"><img id="SFBUqg66rMWFMDAdpLWWFc" name="Vera Benchmarks - SQLite" alt="Vera SQLite benchmarks" src="https://cdn.mos.cms.futurecdn.net/SFBUqg66rMWFMDAdpLWWFc-1920-80.png" mos="" align="middle" fullscreen="" width="1999" height="1542" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Tom's Hardware)</span></figcaption></figure><p>For instance, SQLite performance is important for any server querying a large database, regardless of its written by logic, a person, or an AI agent. Nvidia’s second-gen Scalable Coherency Fabric (SCF), which we saw at play in Stream, shows up in a real workload here, with these database queries largely bound by data movement across the system. </p><p>And that is what makes Vera so interesting in the context of Oracle’s announcement. It’s a CPU built for agentic AI, says Nvidia, but the crosstalk with general data center workloads is strong. It’s not just a CPU to use because you need access to Nvidia’s accelerators. Depending on the workload, it may even be a better choice than the x86 guard. </p><p>Ultimately, we have to see how customers of these hyperscalers use (or don’t use) Vera in production. It’s demonstrated its usefulness in agentic AI loops, but it’s also showing early promise in certain general-purpose workloads. Oracle’s standalone deployment is a sign that there’s some demand for Vera that isn't exclusively tied at the hip of a Rubin GPU. </p><p>Oracle is approaching cautiously, however, as any hyperscaler should be when bringing up a multi-gigawatt deployment. The company said it will also offer Nvidia’s HGX NVL8, bringing eight Rubin GPUs together in a cluster, connected to a dual, 64-core Intel Xeon 6 host with 4TB of DDR5 memory. </p><p>Our testing of Vera shared in this article is part of a larger benchmarking session we had with the chip, and we'll share the technical details of the test setup in our forthcoming review. We’ll be publishing our full results shortly on <em>Tom’s Hardware. </em></p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Nvidia is bankrolling AI accelerator supply chain buildout to reserve production capacity ]]></title>
                                                                                                <dc:content><![CDATA[ <p>On October 21, during GTC Berlin, Nvidia <a href="https://nvidianews.nvidia.com/news/nvidia-and-supply-chain-partners-expand-capacity-to-accelerate-ai-infrastructure-buildout" target="_blank">announced</a> several massive, multi-billion-dollar deals with supply chain partners across substrates, advanced packaging, semiconductor testing, and optical engines to expand production capacity. Aimed at overcoming severe bottlenecks restricting the current and future AI infrastructure buildouts, the deals feature commitments and production roadmaps stretching through 2028, securing capacity for Nvidia and ensuring resiliency and redundancy in its AI accelerator supply chain. “We are in the midst of the largest infrastructure buildout in human history. Demand is extraordinary. Capacity is the constraint,” said Nvidia CEO Jensen Huang. “Our supply chain partners are the world’s best manufacturers. I’m grateful they are expanding capacity around the globe to scale production and help us deliver the AI supercomputing systems powering this new industrial revolution.”</p><p>The deals cover two major AI accelerator supply chain categories: hardware interconnects and OSAT (Outsourced Semiconductor Assembly and Test). Under hardware interconnects, Nvidia is securing capacity for both substrates and co-packaged optics. Nvidia has signed deals with Ibiden, Unimicron, and Samsung Electro-Mechanics — three of the world's five largest substrate manufacturers — to support its production capacity expansion and reserve significant portions of that capacity for the chipmaker. The company says it plans to secure 2.5x dedicated substrate capacity over the next three years. Nvidia has also signed agreements with SPIL (Siliconware Precision Industries Co., Ltd) to expand manufacturing capacity for the optical engines used in <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/co-packaged-optics-cpo-foundry-roadmaps-breaking-down-tsmc-intel-samsung-and-globalfoundries-approach-to-next-generation-scale-up-connectivity">co-packaged optics</a> by 15x by mid-2027.</p><p>On the OSAT side, Nvidia and Amkor are targeting a 300% increase in interposer wafer capacity by the end of 2028 through expanded advanced packaging capacity in South Korea, to support higher production volumes for Nvidia's platform. The company also plans to double its semiconductor testing capacity by 2027, investing $1.1 billion to reserve capacity from KYEC (King Yuan Electronics Co., Ltd.), the world's largest semiconductor testing company, which currently handles nearly all of Nvidia's chip testing.</p><p>Nvidia currently leads the AI chip market by a considerable margin, with its <a href="https://www.tomshardware.com/pc-components/gpus/nvidias-vera-rubin-platform-in-depth-inside-nvidias-most-complex-ai-and-hpc-platform-to-date">extremely powerful accelerators</a> sitting in a vast majority of AI data centers globally.  These deals target major bottlenecks in the AI chip supply chain, suggesting Nvidia doesn't plan to relinquish its position soon. So, why these specific deals? What are the exact details? What's the broader play? And where do Nvidia's competitors and the rest of the AI industry stand?</p><h2 id="uncorking-bottlenecks">Uncorking bottlenecks </h2><p>The substrate base, a specialized intermediary layer in every accelerator package that connects the tiny silicon chips above it to the board below, is one of the most critical bottlenecks in the AI chip supply chain. As we detailed in our <a href="https://www.tomshardware.com/tech-industry/semiconductors/the-state-of-abf-substrates-in-data-center-silicon-in-2026-solving-the-supply-crunch-and-material-wall-beneath-every-ai-accelerator" target="_blank">comprehensive substrates roadmap</a>, this indispensable physical layer, manufactured by only a handful of companies, is expanding in both the X-Y and Z axes as chip makers <a href="https://www.tomshardware.com/tech-industry/semiconductors/ai-chip-design-is-pushing-2-5d-packaging-to-its-limits" target="_blank">package an increasing number of components onto one board</a>. Nvidia's Blackwell generation places two reticle-sized GPU dies and eight HBM3E stacks on a single package. Its Rubin and Rubin Ultra architectures mount even more components. As a result, each substrate now consumes much more material and production capacity, even as demand for individual substrates soars.</p><p>Fubon Research <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/google-could-build-more-ai-accelerators-than-nvidia-sells-in-2028-analyst-claims-could-push-the-company-to-use-intel-foundry-to-meet-its-goals" target="_blank">estimates</a> that Nvidia would supply around 12.4 million data-center AI GPUs in 2028, up from a projected 8.2 million in 2026. Each one would feature a substrate layer. Nvidia is staying ahead of the curve, with a planned 2.5x capacity expansion over the next three years, led by Ibiden. Earlier this year, the company signed a deal to provide a multibillion-dollar capacity reservation advance to support Ibiden's Ono Plant in Japan, targeting 3x the plant’s current individual capacity.</p><p>To reach its 2.5x target, Nvidia has also brought on Samsung Electro-Mechanics, Samsung's substrate-manufacturing subsidiary, as a supplier. Beyond the immediate capacity from Samsung Electro-Mechanics’ existing operations, the deal will support facility upgrades and new plants in Busan and Sejong. Nvidia is projected to account for a significant portion of the substrate manufacturer’s output by 2028. To build even more redundancy, Nvidia also signed a capacity reservation agreement with Unimicron, supporting the latter's substrate manufacturing expansion in Taiwan. The expansion will reach full capacity by mid-2028 but is expected to start providing capacity by Q2 2027.</p><p>Next on Nvidia's to-expand-capacity list announced on October 21 is co-packaged optics. As next-generation speeds surpass 224 Gbps per lane, traditional copper traces on the board consume considerable power to move data even a few inches from the processor to the edge of the board. To address this, the industry is moving the optical engine into the accelerator package itself via <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/co-packaged-optics-cpo-foundry-roadmaps-breaking-down-tsmc-intel-samsung-and-globalfoundries-approach-to-next-generation-scale-up-connectivity" target="_blank">co-packaged optics</a><strong>. </strong><a href="https://www.tomshardware.com/tech-industry/photonics/ai-data-center-optical-interconnect-market-to-hit-usd144-billion-by-2030-an-over-ten-fold-increase-from-2024-figures-according-to-new-projections-silicon-photonics-expected-to-account-for-nearly-two-thirds-of-revenue-driven-by-co-packaged-optics" target="_blank">Nvidia had long detailed its plans for silicon photonics and co-packaged optics</a> as the standard for future AI accelerator packages. The company has signed a deal with Taiwanese optical engine manufacturer SPIL to expand production capacity by 1,400% by mid-2027, with plans to double again by the end of 2028.</p><h2 id="a-4x-increase-in-packaging-double-the-testing-and-1-1-billion-specifically-for-u-s-facilities">A 4x increase in packaging, double the testing, and $1.1 billion specifically for U.S. facilities</h2><p>Interestingly, the world's biggest semiconductor company, Nvidia, doesn't own a single Fab. After TSMC manufactures the logic chip and binds it with memory chips on a silicon interposer, final-assembly companies such as Amkor package that integrated module with the substrate and protective casing into the final accelerator package. These final steps are increasingly incorporating co-packaged photonics chips. Nvidia’s deal with Amkor aims to expand the latter's interposer wafer capacity by more than 300% by the end of 2028, through advanced packaging capacity expansions in South Korea. Amkor confirmed the deal is worth $1.3 billion, bringing Nvidia's total investment in Amkor across its <a href="https://www.tomshardware.com/tech-industry/semiconductors/amkor-breaks-ground-on-arizona-advanced-packaging-campus" target="_blank">Arizona and South Korea facilities</a> to $3 billion.</p><p>After final assembly, finished Nvidia AI accelerator packages are shipped to KYEC for extreme testing. KYEC currently handles nearly all of Nvidia's chip testing. It therefore comes as no surprise that Nvidia expects semiconductor testing capacity to double from calendar 2026 to 2027 through buildouts across KYEC’s facilities in the U.S., Taiwan, and Singapore. On October 13, Nvidia signed a deal to pay KYEC $1.1 billion to reserve testing capacity at KYEC's new $1.4 billion U.S. facility, where high-volume production is slated to begin in April 2027. The funds will build domestic testing capacity and bolster the domestic supply chain. At the same time, KYEC is aggressively expanding operations in Singapore, where it expects high-volume production to begin in February 2027.</p><h2 id="the-bigger-picture">The bigger picture</h2><p>Nvidia has been aggressively securing production capacity and long-term supply commitments to meet anticipated demand for its next-generation AI accelerators. According to its August 2026 financial disclosures, the company had <a href="https://www.tomshardware.com/tech-industry/big-tech/nvidia-revenue-tops-usd96-billion-as-memory-commitments-soar-to-usd160-billion-ceo-jensen-huang-says-ai-has-reached-its-inflection-point" target="_blank">accumulated $279 billion in outstanding supply and manufacturing-capacity commitments</a> by July 2026. The company has since signed several additional deals. These commitments cover future component purchases and manufacturing capacity reservations, particularly for memory and chip production. Combined with its more recent agreements, Nvidia is securing nearly every stage of its AI accelerator supply chain, from semiconductor fabrication and HBM memory to substrates, advanced packaging, optical interconnects, final assembly, and testing.</p><p>Of course, the rest of the industry isn't standing by. On May 21, 2026, AMD announced over $10 billion in investments across Taiwan's semiconductor ecosystem, specifically targeting advanced packaging and manufacturing partnerships. On October 6, CEO Lisa Su confirmed that AMD intends to increase those investments further. The company is working with TSMC, packaging partners, substrate manufacturers, and memory suppliers to increase production substantially in 2027. Google and Amazon are also aggressively expanding their <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/googles-tpus-for-ai-workloads-now-run-in-at-metas-data-centers-tpus-sold-to-external-customers-for-the-first-time-threaten-nvidias-market-dominance">custom TPU </a>and <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/amazon-launches-trainium3-ai-accelerator-competing-directly-against-blackwell-ultra-in-fp8-performance-new-trn3-gen2-ultraserver-takes-vertical-scaling-notes-from-nvidias-playbook">Trainium</a> accelerator programs.</p><p>Intel, meanwhile, is positioning its foundry and advanced packaging operations as alternatives to TSMC's increasingly constrained production capacity. The company is expanding its <a href="https://www.tomshardware.com/tech-industry/semiconductors/intels-emib-packaging-gains-traction-as-chip-designers-look-to-skirt-tsmcs-cowos-constraints-googles-reported-decision-for-9th-gen-tpus-highlights-intels-attractive-alternative" target="_blank">EMIB-T packaging technology</a>, with reports indicating that Google has already reserved Intel capacity to package more than three million TPUs in 2028. Intel is also advancing its <a href="https://www.tomshardware.com/tech-industry/semiconductors/intels-fab-roadmap-examined" target="_blank">18A and 14A manufacturing processes</a> as it attempts to attract major external customers.</p><p>Samsung is expanding its semiconductor operations as well. The Korean company, which held approximately 5.9% of the global foundry market by revenue in Q2 2026, is ramping up its 2nm manufacturing technology and preparing its Taylor, Texas, facility for volume production in 2027. As detailed in our comprehensive Samsung foundry roadmap, the company is investing heavily in its manufacturing operations as it attempts to close the considerable gap with TSMC.</p><p>Perhaps the most ambitious player is new entrant Elon Musk's Terafab, a joint semiconductor manufacturing project involving Tesla, SpaceX, and xAI. The project aims to eventually manufacture the equivalent of one terawatt of AI computing capacity annually, an extraordinarily ambitious target given the industry's current scale. Intel joined the project in April, and Musk later confirmed plans to use the company's 14A fabrication technology. The project aims to deliver<a href="https://www.tomshardware.com/tech-industry/semiconductors/elon-musk-says-asml-and-other-critical-semiconductor-equipment-suppliers-will-expedite-shipments-to-terafab-facility-targets-volume-chip-production-by-q3-q4-2028-in-highly-aspirational-timeline" target="_blank"> its first experimental chips in 2027 and begin high-volume production by the third or fourth quarter of 2028</a>.</p><p>Nvidia and other semiconductor firms are continuously fighting against emerging bottlenecks, but compute doesn't come online overnight, and fabs aren't built in a day. So, while almost every inch of the semiconductor supply chain remains constrained, it remains likely that big-name players will continue to invest heavily to ensure that products continue to flow into neoclouds and data centers alike. </p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/semiconductors/nvidia-is-bankrolling-ai-accelerator-supply-chain-buildout-to-reserve-production-capacity-announces-billion-dollar-deals-with-six-major-suppliers-across-substrates-advanced-packaging-semiconductor-testing-and-optical-engines</link>
                                                                            <description>
                            <![CDATA[ Nvidia is committing billions to expand substrate, co-packaged optics, advanced packaging, and chip-testing capacity through 2028, reserving capacity at nearly every stage of its AI accelerator supply chain. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">KcNNFsQt7R9eKUfLT8Dkbd</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/zkoCNa9uUdXnRAcBKxyasn-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Tue, 27 Oct 2026 11:00:00 +0000</pubDate>                                                                                                                                <updated>Tue, 27 Oct 2026 18:48:50 +0000</updated>
                                                                                                                                            <category><![CDATA[Semiconductors]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                    <category><![CDATA[Manufacturing]]></category>
                                                                                                                    <dc:creator><![CDATA[ Etiido Uko ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/BBrMt7jWtSo2Dc3iKoroyD-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Etiido Uko is a mechanical engineer and senior technical writer with over nine years of experience in documentation and reporting. He is deeply passionate about all things engineering and technology, and is an expert in gadgets, manufacturing, robotics, automotive, and aerospace. His work spans content creation for industry leaders across multiple sectors, including Autodesk, Siemens, Xometry, Telus, and Coca-Cola. When he is not writing or keeping up with the latest innovations, you can find him exploring lands unknown. Check out more of his work at etiidowrites.com.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/zkoCNa9uUdXnRAcBKxyasn-1920-80.jpg">
                                                            <media:credit><![CDATA[Nvidia]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Nvidia supply chain partners]]></media:description>                                                            <media:text><![CDATA[Nvidia supply chain partners]]></media:text>
                                <media:title type="plain"><![CDATA[Nvidia supply chain partners]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/zkoCNa9uUdXnRAcBKxyasn-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>On October 21, during GTC Berlin, Nvidia <a href="https://nvidianews.nvidia.com/news/nvidia-and-supply-chain-partners-expand-capacity-to-accelerate-ai-infrastructure-buildout" target="_blank">announced</a> several massive, multi-billion-dollar deals with supply chain partners across substrates, advanced packaging, semiconductor testing, and optical engines to expand production capacity. Aimed at overcoming severe bottlenecks restricting the current and future AI infrastructure buildouts, the deals feature commitments and production roadmaps stretching through 2028, securing capacity for Nvidia and ensuring resiliency and redundancy in its AI accelerator supply chain. “We are in the midst of the largest infrastructure buildout in human history. Demand is extraordinary. Capacity is the constraint,” said Nvidia CEO Jensen Huang. “Our supply chain partners are the world’s best manufacturers. I’m grateful they are expanding capacity around the globe to scale production and help us deliver the AI supercomputing systems powering this new industrial revolution.”</p><p>The deals cover two major AI accelerator supply chain categories: hardware interconnects and OSAT (Outsourced Semiconductor Assembly and Test). Under hardware interconnects, Nvidia is securing capacity for both substrates and co-packaged optics. Nvidia has signed deals with Ibiden, Unimicron, and Samsung Electro-Mechanics — three of the world's five largest substrate manufacturers — to support its production capacity expansion and reserve significant portions of that capacity for the chipmaker. The company says it plans to secure 2.5x dedicated substrate capacity over the next three years. Nvidia has also signed agreements with SPIL (Siliconware Precision Industries Co., Ltd) to expand manufacturing capacity for the optical engines used in <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/co-packaged-optics-cpo-foundry-roadmaps-breaking-down-tsmc-intel-samsung-and-globalfoundries-approach-to-next-generation-scale-up-connectivity">co-packaged optics</a> by 15x by mid-2027.</p><p>On the OSAT side, Nvidia and Amkor are targeting a 300% increase in interposer wafer capacity by the end of 2028 through expanded advanced packaging capacity in South Korea, to support higher production volumes for Nvidia's platform. The company also plans to double its semiconductor testing capacity by 2027, investing $1.1 billion to reserve capacity from KYEC (King Yuan Electronics Co., Ltd.), the world's largest semiconductor testing company, which currently handles nearly all of Nvidia's chip testing.</p><p>Nvidia currently leads the AI chip market by a considerable margin, with its <a href="https://www.tomshardware.com/pc-components/gpus/nvidias-vera-rubin-platform-in-depth-inside-nvidias-most-complex-ai-and-hpc-platform-to-date">extremely powerful accelerators</a> sitting in a vast majority of AI data centers globally.  These deals target major bottlenecks in the AI chip supply chain, suggesting Nvidia doesn't plan to relinquish its position soon. So, why these specific deals? What are the exact details? What's the broader play? And where do Nvidia's competitors and the rest of the AI industry stand?</p><h2 id="uncorking-bottlenecks">Uncorking bottlenecks </h2><p>The substrate base, a specialized intermediary layer in every accelerator package that connects the tiny silicon chips above it to the board below, is one of the most critical bottlenecks in the AI chip supply chain. As we detailed in our <a href="https://www.tomshardware.com/tech-industry/semiconductors/the-state-of-abf-substrates-in-data-center-silicon-in-2026-solving-the-supply-crunch-and-material-wall-beneath-every-ai-accelerator" target="_blank">comprehensive substrates roadmap</a>, this indispensable physical layer, manufactured by only a handful of companies, is expanding in both the X-Y and Z axes as chip makers <a href="https://www.tomshardware.com/tech-industry/semiconductors/ai-chip-design-is-pushing-2-5d-packaging-to-its-limits" target="_blank">package an increasing number of components onto one board</a>. Nvidia's Blackwell generation places two reticle-sized GPU dies and eight HBM3E stacks on a single package. Its Rubin and Rubin Ultra architectures mount even more components. As a result, each substrate now consumes much more material and production capacity, even as demand for individual substrates soars.</p><p>Fubon Research <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/google-could-build-more-ai-accelerators-than-nvidia-sells-in-2028-analyst-claims-could-push-the-company-to-use-intel-foundry-to-meet-its-goals" target="_blank">estimates</a> that Nvidia would supply around 12.4 million data-center AI GPUs in 2028, up from a projected 8.2 million in 2026. Each one would feature a substrate layer. Nvidia is staying ahead of the curve, with a planned 2.5x capacity expansion over the next three years, led by Ibiden. Earlier this year, the company signed a deal to provide a multibillion-dollar capacity reservation advance to support Ibiden's Ono Plant in Japan, targeting 3x the plant’s current individual capacity.</p><p>To reach its 2.5x target, Nvidia has also brought on Samsung Electro-Mechanics, Samsung's substrate-manufacturing subsidiary, as a supplier. Beyond the immediate capacity from Samsung Electro-Mechanics’ existing operations, the deal will support facility upgrades and new plants in Busan and Sejong. Nvidia is projected to account for a significant portion of the substrate manufacturer’s output by 2028. To build even more redundancy, Nvidia also signed a capacity reservation agreement with Unimicron, supporting the latter's substrate manufacturing expansion in Taiwan. The expansion will reach full capacity by mid-2028 but is expected to start providing capacity by Q2 2027.</p><p>Next on Nvidia's to-expand-capacity list announced on October 21 is co-packaged optics. As next-generation speeds surpass 224 Gbps per lane, traditional copper traces on the board consume considerable power to move data even a few inches from the processor to the edge of the board. To address this, the industry is moving the optical engine into the accelerator package itself via <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/co-packaged-optics-cpo-foundry-roadmaps-breaking-down-tsmc-intel-samsung-and-globalfoundries-approach-to-next-generation-scale-up-connectivity" target="_blank">co-packaged optics</a><strong>. </strong><a href="https://www.tomshardware.com/tech-industry/photonics/ai-data-center-optical-interconnect-market-to-hit-usd144-billion-by-2030-an-over-ten-fold-increase-from-2024-figures-according-to-new-projections-silicon-photonics-expected-to-account-for-nearly-two-thirds-of-revenue-driven-by-co-packaged-optics" target="_blank">Nvidia had long detailed its plans for silicon photonics and co-packaged optics</a> as the standard for future AI accelerator packages. The company has signed a deal with Taiwanese optical engine manufacturer SPIL to expand production capacity by 1,400% by mid-2027, with plans to double again by the end of 2028.</p><h2 id="a-4x-increase-in-packaging-double-the-testing-and-1-1-billion-specifically-for-u-s-facilities">A 4x increase in packaging, double the testing, and $1.1 billion specifically for U.S. facilities</h2><p>Interestingly, the world's biggest semiconductor company, Nvidia, doesn't own a single Fab. After TSMC manufactures the logic chip and binds it with memory chips on a silicon interposer, final-assembly companies such as Amkor package that integrated module with the substrate and protective casing into the final accelerator package. These final steps are increasingly incorporating co-packaged photonics chips. Nvidia’s deal with Amkor aims to expand the latter's interposer wafer capacity by more than 300% by the end of 2028, through advanced packaging capacity expansions in South Korea. Amkor confirmed the deal is worth $1.3 billion, bringing Nvidia's total investment in Amkor across its <a href="https://www.tomshardware.com/tech-industry/semiconductors/amkor-breaks-ground-on-arizona-advanced-packaging-campus" target="_blank">Arizona and South Korea facilities</a> to $3 billion.</p><p>After final assembly, finished Nvidia AI accelerator packages are shipped to KYEC for extreme testing. KYEC currently handles nearly all of Nvidia's chip testing. It therefore comes as no surprise that Nvidia expects semiconductor testing capacity to double from calendar 2026 to 2027 through buildouts across KYEC’s facilities in the U.S., Taiwan, and Singapore. On October 13, Nvidia signed a deal to pay KYEC $1.1 billion to reserve testing capacity at KYEC's new $1.4 billion U.S. facility, where high-volume production is slated to begin in April 2027. The funds will build domestic testing capacity and bolster the domestic supply chain. At the same time, KYEC is aggressively expanding operations in Singapore, where it expects high-volume production to begin in February 2027.</p><h2 id="the-bigger-picture">The bigger picture</h2><p>Nvidia has been aggressively securing production capacity and long-term supply commitments to meet anticipated demand for its next-generation AI accelerators. According to its August 2026 financial disclosures, the company had <a href="https://www.tomshardware.com/tech-industry/big-tech/nvidia-revenue-tops-usd96-billion-as-memory-commitments-soar-to-usd160-billion-ceo-jensen-huang-says-ai-has-reached-its-inflection-point" target="_blank">accumulated $279 billion in outstanding supply and manufacturing-capacity commitments</a> by July 2026. The company has since signed several additional deals. These commitments cover future component purchases and manufacturing capacity reservations, particularly for memory and chip production. Combined with its more recent agreements, Nvidia is securing nearly every stage of its AI accelerator supply chain, from semiconductor fabrication and HBM memory to substrates, advanced packaging, optical interconnects, final assembly, and testing.</p><p>Of course, the rest of the industry isn't standing by. On May 21, 2026, AMD announced over $10 billion in investments across Taiwan's semiconductor ecosystem, specifically targeting advanced packaging and manufacturing partnerships. On October 6, CEO Lisa Su confirmed that AMD intends to increase those investments further. The company is working with TSMC, packaging partners, substrate manufacturers, and memory suppliers to increase production substantially in 2027. Google and Amazon are also aggressively expanding their <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/googles-tpus-for-ai-workloads-now-run-in-at-metas-data-centers-tpus-sold-to-external-customers-for-the-first-time-threaten-nvidias-market-dominance">custom TPU </a>and <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/amazon-launches-trainium3-ai-accelerator-competing-directly-against-blackwell-ultra-in-fp8-performance-new-trn3-gen2-ultraserver-takes-vertical-scaling-notes-from-nvidias-playbook">Trainium</a> accelerator programs.</p><p>Intel, meanwhile, is positioning its foundry and advanced packaging operations as alternatives to TSMC's increasingly constrained production capacity. The company is expanding its <a href="https://www.tomshardware.com/tech-industry/semiconductors/intels-emib-packaging-gains-traction-as-chip-designers-look-to-skirt-tsmcs-cowos-constraints-googles-reported-decision-for-9th-gen-tpus-highlights-intels-attractive-alternative" target="_blank">EMIB-T packaging technology</a>, with reports indicating that Google has already reserved Intel capacity to package more than three million TPUs in 2028. Intel is also advancing its <a href="https://www.tomshardware.com/tech-industry/semiconductors/intels-fab-roadmap-examined" target="_blank">18A and 14A manufacturing processes</a> as it attempts to attract major external customers.</p><p>Samsung is expanding its semiconductor operations as well. The Korean company, which held approximately 5.9% of the global foundry market by revenue in Q2 2026, is ramping up its 2nm manufacturing technology and preparing its Taylor, Texas, facility for volume production in 2027. As detailed in our comprehensive Samsung foundry roadmap, the company is investing heavily in its manufacturing operations as it attempts to close the considerable gap with TSMC.</p><p>Perhaps the most ambitious player is new entrant Elon Musk's Terafab, a joint semiconductor manufacturing project involving Tesla, SpaceX, and xAI. The project aims to eventually manufacture the equivalent of one terawatt of AI computing capacity annually, an extraordinarily ambitious target given the industry's current scale. Intel joined the project in April, and Musk later confirmed plans to use the company's 14A fabrication technology. The project aims to deliver<a href="https://www.tomshardware.com/tech-industry/semiconductors/elon-musk-says-asml-and-other-critical-semiconductor-equipment-suppliers-will-expedite-shipments-to-terafab-facility-targets-volume-chip-production-by-q3-q4-2028-in-highly-aspirational-timeline" target="_blank"> its first experimental chips in 2027 and begin high-volume production by the third or fourth quarter of 2028</a>.</p><p>Nvidia and other semiconductor firms are continuously fighting against emerging bottlenecks, but compute doesn't come online overnight, and fabs aren't built in a day. So, while almost every inch of the semiconductor supply chain remains constrained, it remains likely that big-name players will continue to invest heavily to ensure that products continue to flow into neoclouds and data centers alike. </p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Cadence reveals Maestro, fine-tuned AI models built to run specialized chip design agents, with early access coming for 'select customers' in Q4 2026  ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Cadence, one of the largest electronic design automation (EDA) companies, has unveiled Maestro, its new portfolio of domain-specialized AI models tailored for its chip design agents, in <a href="https://newsroom.cadence.com">a press release</a> sent just ahead of earnings. In initial results compared with general-purpose AI models, Cadence claims up to a 30% improvement in the quality of results and that Maestro effectively uses fewer tokens for the same task, with up to 5x the token efficiency. The first models “are available to select early access customers” in Q4 2026, and customers may run them on Cadence’s cloud, public clouds, or customer-controlled on-premises environments and post-train them on their own “proprietary data, designs, and workflows.” To learn more about Maestro, we spoke with two Cadence executives: Rob Knoth, senior group director of strategy and new ventures, and Paul Cunningham, senior vice president and general manager.</p><p>This approach is distinct from <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-and-synopsys-partner-to-build-gpt-synopsys-for-autonomous-chip-design-specialized-ai-model-will-operate-eda-tools-allowing-engineers-to-deliver-more-sophisticated-chips-faster">Synopsys’ own specialized model, GPT-Synopsys</a>, which will run on OpenAI-hosted infrastructure, without using customer data to train the model. The models of both EDA giants, announced less than a month apart, now split on where a customer’s design data goes. Where they might not differ is why: Paul Cunningham, senior vice president and general manager at Cadence, told <em>Tom’s Hardware Premium</em> that agent costs have “really started to explode,” adding, “And I think the world realized so.” That makes it a natural step, he said, to move more of the agents’ iterating “into the weights of the models themselves.”</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="RsAc8W5GdnGY4vrVtXnKEc" name="cadence-maestro-three-pillars" alt="Cadence diagram of Maestro’s three parts: domain-specialized models, routing between foundation and Maestro models, and customer-specialized models post-trained on customer data." src="https://cdn.mos.cms.futurecdn.net/RsAc8W5GdnGY4vrVtXnKEc-1920-80.png" mos="" align="middle" fullscreen="" width="2560" height="1440" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Cadence)</span></figcaption></figure><p>Maestro takes this step through Cadence’s <a href="https://www.tomshardware.com/tech-industry/semiconductors/cadence-embeds-ai-across-its-eda-portfolio">AI Super Agents</a>, which “interpret design intent, decompose tasks” and “orchestrate the Maestro portfolio with other leading frontier and open-weight models,” evaluating trade-offs to guide decisions with domain expertise. It also uses reinforcement learning (RL), where Cadence’s “deterministic solvers evaluate outputs against engineering objectives, constraints and ground truth,” and, importantly, the results feed back into the models for improvement.</p><p>The models include <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidias-nemoclaw-coalition-brings-eight-ai-labs-together-to-build-open-frontier-models">Nvidia’s Nemotron</a>, and customers can run the customer-data training step themselves with Cadence’s fine-tuning harness, potentially “on their own GPUs, on their own network,” Cunningham said. He added that Nemotron, specifically, was “giving really solid results and actually improving very, very fast.”</p><p>One constant that keeps popping up with agentic EDA is the need to nail down what the numbers actually measure. In chip design, quality of results usually refers to the design’s power, performance, and area, though Cadence didn’t specify what its 30% figure measures. Both it and the 5x token efficiency figure are measured against “a general-purpose foundation model,” which Cadence told us is the open-source model Maestro started from before its training. The company wouldn’t name the foundational AI model that Maestro is derived from beyond mentioning that it is “one of the leading open source models that’s very competitive with commercial frontier models.”</p><p>Cadence credits the token savings to balancing its EDA and system design automation (SDA) “tool calls with LLM invocations.” As an analogy, Cunningham said, “You don’t need to drive to Safeway in the jet engine,” and you want to use “the right model at the right time for the right task.”</p><p>Runtime does matter, too: as Rob Knoth, senior group director of strategy and new ventures at Cadence, told <em>Tom’s Hardware Premium</em> in September that a tool call could take anywhere from hours to days to return a value, while the AI industry’s training loops are “used to minutes.” Efficiency could also lower costs, since Maestro use counts toward Cadence’s agent pricing, a subscription with a consumption allowance and charges for use beyond it, Cunningham said. He didn’t say whether the token savings cut a customer’s bill.</p><p>Being flexible for the customer is a key element of any solution, and for Cadence’s competitor Synopsys, <a href="https://www.tomshardware.com/tech-industry/semiconductors/synopsys-debuts-autopilot-platform-for-developing-chips-autonomously-using-ai-new-agentengineer-platform-is-poised-for-general-availability-by-the-end-of-2026">this lies in Autopilot</a>, which Synopsys says allows customers to “choose among commercial, open-source, and fine-tuned language models” across “Synopsys-managed and customer-managed deployments,” with cloud or on-premises infrastructure. For agent security, such as sandboxing, Cadence says it’s working with Nvidia on an integration with Nvidia’s OpenShell Reference Architecture, which Nvidia launched last week, naming both Cadence and Synopsys. These options can help protect customer intellectual property.</p><div ><table><caption>Cadence Maestro vs. GPT-Synopsys</caption><thead><tr><th class="firstcol empty" ></th><th  ><p>Cadence Maestro</p></th><th  ><p>GPT-Synopsys</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Announced</p></td><td  ><p>Oct. 26, 2026</p></td><td  ><p>Sept. 30, 2026</p></td></tr><tr><td class="firstcol " ><p>What it is</p></td><td  ><p>A portfolio of domain-specialized AI models for Cadence’s AI Super Agents</p></td><td  ><p>A specialized model that operates Synopsys’ tools directly</p></td></tr><tr><td class="firstcol " ><p>Model partner</p></td><td  ><p>Nvidia (Nemotron, OpenShell)</p></td><td  ><p>OpenAI</p></td></tr><tr><td class="firstcol " ><p>Models</p></td><td  ><p>Post-trained LLMs, trained decision models and agent skills, blended with frontier and open-weight models, including Nemotron</p></td><td  ><p>OpenAI frontier models</p></td></tr><tr><td class="firstcol " ><p>Where it runs</p></td><td  ><p>Cadence’s cloud, public clouds or customer-controlled on-premises environments</p></td><td  ><p>OpenAI-hosted infrastructure</p></td></tr><tr><td class="firstcol " ><p>Customer data</p></td><td  ><p>Customers can post-train it on their own data, designs and workflows, with Cadence’s fine-tuning tools</p></td><td  ><p>Customer data isn’t used to train the model</p></td></tr><tr><td class="firstcol " ><p>Claimed results</p></td><td  ><p>Up to 30% higher quality of results and up to 5x token efficiency vs. a general-purpose foundation model, the unnamed open-source model Maestro started from (initial results; measure not stated)</p></td><td  ><p>None disclosed</p></td></tr><tr><td class="firstcol " ><p>Availability</p></td><td  ><p>First models to select early-access customers in Q4 2026</p></td><td  ><p>No date; early technology engagements underway</p></td></tr><tr><td class="firstcol " ><p>Customers named</p></td><td  ><p>Socionext, Silicon Labs (Nvidia as partner), frontier AI labs</p></td><td  ><p>None</p></td></tr><tr><td class="firstcol " ><p>Pricing</p></td><td  ><p>Subscription with a usage allowance, then charged by use</p></td><td  ><p>Not disclosed; Synopsys and OpenAI share revenue under their deal</p></td></tr></tbody></table></div><p>Cadence has had a busy 12 months with customers and agents alike. During Cadence’s July earnings call, CEO Anirudh Devgan said ChipStack, its design and verification agent, “has more than 20 customer engagements,” and ViraStack, its custom and analog agent, “more than 25 customer engagements.” Socionext said just last week that it was applying InnoStack. Silicon Labs has also credited ChipStack with accelerating debugging, and both companies say they are watching Maestro next.</p><p>Cadence's agentic consolidation began with its acquisition of ChipStack back in November 2025. In June 2026, the company claimed that its agent had reached Level 5 autonomy, which <a href="https://www.tomshardware.com/tech-industry/semiconductors/the-state-of-agentic-ai-in-chip-design-tools-in-2026-cadence-synopsys-and-siemens-all-pitch-autonomous-engineers">we explored in detail</a>. In September, Knoth teased Maestro’s ingredients to <em>Tom's Hardware Premium</em> a month early, two days before the GPT-Synopsys announcement: “reinforcement learning loops, domain-specific post-training of open-weight models, model routing.” Maestro is now sitting under all four super agents, targeting a similar year-end window to Synopsys’ Autopilot; the table below has the full timeline.</p><div ><table><caption>Cadence’s agentic AI roadmap</caption><thead><tr><th class="firstcol " ><p>When</p></th><th  ><p>Product</p></th><th  ><p>Cadence’s claim</p></th><th  ><p>Status and what’s due</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>November 2025</p></td><td  ><p>ChipStack acquisition (agentic verification startup)</p></td><td  ><p>—</p></td><td  ><p>Done; co-founder Kartik Hegde now leads Cadence’s core AI research group</p></td></tr><tr><td class="firstcol " ><p>February 2026</p></td><td  ><p>ChipStack AI Super Agent</p></td><td  ><p>Up to 10x productivity for coding designs and testbenches, test plans, regressions and debug</p></td><td  ><p>In early access since launch</p></td></tr><tr><td class="firstcol " ><p>April 2026</p></td><td  ><p>ViraStack (custom and analog), InnoStack (digital implementation) and AgentStack (head agent)</p></td><td  ><p>—</p></td><td  ><p>Socionext is deploying InnoStack in production; others not stated</p></td></tr><tr><td class="firstcol " ><p>June 2026</p></td><td  ><p>ChipStack at Level 5, with Nvidia’s Nemotron and OpenShell</p></td><td  ><p>Over 40x faster RTL validation cycles in Nvidia’s deployment</p></td><td  ><p>Level 5 and AgentStack due to early-access customers in the second half of 2026</p></td></tr><tr><td class="firstcol " ><p>July 2026</p></td><td  ><p>AuraStack (PCB and advanced packaging)</p></td><td  ><p>Up to 2x faster time to market and 15x higher productivity</p></td><td  ><p>Available on Cadence’s Millennium M2000 supercomputer; Nokia is deploying it (Oct. 21)</p></td></tr><tr><td class="firstcol " ><p>September 2026</p></td><td  ><p>ChipStack RTL Generation Agent</p></td><td  ><p>An average of 24% less area and 18% less power than pure foundation-model code, in early evaluations</p></td><td  ><p>Due to select early-access customers in Q4 2026; early collaboration with Honda R&amp;D</p></td></tr><tr><td class="firstcol " ><p>Oct. 26, 2026</p></td><td  ><p>Maestro AI Model Portfolio</p></td><td  ><p>Up to 30% higher quality of results and up to 5x token efficiency vs. a general-purpose foundation model, in initial results</p></td><td  ><p>First models available to select early-access customers in Q4 2026; about a dozen engagements</p></td></tr></tbody></table></div><p>This rapid pace of progress means that a clear distinction has to be made between early access and general availability. Both Cadence and Synopsys can point to customer engagements, and, at least in the latter’s case for the AgentEngineer portfolio and Autopilot platform, general availability remains a goal by year’s end; Cadence hasn’t given a general availability date for Maestro. For GPT-Synopsys, though, only “early technology engagements are underway,” according to its partners. The platforms and agents remain works in progress.</p><p>For Maestro, Cunningham hopes for “more case studies, more quantifiable benefits” through Q4. The real story, as he aptly put it, is that “we’re going to bring agents and agentic AI natively into our vertical, not that our vertical is going to get embedded into general-purpose agents.” Only customer results on real workflows will test Cadence’s 30% number, but going by what both companies have shown so far, their own early figures are extremely promising and exciting for the EDA industry.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/semiconductors/cadence-reveals-maestro-fine-tuned-ai-models-built-to-run-specialized-chip-design-agents-with-early-access-coming-for-select-customers-in-q4-2026-claims-up-to-30-percent-higher-quality-results-and-up-to-5x-token-efficiency-over-a-general-purpose-ai-model</link>
                                                                            <description>
                            <![CDATA[ Cadence unveiled Maestro, domain-specialized AI models for its chip design agents, with early access for select customers in Q4 2026. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">9vrvg7oWN8NC7kaZpoRBoM</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/P8XPrtLpMc8eqkqoGb66Vo-1920-80.png" type="image/png" length="0"></enclosure>
                                                                        <pubDate>Mon, 26 Oct 2026 13:20:00 +0000</pubDate>                                                                                                                                <updated>Mon, 26 Oct 2026 15:24:46 +0000</updated>
                                                                                                                                            <category><![CDATA[Semiconductors]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                    <category><![CDATA[Manufacturing]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/png" url="https://cdn.mos.cms.futurecdn.net/P8XPrtLpMc8eqkqoGb66Vo-1920-80.png">
                                                            <media:credit><![CDATA[Cadence]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Cadence diagram of AgentStack over five AI Super Agents, with the new Maestro AI models layer, the JedAI data and AI platform, and cloud or on-premises compute beneath.]]></media:description>                                                            <media:text><![CDATA[Cadence diagram of AgentStack over five AI Super Agents, with the new Maestro AI models layer, the JedAI data and AI platform, and cloud or on-premises compute beneath.]]></media:text>
                                <media:title type="plain"><![CDATA[Cadence diagram of AgentStack over five AI Super Agents, with the new Maestro AI models layer, the JedAI data and AI platform, and cloud or on-premises compute beneath.]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/P8XPrtLpMc8eqkqoGb66Vo-1920-80.png" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Cadence, one of the largest electronic design automation (EDA) companies, has unveiled Maestro, its new portfolio of domain-specialized AI models tailored for its chip design agents, in <a href="https://newsroom.cadence.com">a press release</a> sent just ahead of earnings. In initial results compared with general-purpose AI models, Cadence claims up to a 30% improvement in the quality of results and that Maestro effectively uses fewer tokens for the same task, with up to 5x the token efficiency. The first models “are available to select early access customers” in Q4 2026, and customers may run them on Cadence’s cloud, public clouds, or customer-controlled on-premises environments and post-train them on their own “proprietary data, designs, and workflows.” To learn more about Maestro, we spoke with two Cadence executives: Rob Knoth, senior group director of strategy and new ventures, and Paul Cunningham, senior vice president and general manager.</p><p>This approach is distinct from <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-and-synopsys-partner-to-build-gpt-synopsys-for-autonomous-chip-design-specialized-ai-model-will-operate-eda-tools-allowing-engineers-to-deliver-more-sophisticated-chips-faster">Synopsys’ own specialized model, GPT-Synopsys</a>, which will run on OpenAI-hosted infrastructure, without using customer data to train the model. The models of both EDA giants, announced less than a month apart, now split on where a customer’s design data goes. Where they might not differ is why: Paul Cunningham, senior vice president and general manager at Cadence, told <em>Tom’s Hardware Premium</em> that agent costs have “really started to explode,” adding, “And I think the world realized so.” That makes it a natural step, he said, to move more of the agents’ iterating “into the weights of the models themselves.”</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="RsAc8W5GdnGY4vrVtXnKEc" name="cadence-maestro-three-pillars" alt="Cadence diagram of Maestro’s three parts: domain-specialized models, routing between foundation and Maestro models, and customer-specialized models post-trained on customer data." src="https://cdn.mos.cms.futurecdn.net/RsAc8W5GdnGY4vrVtXnKEc-1920-80.png" mos="" align="middle" fullscreen="" width="2560" height="1440" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Cadence)</span></figcaption></figure><p>Maestro takes this step through Cadence’s <a href="https://www.tomshardware.com/tech-industry/semiconductors/cadence-embeds-ai-across-its-eda-portfolio">AI Super Agents</a>, which “interpret design intent, decompose tasks” and “orchestrate the Maestro portfolio with other leading frontier and open-weight models,” evaluating trade-offs to guide decisions with domain expertise. It also uses reinforcement learning (RL), where Cadence’s “deterministic solvers evaluate outputs against engineering objectives, constraints and ground truth,” and, importantly, the results feed back into the models for improvement.</p><p>The models include <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidias-nemoclaw-coalition-brings-eight-ai-labs-together-to-build-open-frontier-models">Nvidia’s Nemotron</a>, and customers can run the customer-data training step themselves with Cadence’s fine-tuning harness, potentially “on their own GPUs, on their own network,” Cunningham said. He added that Nemotron, specifically, was “giving really solid results and actually improving very, very fast.”</p><p>One constant that keeps popping up with agentic EDA is the need to nail down what the numbers actually measure. In chip design, quality of results usually refers to the design’s power, performance, and area, though Cadence didn’t specify what its 30% figure measures. Both it and the 5x token efficiency figure are measured against “a general-purpose foundation model,” which Cadence told us is the open-source model Maestro started from before its training. The company wouldn’t name the foundational AI model that Maestro is derived from beyond mentioning that it is “one of the leading open source models that’s very competitive with commercial frontier models.”</p><p>Cadence credits the token savings to balancing its EDA and system design automation (SDA) “tool calls with LLM invocations.” As an analogy, Cunningham said, “You don’t need to drive to Safeway in the jet engine,” and you want to use “the right model at the right time for the right task.”</p><p>Runtime does matter, too: as Rob Knoth, senior group director of strategy and new ventures at Cadence, told <em>Tom’s Hardware Premium</em> in September that a tool call could take anywhere from hours to days to return a value, while the AI industry’s training loops are “used to minutes.” Efficiency could also lower costs, since Maestro use counts toward Cadence’s agent pricing, a subscription with a consumption allowance and charges for use beyond it, Cunningham said. He didn’t say whether the token savings cut a customer’s bill.</p><p>Being flexible for the customer is a key element of any solution, and for Cadence’s competitor Synopsys, <a href="https://www.tomshardware.com/tech-industry/semiconductors/synopsys-debuts-autopilot-platform-for-developing-chips-autonomously-using-ai-new-agentengineer-platform-is-poised-for-general-availability-by-the-end-of-2026">this lies in Autopilot</a>, which Synopsys says allows customers to “choose among commercial, open-source, and fine-tuned language models” across “Synopsys-managed and customer-managed deployments,” with cloud or on-premises infrastructure. For agent security, such as sandboxing, Cadence says it’s working with Nvidia on an integration with Nvidia’s OpenShell Reference Architecture, which Nvidia launched last week, naming both Cadence and Synopsys. These options can help protect customer intellectual property.</p><div ><table><caption>Cadence Maestro vs. GPT-Synopsys</caption><thead><tr><th class="firstcol empty" ></th><th  ><p>Cadence Maestro</p></th><th  ><p>GPT-Synopsys</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Announced</p></td><td  ><p>Oct. 26, 2026</p></td><td  ><p>Sept. 30, 2026</p></td></tr><tr><td class="firstcol " ><p>What it is</p></td><td  ><p>A portfolio of domain-specialized AI models for Cadence’s AI Super Agents</p></td><td  ><p>A specialized model that operates Synopsys’ tools directly</p></td></tr><tr><td class="firstcol " ><p>Model partner</p></td><td  ><p>Nvidia (Nemotron, OpenShell)</p></td><td  ><p>OpenAI</p></td></tr><tr><td class="firstcol " ><p>Models</p></td><td  ><p>Post-trained LLMs, trained decision models and agent skills, blended with frontier and open-weight models, including Nemotron</p></td><td  ><p>OpenAI frontier models</p></td></tr><tr><td class="firstcol " ><p>Where it runs</p></td><td  ><p>Cadence’s cloud, public clouds or customer-controlled on-premises environments</p></td><td  ><p>OpenAI-hosted infrastructure</p></td></tr><tr><td class="firstcol " ><p>Customer data</p></td><td  ><p>Customers can post-train it on their own data, designs and workflows, with Cadence’s fine-tuning tools</p></td><td  ><p>Customer data isn’t used to train the model</p></td></tr><tr><td class="firstcol " ><p>Claimed results</p></td><td  ><p>Up to 30% higher quality of results and up to 5x token efficiency vs. a general-purpose foundation model, the unnamed open-source model Maestro started from (initial results; measure not stated)</p></td><td  ><p>None disclosed</p></td></tr><tr><td class="firstcol " ><p>Availability</p></td><td  ><p>First models to select early-access customers in Q4 2026</p></td><td  ><p>No date; early technology engagements underway</p></td></tr><tr><td class="firstcol " ><p>Customers named</p></td><td  ><p>Socionext, Silicon Labs (Nvidia as partner), frontier AI labs</p></td><td  ><p>None</p></td></tr><tr><td class="firstcol " ><p>Pricing</p></td><td  ><p>Subscription with a usage allowance, then charged by use</p></td><td  ><p>Not disclosed; Synopsys and OpenAI share revenue under their deal</p></td></tr></tbody></table></div><p>Cadence has had a busy 12 months with customers and agents alike. During Cadence’s July earnings call, CEO Anirudh Devgan said ChipStack, its design and verification agent, “has more than 20 customer engagements,” and ViraStack, its custom and analog agent, “more than 25 customer engagements.” Socionext said just last week that it was applying InnoStack. Silicon Labs has also credited ChipStack with accelerating debugging, and both companies say they are watching Maestro next.</p><p>Cadence's agentic consolidation began with its acquisition of ChipStack back in November 2025. In June 2026, the company claimed that its agent had reached Level 5 autonomy, which <a href="https://www.tomshardware.com/tech-industry/semiconductors/the-state-of-agentic-ai-in-chip-design-tools-in-2026-cadence-synopsys-and-siemens-all-pitch-autonomous-engineers">we explored in detail</a>. In September, Knoth teased Maestro’s ingredients to <em>Tom's Hardware Premium</em> a month early, two days before the GPT-Synopsys announcement: “reinforcement learning loops, domain-specific post-training of open-weight models, model routing.” Maestro is now sitting under all four super agents, targeting a similar year-end window to Synopsys’ Autopilot; the table below has the full timeline.</p><div ><table><caption>Cadence’s agentic AI roadmap</caption><thead><tr><th class="firstcol " ><p>When</p></th><th  ><p>Product</p></th><th  ><p>Cadence’s claim</p></th><th  ><p>Status and what’s due</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>November 2025</p></td><td  ><p>ChipStack acquisition (agentic verification startup)</p></td><td  ><p>—</p></td><td  ><p>Done; co-founder Kartik Hegde now leads Cadence’s core AI research group</p></td></tr><tr><td class="firstcol " ><p>February 2026</p></td><td  ><p>ChipStack AI Super Agent</p></td><td  ><p>Up to 10x productivity for coding designs and testbenches, test plans, regressions and debug</p></td><td  ><p>In early access since launch</p></td></tr><tr><td class="firstcol " ><p>April 2026</p></td><td  ><p>ViraStack (custom and analog), InnoStack (digital implementation) and AgentStack (head agent)</p></td><td  ><p>—</p></td><td  ><p>Socionext is deploying InnoStack in production; others not stated</p></td></tr><tr><td class="firstcol " ><p>June 2026</p></td><td  ><p>ChipStack at Level 5, with Nvidia’s Nemotron and OpenShell</p></td><td  ><p>Over 40x faster RTL validation cycles in Nvidia’s deployment</p></td><td  ><p>Level 5 and AgentStack due to early-access customers in the second half of 2026</p></td></tr><tr><td class="firstcol " ><p>July 2026</p></td><td  ><p>AuraStack (PCB and advanced packaging)</p></td><td  ><p>Up to 2x faster time to market and 15x higher productivity</p></td><td  ><p>Available on Cadence’s Millennium M2000 supercomputer; Nokia is deploying it (Oct. 21)</p></td></tr><tr><td class="firstcol " ><p>September 2026</p></td><td  ><p>ChipStack RTL Generation Agent</p></td><td  ><p>An average of 24% less area and 18% less power than pure foundation-model code, in early evaluations</p></td><td  ><p>Due to select early-access customers in Q4 2026; early collaboration with Honda R&amp;D</p></td></tr><tr><td class="firstcol " ><p>Oct. 26, 2026</p></td><td  ><p>Maestro AI Model Portfolio</p></td><td  ><p>Up to 30% higher quality of results and up to 5x token efficiency vs. a general-purpose foundation model, in initial results</p></td><td  ><p>First models available to select early-access customers in Q4 2026; about a dozen engagements</p></td></tr></tbody></table></div><p>This rapid pace of progress means that a clear distinction has to be made between early access and general availability. Both Cadence and Synopsys can point to customer engagements, and, at least in the latter’s case for the AgentEngineer portfolio and Autopilot platform, general availability remains a goal by year’s end; Cadence hasn’t given a general availability date for Maestro. For GPT-Synopsys, though, only “early technology engagements are underway,” according to its partners. The platforms and agents remain works in progress.</p><p>For Maestro, Cunningham hopes for “more case studies, more quantifiable benefits” through Q4. The real story, as he aptly put it, is that “we’re going to bring agents and agentic AI natively into our vertical, not that our vertical is going to get embedded into general-purpose agents.” Only customer results on real workflows will test Cadence’s 30% number, but going by what both companies have shown so far, their own early figures are extremely promising and exciting for the EDA industry.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Deepseek's latest AI model is smarter than ever, but its low running costs come at a price ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Deepseek just upgraded its V4.1 Flash open-weight model to revision 1017. What looks like a mere point revision may actually be a significant step forward in model efficiency. On the <a href="https://api-docs.deepseek.com/news/news261017" target="_blank">product page</a>, Deepseek uses the slogan "wish-driven programming made real," a likely reference to the colloquial "vibe coding," implying the model is a good fit for open-ended tasks with broad descriptions. However, while the model is indeed cheap and scores well, it can be hallucination-prone.</p><p>LLMs are roughly divided into two categories: closed commercial efforts from frontier labs, whose models are expensive to run but deliver excellent results, and Chinese open-weight competitors that broadly exchange reliability for a <em>much</em> cheaper running cost, particularly for long-running agentic tasks. </p><p>Deepseek V4.1 Flash (DSF) is a Mixture-of-Experts (MoE) model with 552B total parameters, 16B of which are active during runs. The model is multimodal, accepts both text and images as inputs, and its context window length measures the now-standard 1 million tokens.</p><p>According to Deepseek's internal benchmarks, the model slots in neatly between GPT-5.6 Sol and Claude Opus 5, even beating Anthropic's LLM in task tests. That's already a strong showing, but this latest revision of DSF shines in the performance-per-task base and token efficiency. The upgrade from V4 Flash to V4.1 Flash significantly slashed hardware requirements, with V4.1 using only one-quarter of the HBM and one-eighth of SSD storage for its KV cache compared to its predecessor, spending only 890 bytes per token entry.</p><p>The company illustrates the point with a cost-per-task graph based on <a href="https://deepswe.datacurve.ai/" target="_blank">DeepSWE's dataset</a> from September 2, adding its own data points for both DSF and DSF-1017. Both variants cluster in the desirable top-left area, and even low-effort runs of DSF-1017 offer a task completion rate in the ballpark of "max" settings for pricey competitors. It's worth noting that DeepSWE is only one of many benchmarks, but this is nonetheless an impressive result.</p><figure class="van-image-figure  extended-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:86.17%;"><img id="QJnQJ99DDimGvHPGhSi6oB" name="DeepSWE - Deepseek V4.1 Flash - 1017" alt="DeepSWE - Deepseek V4.1 Flash - 1017" src="https://cdn.mos.cms.futurecdn.net/QJnQJ99DDimGvHPGhSi6oB-1920-80.png" mos="" align="middle" fullscreen="1" width="2560" height="2206" attribution="" endorsement="" class="extended expandable" /><a href='https://cdn.mos.cms.futurecdn.net/QJnQJ99DDimGvHPGhSi6oB-1920-80.png' target='_blank' class='expand-button icon-expand-image icon' ></a></p></div></div><figcaption itemprop="caption description" class=" extended-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Deepseek)</span></figcaption></figure><p>Even at Deepseek's <a href="https://api-docs.deepseek.com/quick_start/pricing/" target="_blank">own pricing</a> (remember, this is an open-weight model you can run anywhere), the price at off-peak times per million tokens is just $0.15 for input and $0.6 for output. On cache hits, the input token cost drops to a mere $0.003. It's worth keeping in mind that third-party providers have undercut Deepseek API in the past, so those figures could go lower if you shop around inference providers. Deepseek also offers search functionality at $0.75 per 1,000 queries, an amount that <a href="https://thenewstack.io/deepseek-web-search-api/" target="_blank">undercuts competitors</a>, with the closest one being Parallel at $1.</p><p>DeepSWE isn't the only benchmark around, either. Artificial Analysis ranks DSF-1017 (max) as only #2 on its intelligence index within the large open-weight class, with a score of 45. For reference, the most seemingly intelligent model, Claude Opus 5.5 (max with fallback), rings in at 58. It's one of the fastest models in its category, too, and its cost-per-task is edged out only by Xiaomi's MiMo V2.6-Pro.</p><p>This all sounds too good to be true, but everyone knows the adage of "fast, cheap, and reliable, pick two." So what's the catch? In one word: hallucination. Just because a model is smart doesn't mean it's reliable. Broadly speaking, users prefer models that don't answer when they lack supporting facts, and DSF-1017 doesn't seem very good at knowing what it doesn't know.</p><p>Artificial Analysis' <a href="https://artificialanalysis.ai/models/deepseek-v4-1-flash-1017?intelligence-per-task=intelligence-vs-output-tokens-per-task&amp;model-filters=large-models#omniscience-tabs" target="_blank">Omniscience benchmark</a> is quite clear on this matter, ranking the model among the lowest. It answered questions correctly 43% of the time, which is among the best for open-weight models, but inferior to the proprietary competition. That would be perfectly acceptable if it skipped answers without a foundation, but it hallucinated 67% of non-answers.</p><p>The benchmark's main index grants points for good answers and deducts them for wrong ones, and non-answers don't count. DSF-1017 gets 5 points, compared to Kimi K3-max's 20, with the frontier models coming in at 42+. Ultimately, this means DFS-1017 is by no means incompetent — much the opposite —but one should be careful about which tasks to assign to the model.</p><p>Tasks with verifiable outputs like programming, OCR with review, and format conversions will be fine, as the agent can run verification steps against its work and try again. However, open-ended research, looking up facts, and interpreting or drafting highly precise documents are tasks better suited to a frontier model — or, you know, a human — as hallucinations likely won't be properly reined in and results won't be very trustworthy.</p><p>"Fast and loose" seems to be a forming trend for highly competent open-weight models, too. Several months ago, when Kimi K3 <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-ai-releases-weights-for-kimi-k3-firing-a-shot-across-the-bow-of-openai-and-anthropic-open-weight-model-performs-almost-as-well-as-frontier-models-while-being-2-3x-easier-to-run">was released</a>, everyone raved about how effective and cheap it was, and how it proposed to offer frontier-level intelligence for a fraction of the price. My personal experience with Kimi was short-lived, as even short but precise fact look-ups came up with hallucinated information, and Artificial Analysis' data seems to back this observation up.</p><p>Nevertheless, open-weight models remain important both to make AI agents more accessible in absolute dollar terms and potentially as a tool to <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/frontier-ai-faces-pricing-reckoning-as-token-volume-explodes-25-fold-mid-tier-models-deliver-90-percent-of-flagship-capability-at-one-sixth-the-cost">drive down the cost of intelligence </a>when facing down potential price rises from frontier labs.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/artificial-intelligence/deepseeks-latest-ai-model-is-smarter-than-ever-but-its-low-running-costs-come-at-a-price-open-weight-model-is-fit-for-agentic-workloads-but-is-more-prone-to-hallucination-than-proprietary-competitors</link>
                                                                            <description>
                            <![CDATA[ Deepseek 4.1 Flash 1017 drives cost-per-task further down still — open-weight model is fit for agentic workloads and exceedingly cheap to run ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">rapuEnHrrAWLoj9AGDJEz3</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/QzsaWZn9raSdwCW8ecQ7gi-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Fri, 23 Oct 2026 12:00:00 +0000</pubDate>                                                                                                                                <updated>Fri, 23 Oct 2026 16:28:39 +0000</updated>
                                                                                                                                            <category><![CDATA[Artificial Intelligence]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                <author><![CDATA[ editors@tomshardware.com (Bruno Ferreira) ]]></author>                    <dc:creator><![CDATA[ Bruno Ferreira ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/ZQiPPaXaAuQ4VrVEYnnR7G-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Bruno Ferreira&#039;s journey kicked off with the venerable ZX Spectrum, a cassette player, and his hopes and dreams. He quickly realized he had more fun figuring out how computers work than he did actually using the things. Kicking off a developer career with C and Assembly before moving to scripting languages, he&#039;s worn many hats, including both database architect and systems administration. As a teen, Bruno co-founded a web development outfit where he was for 17 years before moving on to spend nearly a decade at The Tech Report as a writer, editor, and (of course) developer. In this decade, he&#039;s been at Asus, MLCommons, and HotHardware, among others. When not fiddling with computers and games, his love for music and production sends him off to live shows and festivals. Occasionally, he pretends he can play the guitar and bass.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/QzsaWZn9raSdwCW8ecQ7gi-1920-80.jpg">
                                                            <media:credit><![CDATA[Getty Images / SOPA Images]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[DeepSeek logo displayed on a smartphone screen]]></media:description>                                                            <media:text><![CDATA[DeepSeek logo displayed on a smartphone screen]]></media:text>
                                <media:title type="plain"><![CDATA[DeepSeek logo displayed on a smartphone screen]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/QzsaWZn9raSdwCW8ecQ7gi-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Deepseek just upgraded its V4.1 Flash open-weight model to revision 1017. What looks like a mere point revision may actually be a significant step forward in model efficiency. On the <a href="https://api-docs.deepseek.com/news/news261017" target="_blank">product page</a>, Deepseek uses the slogan "wish-driven programming made real," a likely reference to the colloquial "vibe coding," implying the model is a good fit for open-ended tasks with broad descriptions. However, while the model is indeed cheap and scores well, it can be hallucination-prone.</p><p>LLMs are roughly divided into two categories: closed commercial efforts from frontier labs, whose models are expensive to run but deliver excellent results, and Chinese open-weight competitors that broadly exchange reliability for a <em>much</em> cheaper running cost, particularly for long-running agentic tasks. </p><p>Deepseek V4.1 Flash (DSF) is a Mixture-of-Experts (MoE) model with 552B total parameters, 16B of which are active during runs. The model is multimodal, accepts both text and images as inputs, and its context window length measures the now-standard 1 million tokens.</p><p>According to Deepseek's internal benchmarks, the model slots in neatly between GPT-5.6 Sol and Claude Opus 5, even beating Anthropic's LLM in task tests. That's already a strong showing, but this latest revision of DSF shines in the performance-per-task base and token efficiency. The upgrade from V4 Flash to V4.1 Flash significantly slashed hardware requirements, with V4.1 using only one-quarter of the HBM and one-eighth of SSD storage for its KV cache compared to its predecessor, spending only 890 bytes per token entry.</p><p>The company illustrates the point with a cost-per-task graph based on <a href="https://deepswe.datacurve.ai/" target="_blank">DeepSWE's dataset</a> from September 2, adding its own data points for both DSF and DSF-1017. Both variants cluster in the desirable top-left area, and even low-effort runs of DSF-1017 offer a task completion rate in the ballpark of "max" settings for pricey competitors. It's worth noting that DeepSWE is only one of many benchmarks, but this is nonetheless an impressive result.</p><figure class="van-image-figure  extended-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:86.17%;"><img id="QJnQJ99DDimGvHPGhSi6oB" name="DeepSWE - Deepseek V4.1 Flash - 1017" alt="DeepSWE - Deepseek V4.1 Flash - 1017" src="https://cdn.mos.cms.futurecdn.net/QJnQJ99DDimGvHPGhSi6oB-1920-80.png" mos="" align="middle" fullscreen="1" width="2560" height="2206" attribution="" endorsement="" class="extended expandable" /><a href='https://cdn.mos.cms.futurecdn.net/QJnQJ99DDimGvHPGhSi6oB-1920-80.png' target='_blank' class='expand-button icon-expand-image icon' ></a></p></div></div><figcaption itemprop="caption description" class=" extended-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Deepseek)</span></figcaption></figure><p>Even at Deepseek's <a href="https://api-docs.deepseek.com/quick_start/pricing/" target="_blank">own pricing</a> (remember, this is an open-weight model you can run anywhere), the price at off-peak times per million tokens is just $0.15 for input and $0.6 for output. On cache hits, the input token cost drops to a mere $0.003. It's worth keeping in mind that third-party providers have undercut Deepseek API in the past, so those figures could go lower if you shop around inference providers. Deepseek also offers search functionality at $0.75 per 1,000 queries, an amount that <a href="https://thenewstack.io/deepseek-web-search-api/" target="_blank">undercuts competitors</a>, with the closest one being Parallel at $1.</p><p>DeepSWE isn't the only benchmark around, either. Artificial Analysis ranks DSF-1017 (max) as only #2 on its intelligence index within the large open-weight class, with a score of 45. For reference, the most seemingly intelligent model, Claude Opus 5.5 (max with fallback), rings in at 58. It's one of the fastest models in its category, too, and its cost-per-task is edged out only by Xiaomi's MiMo V2.6-Pro.</p><p>This all sounds too good to be true, but everyone knows the adage of "fast, cheap, and reliable, pick two." So what's the catch? In one word: hallucination. Just because a model is smart doesn't mean it's reliable. Broadly speaking, users prefer models that don't answer when they lack supporting facts, and DSF-1017 doesn't seem very good at knowing what it doesn't know.</p><p>Artificial Analysis' <a href="https://artificialanalysis.ai/models/deepseek-v4-1-flash-1017?intelligence-per-task=intelligence-vs-output-tokens-per-task&amp;model-filters=large-models#omniscience-tabs" target="_blank">Omniscience benchmark</a> is quite clear on this matter, ranking the model among the lowest. It answered questions correctly 43% of the time, which is among the best for open-weight models, but inferior to the proprietary competition. That would be perfectly acceptable if it skipped answers without a foundation, but it hallucinated 67% of non-answers.</p><p>The benchmark's main index grants points for good answers and deducts them for wrong ones, and non-answers don't count. DSF-1017 gets 5 points, compared to Kimi K3-max's 20, with the frontier models coming in at 42+. Ultimately, this means DFS-1017 is by no means incompetent — much the opposite —but one should be careful about which tasks to assign to the model.</p><p>Tasks with verifiable outputs like programming, OCR with review, and format conversions will be fine, as the agent can run verification steps against its work and try again. However, open-ended research, looking up facts, and interpreting or drafting highly precise documents are tasks better suited to a frontier model — or, you know, a human — as hallucinations likely won't be properly reined in and results won't be very trustworthy.</p><p>"Fast and loose" seems to be a forming trend for highly competent open-weight models, too. Several months ago, when Kimi K3 <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-ai-releases-weights-for-kimi-k3-firing-a-shot-across-the-bow-of-openai-and-anthropic-open-weight-model-performs-almost-as-well-as-frontier-models-while-being-2-3x-easier-to-run">was released</a>, everyone raved about how effective and cheap it was, and how it proposed to offer frontier-level intelligence for a fraction of the price. My personal experience with Kimi was short-lived, as even short but precise fact look-ups came up with hallucinated information, and Artificial Analysis' data seems to back this observation up.</p><p>Nevertheless, open-weight models remain important both to make AI agents more accessible in absolute dollar terms and potentially as a tool to <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/frontier-ai-faces-pricing-reckoning-as-token-volume-explodes-25-fold-mid-tier-models-deliver-90-percent-of-flagship-capability-at-one-sixth-the-cost">drive down the cost of intelligence </a>when facing down potential price rises from frontier labs.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ One easy change makes Microsoft's Surface Laptop Ultra with RTX Spark a better gaming system ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Nvidia's <a href="https://www.tomshardware.com/tablets/microsoft-surface/microsoft-surface-laptop-ultra-rtx-spark-review">RTX Spark</a> platform puts local AI inference and content creation first, but Microsoft's Surface Laptop Ultra really wants to game, too. The vented keyboard keeps the critical WASD keys cool under load by design, the smooth 120Hz screen can get as bright as 2000 nits with HDR games, and the built-in speakers may be the best I've ever heard on a laptop, period. </p><p>There's just one problem: that 3270x2180 screen asks a lot of DLSS upscaling. The computation that DLSS performs doesn't come for free, and the load increases with higher output resolutions. This is especially relevant with the Surface Laptop Ultra's high-density panel, which has just 14% fewer pixels than a 16:9 4K display. </p><p>The Surface's unusual 3:2 aspect ratio (at least compared to the average gaming laptop) also means that scaling content up to fill the screen isn't straightforward in many games. 16:9 and 16:10 content gets stretched vertically, causing obvious distortion of character models and objects like tires. And if you use windowed mode instead, playing in a 16:9 or 16:10 window on a relatively small screen like this quickly tires the eyes. </p><p>The thing is, all this is easy to solve—easy enough, in fact, that I'm surprised I had to do it myself at all. </p><p>Microsoft seems aware that lower 3:2 resolutions are useful for this system, as it put a 1600x1024 custom resolution in the list of display modes that's slightly less than half of native on both axes. But at roughly 1.6 megapixels in total, that image offers less resolution than even a 1920x1080 display does. Scaling up from such a low-res baseline on the Surface screen results in a noticeably soft image even before we dial in DLSS. </p><p>What we really want, for reasons of both performance and adequate sharpness, is a 3:2 resolution that comes close to the 2.07 megapixels that 1920x1080 yields. Some quick math and rounding to whole numbers gives us 1764x1176 as an optimal target. </p><p>With those ideal dimensions in hand, the next step was to create a custom resolution to match. The Nvidia Control Panel doesn't work on Windows on Arm, but we can still use the Nvidia App to set custom resolutions. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1980px;"><p class="vanilla-image-block" style="padding-top:56.26%;"><img id="ZfF4UNYUebu8NXBdkycTw7" name="nvidia app" alt="Custom resolution selection in the Nvidia App" src="https://cdn.mos.cms.futurecdn.net/ZfF4UNYUebu8NXBdkycTw7-1920-80.png" mos="" align="middle" fullscreen="1" width="1980" height="1114" attribution="" endorsement="" class="inline expandable" /><a href='https://cdn.mos.cms.futurecdn.net/ZfF4UNYUebu8NXBdkycTw7-1920-80.png' target='_blank' class='expand-button icon-expand-image icon' ></a></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Future)</span></figcaption></figure><p>After diving into the System tab and expanding the Display Properties pane, I found the "Customize Resolution" button to open an editor to that end and added 1764x1176 at 120Hz. I then tested it and confirmed that it worked as a supported display mode. (If you want to match the 3.7 MP of 2560x1440, set up a 2352x1568 resolution while you’re at it.)</p><p>If for some reason you don't want non-3:2 content to fit the entire Surface Laptop Ultra screen, there is a Scaling pane in the Nvidia App that lets you control that behavior. The "Aspect Ratio" scaling mode will letterbox 16:9 and 16:10 material if you set the display resolution accordingly. But if you're paying $5899 for this system, I presume you want to use every one of its pricey pixels rather than leaving some of them dark.</p><p>So what does that bit of tweaking buy us compared to our initial gaming results with <a href="https://www.tomshardware.com/video-games/pc-gaming/control-resonant-pc-performance-tested-28-gpus-take-us-back-to-the-oldest-house-and-a-warped-manhattan-cityscape"><em>Control Resonant</em></a>?</p><p>As someone who primarily games on our <a href="https://www.tomshardware.com/pc-components/cpus/amd-ryzen-7-9850x3d-review">Ryzen 7 9800X3D</a> GPU testing system and not a laptop, I wanted to make sure I wasn't setting unreasonable expectations for the Surface Laptop for input latency or CPU performance. So I hauled out an older 17" gaming laptop with a Core i9-11980HK and an RTX 3080 Mobile 16GB inside to get some perspective. </p><p>To level the playing field with the Surface Laptop Ultra and its native support for frame generation, I modded Optiscaler into <em>Control Resonant </em>on that system to enable Intel's Xe Frame Generation and Xe Low Latency features. </p><p>I conducted my testing with <em>Control Resonant'</em>s High raster preset and medium RT preset, which gets you a noticeable boost in image quality thanks in part to <a href="https://www.tomshardware.com/video-games/pc-gaming/dlss-4-5-ray-reconstruction-update-arrives-in-august-for-better-ray-tracing-visuals-broader-training-data-set-and-second-gen-transformer-architecture-combine-for-improved-image-quality">DLSS Ray Reconstruction</a>.</p><div ><table><caption>Control Resonant performance - High raster, Medium RT, DLSS Ultra Performance, FG 2X</caption><tbody><tr><td class="firstcol empty" ></td><td  ><p><strong>Average FPS</strong></p></td><td  ><p><strong>1% Low FPS</strong></p></td><td  ><p><strong>Average input latency (ms)</strong></p></td></tr><tr><td class="firstcol " ><p><strong>Surface Laptop Ultra - 3270x2180</strong></p></td><td  ><p>63.5</p></td><td  ><p>38.1</p></td><td  ><p>87. 9</p></td></tr></tbody></table></div><p>You can see that with these settings, trying to upscale Control Resonant to the Surface Laptop Ultra's native resolution is too much to ask for a smooth or responsive gaming experience. Even with DLSS Ultra Performance (an input resolution of 1090x727) and 2X frame generation, we only just break 60 FPS, and latency is at an unacceptable 87 ms average. There's no latency budget left to turn up MFG without turning down other image quality settings. </p><div ><table><caption>Control Resonant performance - High raster, Medium RT, DLSS Quality, FG 2X</caption><tbody><tr><td class="firstcol empty" ></td><td  ><p><strong>Average FPS</strong></p></td><td  ><p><strong>1% Low FPS</strong></p></td><td  ><p><strong>Average input latency (ms)</strong></p></td></tr><tr><td class="firstcol " ><p><strong>Surface Laptop Ultra - 1764x1176</strong></p></td><td  ><p>83.4 </p></td><td  ><p>42.9</p></td><td  ><p>67.8</p></td></tr><tr><td class="firstcol " ><p><strong>RTX 3080 Mobile + 11900HK - 1920x1080 </strong></p></td><td  ><p>87.1</p></td><td  ><p>53</p></td><td  ><p>69</p></td></tr></tbody></table></div><p>With our 1764x1176 custom resolution, however, we're off to a great start with DLSS Quality upscaling (an input resolution of 1176x874). The resulting 83.4 FPS average, 42.9 1% low FPS, and an average input latency of 67.8 ms is a fine experience. We still have room to dial in MFG 3X, leading to a 123 FPS average and 55 FPS 1% lows while keeping latency a reasonable 71 ms. </p><p>The RTX 3080 Mobile system with XeFG and XeLL, meanwhile, turns in an 87.1 FPS average, 53 FPS 1% lows, and a 69 ms average input latency at 1920x1080 with the same image quality settings. Its key advantage is in 1% lows, but that win comes with significant trade-offs. </p><p>Plugged into our Watts Up meter, the RTX 3080 system draws 250 W from the wall to deliver the experience it does, but according to a USB-C power meter, the Surface Laptop Ultra pulls between 110 W and 120 W under load. Silicon has come a long way since Tiger Lake and Ampere. </p><p>The RTX 3080 system and its enormous power adapter together also weigh over nine pounds (4.1 kg). It's technically portable but practically a pain to haul around. The Surface Laptop Ultra and its 140W adapter weigh just 5.3 lb (2.39 kg), which isn't featherweight but won't break your back. </p><p>And the Surface will actually fit on a tray table or your lap, with enough battery life that you can use it through a transcontinental U.S. flight, say, and still have plenty of juice left for either side of the trip. </p><p>So at least in <em>Control Resonant</em>, the Surface Laptop Ultra is delivering largely the same gaming experience as that high-end RTX 3080 Mobile system, but in a much smaller and lighter total package that draws less than half the wall power. And if you want to stretch frame rates further with MFG 3X, the Surface has MFG support that the older system doesn't. Just add in the 1764x1176 intermediate resolution that Microsoft forgot, and you’ll be well on your way to a better gaming experience. </p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/video-games/pc-gaming/one-easy-change-makes-microsofts-surface-laptop-ultra-with-rtx-spark-a-better-gaming-system-custom-resolutions-make-dlss-easier-on-the-device</link>
                                                                            <description>
                            <![CDATA[ Microsoft's Surface Laptop Ultra with RTX Spark might not be a gaming laptop first and foremost, but it certainly wants to game. We found that creating a custom 1920x1080-equivalent resolution for its 3:2 screen, missing from its out-of-the-box config, makes for a better experience. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">tpJYDWEKXcJ4XvabEbk95Z</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/eaoVZGZzRqzXmF7J97SBbE-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Thu, 22 Oct 2026 13:30:00 +0000</pubDate>                                                                                                                                                                                                                                <category><![CDATA[PC Gaming]]></category>
                                                    <category><![CDATA[Video Games]]></category>
                                                                                                                    <dc:creator><![CDATA[ Jeffrey Kampman ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/8JCjGs5yVZds2YdKmzjUDE-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Jeff Kampman has been playing PC games ever since he learned how to fire up freeware CDs from the DOS command line. He started building his own PCs in the mid-aughts and later turned that passion into a career, working as a news and guides writer, reviewer, and ultimately Editor-in-Chief at The Tech Report, where he dove deep on CPUs and GPUs (and more) in pursuit of the smoothest gaming experiences around. Jeff later took on roles at Asus and Intel as a technical marketer before joining Tom&#039;s Hardware. As Senior Analyst, Graphics, Jeff covers everything from integrated graphics processors to discrete graphics cards to the massive data center GPU installations powering our AI future. Jeff is also a hobbyist photographer, Twitch streamer, espresso enthusiast, and runner.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/eaoVZGZzRqzXmF7J97SBbE-1920-80.jpg">
                                                            <media:credit><![CDATA[Future]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[A large gaming laptop next to a smaller, thinner creator laptop]]></media:description>                                                            <media:text><![CDATA[A large gaming laptop next to a smaller, thinner creator laptop]]></media:text>
                                <media:title type="plain"><![CDATA[A large gaming laptop next to a smaller, thinner creator laptop]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/eaoVZGZzRqzXmF7J97SBbE-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Nvidia's <a href="https://www.tomshardware.com/tablets/microsoft-surface/microsoft-surface-laptop-ultra-rtx-spark-review">RTX Spark</a> platform puts local AI inference and content creation first, but Microsoft's Surface Laptop Ultra really wants to game, too. The vented keyboard keeps the critical WASD keys cool under load by design, the smooth 120Hz screen can get as bright as 2000 nits with HDR games, and the built-in speakers may be the best I've ever heard on a laptop, period. </p><p>There's just one problem: that 3270x2180 screen asks a lot of DLSS upscaling. The computation that DLSS performs doesn't come for free, and the load increases with higher output resolutions. This is especially relevant with the Surface Laptop Ultra's high-density panel, which has just 14% fewer pixels than a 16:9 4K display. </p><p>The Surface's unusual 3:2 aspect ratio (at least compared to the average gaming laptop) also means that scaling content up to fill the screen isn't straightforward in many games. 16:9 and 16:10 content gets stretched vertically, causing obvious distortion of character models and objects like tires. And if you use windowed mode instead, playing in a 16:9 or 16:10 window on a relatively small screen like this quickly tires the eyes. </p><p>The thing is, all this is easy to solve—easy enough, in fact, that I'm surprised I had to do it myself at all. </p><p>Microsoft seems aware that lower 3:2 resolutions are useful for this system, as it put a 1600x1024 custom resolution in the list of display modes that's slightly less than half of native on both axes. But at roughly 1.6 megapixels in total, that image offers less resolution than even a 1920x1080 display does. Scaling up from such a low-res baseline on the Surface screen results in a noticeably soft image even before we dial in DLSS. </p><p>What we really want, for reasons of both performance and adequate sharpness, is a 3:2 resolution that comes close to the 2.07 megapixels that 1920x1080 yields. Some quick math and rounding to whole numbers gives us 1764x1176 as an optimal target. </p><p>With those ideal dimensions in hand, the next step was to create a custom resolution to match. The Nvidia Control Panel doesn't work on Windows on Arm, but we can still use the Nvidia App to set custom resolutions. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1980px;"><p class="vanilla-image-block" style="padding-top:56.26%;"><img id="ZfF4UNYUebu8NXBdkycTw7" name="nvidia app" alt="Custom resolution selection in the Nvidia App" src="https://cdn.mos.cms.futurecdn.net/ZfF4UNYUebu8NXBdkycTw7-1920-80.png" mos="" align="middle" fullscreen="1" width="1980" height="1114" attribution="" endorsement="" class="inline expandable" /><a href='https://cdn.mos.cms.futurecdn.net/ZfF4UNYUebu8NXBdkycTw7-1920-80.png' target='_blank' class='expand-button icon-expand-image icon' ></a></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Future)</span></figcaption></figure><p>After diving into the System tab and expanding the Display Properties pane, I found the "Customize Resolution" button to open an editor to that end and added 1764x1176 at 120Hz. I then tested it and confirmed that it worked as a supported display mode. (If you want to match the 3.7 MP of 2560x1440, set up a 2352x1568 resolution while you’re at it.)</p><p>If for some reason you don't want non-3:2 content to fit the entire Surface Laptop Ultra screen, there is a Scaling pane in the Nvidia App that lets you control that behavior. The "Aspect Ratio" scaling mode will letterbox 16:9 and 16:10 material if you set the display resolution accordingly. But if you're paying $5899 for this system, I presume you want to use every one of its pricey pixels rather than leaving some of them dark.</p><p>So what does that bit of tweaking buy us compared to our initial gaming results with <a href="https://www.tomshardware.com/video-games/pc-gaming/control-resonant-pc-performance-tested-28-gpus-take-us-back-to-the-oldest-house-and-a-warped-manhattan-cityscape"><em>Control Resonant</em></a>?</p><p>As someone who primarily games on our <a href="https://www.tomshardware.com/pc-components/cpus/amd-ryzen-7-9850x3d-review">Ryzen 7 9800X3D</a> GPU testing system and not a laptop, I wanted to make sure I wasn't setting unreasonable expectations for the Surface Laptop for input latency or CPU performance. So I hauled out an older 17" gaming laptop with a Core i9-11980HK and an RTX 3080 Mobile 16GB inside to get some perspective. </p><p>To level the playing field with the Surface Laptop Ultra and its native support for frame generation, I modded Optiscaler into <em>Control Resonant </em>on that system to enable Intel's Xe Frame Generation and Xe Low Latency features. </p><p>I conducted my testing with <em>Control Resonant'</em>s High raster preset and medium RT preset, which gets you a noticeable boost in image quality thanks in part to <a href="https://www.tomshardware.com/video-games/pc-gaming/dlss-4-5-ray-reconstruction-update-arrives-in-august-for-better-ray-tracing-visuals-broader-training-data-set-and-second-gen-transformer-architecture-combine-for-improved-image-quality">DLSS Ray Reconstruction</a>.</p><div ><table><caption>Control Resonant performance - High raster, Medium RT, DLSS Ultra Performance, FG 2X</caption><tbody><tr><td class="firstcol empty" ></td><td  ><p><strong>Average FPS</strong></p></td><td  ><p><strong>1% Low FPS</strong></p></td><td  ><p><strong>Average input latency (ms)</strong></p></td></tr><tr><td class="firstcol " ><p><strong>Surface Laptop Ultra - 3270x2180</strong></p></td><td  ><p>63.5</p></td><td  ><p>38.1</p></td><td  ><p>87. 9</p></td></tr></tbody></table></div><p>You can see that with these settings, trying to upscale Control Resonant to the Surface Laptop Ultra's native resolution is too much to ask for a smooth or responsive gaming experience. Even with DLSS Ultra Performance (an input resolution of 1090x727) and 2X frame generation, we only just break 60 FPS, and latency is at an unacceptable 87 ms average. There's no latency budget left to turn up MFG without turning down other image quality settings. </p><div ><table><caption>Control Resonant performance - High raster, Medium RT, DLSS Quality, FG 2X</caption><tbody><tr><td class="firstcol empty" ></td><td  ><p><strong>Average FPS</strong></p></td><td  ><p><strong>1% Low FPS</strong></p></td><td  ><p><strong>Average input latency (ms)</strong></p></td></tr><tr><td class="firstcol " ><p><strong>Surface Laptop Ultra - 1764x1176</strong></p></td><td  ><p>83.4 </p></td><td  ><p>42.9</p></td><td  ><p>67.8</p></td></tr><tr><td class="firstcol " ><p><strong>RTX 3080 Mobile + 11900HK - 1920x1080 </strong></p></td><td  ><p>87.1</p></td><td  ><p>53</p></td><td  ><p>69</p></td></tr></tbody></table></div><p>With our 1764x1176 custom resolution, however, we're off to a great start with DLSS Quality upscaling (an input resolution of 1176x874). The resulting 83.4 FPS average, 42.9 1% low FPS, and an average input latency of 67.8 ms is a fine experience. We still have room to dial in MFG 3X, leading to a 123 FPS average and 55 FPS 1% lows while keeping latency a reasonable 71 ms. </p><p>The RTX 3080 Mobile system with XeFG and XeLL, meanwhile, turns in an 87.1 FPS average, 53 FPS 1% lows, and a 69 ms average input latency at 1920x1080 with the same image quality settings. Its key advantage is in 1% lows, but that win comes with significant trade-offs. </p><p>Plugged into our Watts Up meter, the RTX 3080 system draws 250 W from the wall to deliver the experience it does, but according to a USB-C power meter, the Surface Laptop Ultra pulls between 110 W and 120 W under load. Silicon has come a long way since Tiger Lake and Ampere. </p><p>The RTX 3080 system and its enormous power adapter together also weigh over nine pounds (4.1 kg). It's technically portable but practically a pain to haul around. The Surface Laptop Ultra and its 140W adapter weigh just 5.3 lb (2.39 kg), which isn't featherweight but won't break your back. </p><p>And the Surface will actually fit on a tray table or your lap, with enough battery life that you can use it through a transcontinental U.S. flight, say, and still have plenty of juice left for either side of the trip. </p><p>So at least in <em>Control Resonant</em>, the Surface Laptop Ultra is delivering largely the same gaming experience as that high-end RTX 3080 Mobile system, but in a much smaller and lighter total package that draws less than half the wall power. And if you want to stretch frame rates further with MFG 3X, the Surface has MFG support that the older system doesn't. Just add in the 1764x1176 intermediate resolution that Microsoft forgot, and you’ll be well on your way to a better gaming experience. </p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ FADU puts the filesystem inside the SSD's flash to bypass host CPUs ]]></title>
                                                                                                <dc:content><![CDATA[ <p>SSD controller designer FADU’s new File-Semantic SSD (FS-SSD) was announced at the 2026 OCP Global Summit. The new device promises to move the file system into the SSD itself.  As detailed in <a href="https://www.fadu.io/files/whitepaper/FSSSD_whitepaper-1.0.2.pdf" target="_blank">a white paper</a>, this enables GPUs, DPUs, and NICs to open and read files with “no host filesystem, no block layer, and no storage-server CPU in the path.” FADU told <em>Tom’s Hardware Premium</em> it expects partner samples by Q3 2027. </p><p>FADU is a Seoul, South Korea-based <a href="https://www.tomshardware.com/news/gen-5-nvme-ssd-fadu-technology">fabless SSD controller designer</a>. Its FS-SSD technology is designed to handle and improve the control path, as the data path side moved to accelerators years ago. FADU’s measurements show that the file system, plus block layers, can consume about 30% of CPU cycles in a <a href="https://www.tomshardware.com/pc-components/ssds/highpoint-enables-gpudirect-storage-with-new-adapter-up-to-64-gb-s-from-storage-to-gpu-without-cpu-involvement">GPUDirect Storage</a> (GDS) random-read workload. It notes that <a href="https://www.tomshardware.com/news/nvidia-unveils-big-accelerator-memory-solid-state-storage-for-gpus">bypass options</a> give up the files, and the company claims its solution promises the best of both worlds.</p><p>In FADU’s own preliminary Network File System (NFS) over NVMe test, on a drive built on its FC6161 Gen6 controller in a PCIe 5.0 host, remote file throughput held steady as the server downshifted from what FADU labels 16 cores to just one, maintaining about 460,000 ops/s.</p><p>The NFSv3 operations are carried inside standard NVMe commands, and the drive handles path lookup, access control, and file-to-flash mapping. FS-SSD’s performance is remarkable when compared to Linux NFS-over-RDMA, which achieved a third of the ops/s with 16 cores and only 40K ops/s with one core. The technology remains “under active development,” and FADU told <em>Tom’s Hardware Premium</em> it expects Gen6-host results by Q1 2027, with partner samples in Q3 2027.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:787px;"><p class="vanilla-image-block" style="padding-top:51.46%;"><img id="n6wajtY4Xd99ZTD5Vhk6F6" name="fadu-fs-ssd-whitepaper-fig3-filebench" alt="FADU chart of Filebench fileserver results: FS-SSD holds about 460,000 ops/s from 16 server cores down to one, while NFS over RDMA falls from about 150,000 to 40,000 ops/s" src="https://cdn.mos.cms.futurecdn.net/n6wajtY4Xd99ZTD5Vhk6F6-1920-80.png" mos="" align="middle" fullscreen="" width="787" height="405" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="caption-text">FS-SSD holds about 460,000 ops/s from 16 server cores down to one, while NFS over RDMA falls from approximately 150,000 to 40,000 ops/s </span><span class="credit" itemprop="copyrightHolder">(Image credit: <a href="https://www.fadu.io/files/whitepaper/FSSSD_whitepaper-1.0.2.pdf">FADU</a>)</span></figcaption></figure><p>FADU chooses to use NFSv3 because it’s “small (21 operations), OS- and filesystem-agnostic, and stateless” with low firmware overhead. FS-SSD supports PCIe today, with <a href="https://www.tomshardware.com/news/ethernet-ssds-come-to-the-fore-as-kioxia-starts-sampling">NVMe-oF</a> over RDMA and TCP working without any new transport needed. Reusing an existing protocol is a smart move, in our view. Under the hood, the file system and flash translation layer are merged with a file layer derived from F2FS. Node and data blocks are aligned with the natural flash pages, sections at corresponding erase units, so that the F2FS node table becomes a “File Map” pointing the node IDs at physical flash addresses.</p><p>One big advantage is that the DRAM mapping metadata overhead is considerably smaller: it uses the same amount of DRAM as a standard SSD, ~0.1% of NAND capacity (about 1 GB per TB), which a standard SSD fills almost entirely with its mapping table. Here, most of that DRAM, ~95%, can act as a hot-path node cache.</p><p>One garbage collector also replaces two, avoiding what FADU calls the “double GC” of host F2FS, where file-system cleanup and the drive’s own garbage collection amplify each other. Write amplification is also 63% lower versus a host F2FS solution on a mixed sequential and random workload with FADU’s research prototype, the company said. FADU told <em>Tom's Hardware Premium</em> that the prototype runs synthetic fio workloads on NVMeVirt, an open-source software SSD emulator from Seoul National University, so it is not a hardware result.</p><p>Most importantly for this method, it ensures host-free access. The GPU builds NFS-over-NVMe commands in queues in its own memory, which could include HBM. Alternatively, a remote SmartNIC or DPU with NVMe-oF target offload can drop commands in the drive’s submission queue. The data goes from SSD to NIC via PCIe peer-to-peer DMA, coming out as RDMA writes. FADU’s explicitly named use cases for this include GPU file reads and checkpoints, DPU-served HTTP content delivery, and a JBoF (just a bunch of flash) file server.</p><p>FADU’s numbers for its internal test come from a client and server Intel Core i5-13600K, a CPU with 20 threads that clocks up to 5.1 GHz with 128GB of DDR5, and a Mellanox ConnectX-6 100 Gbps NIC. Synthetic testing is compared using fio with the baseline F2FS and Ext4 on block firmware versus the FS-SSD via a kernel client and an SPDK user-space client. The sequential test is 1 thread, queue depth (QD) 128, with 128KB blocks, while random is at 4 threads, the same QD128, but with 4KB blocks.</p><p>These block sizes are typical for client workloads in our experience, so they are a fair measure, with detailed preliminary results in the table below. Against SPDK on the raw drive, figures FADU provided to us, FS-SSD’s SPDK client keeps about 93% on 4K random read and 53% on 4K random write, via fio, with around a 7% and 47% cost for the in-drive file layer. Though the raw drive’s 13,240 MB/s random write is above the ~12 GB/s FADU gives as the NAND’s program bandwidth, those runs likely weren’t at steady state. The Filebench result is particularly interesting, as the SPDK random gains apply to a plain block device too.</p><p>This should be the workload most favorable for the technology, as the Filebench file server is metadata-heavy with relatively high host overhead: FADU says it uses 10,000 files in a multi-level tree, mean directory width 20, and a mean file size of 128KB with 50 client threads.</p><div ><table><caption>FADU’s preliminary fio results: Gen6 controller in a PCIe 5.0 host, direct I/O</caption><thead><tr><th class="firstcol " ><p>Setup</p></th><th  ><p>Sequential read (MB/s)</p></th><th  ><p>Sequential write (MB/s)</p></th><th  ><p>4KB random read (MB/s)</p></th><th  ><p>4KB random write (MB/s)</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>F2FS</p></td><td  ><p>14,120</p></td><td  ><p>4,510</p></td><td  ><p>7,850</p></td><td  ><p>3,980</p></td></tr><tr><td class="firstcol " ><p>Ext4</p></td><td  ><p>14,130</p></td><td  ><p>7,050</p></td><td  ><p>9,120</p></td><td  ><p>4,560</p></td></tr><tr><td class="firstcol " ><p>FS-SSD (kernel client)</p></td><td  ><p>14,140</p></td><td  ><p>11,940</p></td><td  ><p>8,070</p></td><td  ><p>4,320</p></td></tr><tr><td class="firstcol " ><p>FS-SSD (SPDK client)</p></td><td  ><p>Not reported</p></td><td  ><p>Not reported</p></td><td  ><p>13,000</p></td><td  ><p>7,030</p></td></tr><tr><td class="firstcol " ><p>Raw block (SPDK)</p></td><td  ><p>Not reported</p></td><td  ><p>Not reported</p></td><td  ><p>13,990</p></td><td  ><p>13,240</p></td></tr></tbody></table></div><p>One important thing to recall is that although this test is on a <a href="https://www.tomshardware.com/tech-industry/the-current-state-of-pcie-6-0-ssds-and-controllers-marvell-phison-and-smi-prepare-controllers-as-drives-finally-come-to-market-following-years-of-delays">Gen6 </a>drive, the test results are on a PCIe 5.0 host. Therefore, every sequential read config tops out at around 14.1 GB/s, which is roughly the link limit. The testbed also uses a consumer-class desktop CPU, while the one-core NFS server is a showcase example. The kernel client ends up 11.5% behind Ext4 on random read and 5.3% behind on random write, again by our math.</p><p>One example of a tough question more generally is: how do concurrent GPU, DPU, and host writers coordinate? How does a kernel mount’s page cache stay coherent with GPU access? NFSv3 storage devices, by the pNFS Flexible File Layout spec (RFC 8435), are “always treated as loosely coupled,” which means “Mandatory byte-range locking cannot be supported.”</p><p>Credentials riding in the NVMe command may also face security issues, as loosely coupled devices get only “a level of security equivalent to AUTH_SYS,” NFS’s basic scheme, in which the server trusts whatever user and group IDs the client sends. FADU’s published material also does not describe features such as snapshots, encryption handling, quotas, ACLs, crash handling, and so forth, related to this. Statelessness is what makes it work, but it doesn’t solve all problems, so the trust factor moves elsewhere.</p><p>In reply to our questioning, FADU said that there is no cross-host locking, as NFS’s lock manager isn’t implemented. The drive applies each write whole, in order. If two writes hit the same spot, the second one wins; this is “last-writer-wins.” Coordination depends on the clients: kernel-mounted ones get the usual advisory locks among themselves, GPU and SPDK clients coordinate in the application, and at scale a pNFS metadata server handles the coordination. Caching is also handled like ordinary NFS, with kernel clients, or is not done at all, for GPU and SPDK clients. This is stock NFSv3 behavior.</p><p>The drive has to check user/group IDs against the file permissions itself. Because those IDs ride inside the command, a client could fake them, but FADU says it closes that issue with capability-based authentication. Each command carries a capability issued by an outside authentication server, plus a code computed with a key shared only between that server and the drive. This is not detailed in its published material. FADU adds that support for the Kerberos authentication protocol is planned. For networked drives, the standard NVMe-oF protections are used, with only pre-approved hosts admitted. A drive trusts a single host’s kernel like any NVMe drive.</p><p>However, as FADU described to us, ACLs and other features are missing in the current release, aside from encryption, which in this case is handled by the FC6161’s standard TCG Opal self-encrypting drive support. Power loss is also typically handled with capacitors flushing metadata, the write buffer, and a log of in-flight operations. On restart, the drive replays the unfinished operations, and every write is durable if the drive acknowledges it.</p><p>FADU’s longer sights are on producing a JBoF storage solution with no server in front: FS-SSDs, a PCIe switch, and a DPU or NIC for handling NVMe-oF in hardware. FADU’s model puts this at about 42% lower power, about 57% lower cost per 16-drive server, and higher rack density with significantly lower five-year total cost of ownership. The drives would slot in behind existing pNFS metadata servers and could scale. FADU’s controllers already ship with drives such as the SN861, which Western Digital launched and SanDisk now sells, but this would expand the options and solutions it can provide.</p><div ><table><caption>FADU’s FS-SSD roadmap</caption><thead><tr><th class="firstcol " ><p>Milestone</p></th><th  ><p>Timing (per FADU)</p></th><th  ><p>Details</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Gen6-host results</p></td><td  ><p>By Q1 2027</p></td><td  ><p>FS-SSD tested in a PCIe 6.0 host; hardware write-amplification figures to come with them</p></td></tr><tr><td class="firstcol " ><p>Partner samples</p></td><td  ><p>Q3 2027</p></td><td  ><p>OEM/ODM talks are at an early stage; active collaboration expected once samples ship</p></td></tr><tr><td class="firstcol " ><p>Hostless JBoF</p></td><td  ><p>By end of 2027</p></td><td  ><p>Reference design built on Xsight Labs E1 DPUs, delivered to OEM/ODM partners</p></td></tr><tr><td class="firstcol " ><p>GPU access</p></td><td  ><p>Now, via GPUDirect Storage</p></td><td  ><p>Works with GDS through the kernel client; FADU intends to join the Accelerated IO SIG (xio-sig), where Nvidia open-sourced cuFile; no direct engagement with Nvidia</p></td></tr><tr><td class="firstcol " ><p>Standards</p></td><td  ><p>Nothing submitted yet</p></td><td  ><p>FADU plans to start a file-semantic command set discussion at NVM Express and OCP Storage; samples and client code come first</p></td></tr></tbody></table></div><p>Next year should bring real numbers and samples, with the JBoF reference design also expected. There may be an NVM Express or OCP Storage proposal down the line for a file-semantic command set to further cement and mature the approach. FADU has a good idea, is building it on an existing framework, and has a testbed with results and a hardware path forward on its Gen6 controller, with a trust model in mind. A Gen6-host run is now needed to support the company’s claims.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/pc-components/ssds/fadu-puts-the-filesystem-inside-the-ssds-flash-to-bypass-host-cpus-file-semantic-pcie-6-0-ssd-delivers-460-000-ops-s-to-gpus-and-dpus-samples-due-by-q3-2027</link>
                                                                            <description>
                            <![CDATA[ FADU’s FS-SSD carries NFSv3 file operations inside NVMe commands, letting GPUs, DPUs, and NICs read files without a host file system. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">tgdhFNS2yM4zWBSpUrQjbC</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/FfxWhJKRcbYcywYyA2ctLL-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Wed, 21 Oct 2026 12:30:00 +0000</pubDate>                                                                                                                                <updated>Wed, 21 Oct 2026 14:40:48 +0000</updated>
                                                                                                                                            <category><![CDATA[SSDs]]></category>
                                                    <category><![CDATA[PC Components]]></category>
                                                    <category><![CDATA[Storage]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/FfxWhJKRcbYcywYyA2ctLL-1920-80.jpg">
                                                            <media:credit><![CDATA[FADU]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[A curved monitor at FADU’s OCP Global Summit booth shows a Filebench dashboard comparing a live FS-SSD run with a recorded NFS-over-RDMA run, beside an open test chassis with an SSD mounted on a riser]]></media:description>                                                            <media:text><![CDATA[A curved monitor at FADU’s OCP Global Summit booth shows a Filebench dashboard comparing a live FS-SSD run with a recorded NFS-over-RDMA run, beside an open test chassis with an SSD mounted on a riser]]></media:text>
                                <media:title type="plain"><![CDATA[A curved monitor at FADU’s OCP Global Summit booth shows a Filebench dashboard comparing a live FS-SSD run with a recorded NFS-over-RDMA run, beside an open test chassis with an SSD mounted on a riser]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/FfxWhJKRcbYcywYyA2ctLL-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>SSD controller designer FADU’s new File-Semantic SSD (FS-SSD) was announced at the 2026 OCP Global Summit. The new device promises to move the file system into the SSD itself.  As detailed in <a href="https://www.fadu.io/files/whitepaper/FSSSD_whitepaper-1.0.2.pdf" target="_blank">a white paper</a>, this enables GPUs, DPUs, and NICs to open and read files with “no host filesystem, no block layer, and no storage-server CPU in the path.” FADU told <em>Tom’s Hardware Premium</em> it expects partner samples by Q3 2027. </p><p>FADU is a Seoul, South Korea-based <a href="https://www.tomshardware.com/news/gen-5-nvme-ssd-fadu-technology">fabless SSD controller designer</a>. Its FS-SSD technology is designed to handle and improve the control path, as the data path side moved to accelerators years ago. FADU’s measurements show that the file system, plus block layers, can consume about 30% of CPU cycles in a <a href="https://www.tomshardware.com/pc-components/ssds/highpoint-enables-gpudirect-storage-with-new-adapter-up-to-64-gb-s-from-storage-to-gpu-without-cpu-involvement">GPUDirect Storage</a> (GDS) random-read workload. It notes that <a href="https://www.tomshardware.com/news/nvidia-unveils-big-accelerator-memory-solid-state-storage-for-gpus">bypass options</a> give up the files, and the company claims its solution promises the best of both worlds.</p><p>In FADU’s own preliminary Network File System (NFS) over NVMe test, on a drive built on its FC6161 Gen6 controller in a PCIe 5.0 host, remote file throughput held steady as the server downshifted from what FADU labels 16 cores to just one, maintaining about 460,000 ops/s.</p><p>The NFSv3 operations are carried inside standard NVMe commands, and the drive handles path lookup, access control, and file-to-flash mapping. FS-SSD’s performance is remarkable when compared to Linux NFS-over-RDMA, which achieved a third of the ops/s with 16 cores and only 40K ops/s with one core. The technology remains “under active development,” and FADU told <em>Tom’s Hardware Premium</em> it expects Gen6-host results by Q1 2027, with partner samples in Q3 2027.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:787px;"><p class="vanilla-image-block" style="padding-top:51.46%;"><img id="n6wajtY4Xd99ZTD5Vhk6F6" name="fadu-fs-ssd-whitepaper-fig3-filebench" alt="FADU chart of Filebench fileserver results: FS-SSD holds about 460,000 ops/s from 16 server cores down to one, while NFS over RDMA falls from about 150,000 to 40,000 ops/s" src="https://cdn.mos.cms.futurecdn.net/n6wajtY4Xd99ZTD5Vhk6F6-1920-80.png" mos="" align="middle" fullscreen="" width="787" height="405" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="caption-text">FS-SSD holds about 460,000 ops/s from 16 server cores down to one, while NFS over RDMA falls from approximately 150,000 to 40,000 ops/s </span><span class="credit" itemprop="copyrightHolder">(Image credit: <a href="https://www.fadu.io/files/whitepaper/FSSSD_whitepaper-1.0.2.pdf">FADU</a>)</span></figcaption></figure><p>FADU chooses to use NFSv3 because it’s “small (21 operations), OS- and filesystem-agnostic, and stateless” with low firmware overhead. FS-SSD supports PCIe today, with <a href="https://www.tomshardware.com/news/ethernet-ssds-come-to-the-fore-as-kioxia-starts-sampling">NVMe-oF</a> over RDMA and TCP working without any new transport needed. Reusing an existing protocol is a smart move, in our view. Under the hood, the file system and flash translation layer are merged with a file layer derived from F2FS. Node and data blocks are aligned with the natural flash pages, sections at corresponding erase units, so that the F2FS node table becomes a “File Map” pointing the node IDs at physical flash addresses.</p><p>One big advantage is that the DRAM mapping metadata overhead is considerably smaller: it uses the same amount of DRAM as a standard SSD, ~0.1% of NAND capacity (about 1 GB per TB), which a standard SSD fills almost entirely with its mapping table. Here, most of that DRAM, ~95%, can act as a hot-path node cache.</p><p>One garbage collector also replaces two, avoiding what FADU calls the “double GC” of host F2FS, where file-system cleanup and the drive’s own garbage collection amplify each other. Write amplification is also 63% lower versus a host F2FS solution on a mixed sequential and random workload with FADU’s research prototype, the company said. FADU told <em>Tom's Hardware Premium</em> that the prototype runs synthetic fio workloads on NVMeVirt, an open-source software SSD emulator from Seoul National University, so it is not a hardware result.</p><p>Most importantly for this method, it ensures host-free access. The GPU builds NFS-over-NVMe commands in queues in its own memory, which could include HBM. Alternatively, a remote SmartNIC or DPU with NVMe-oF target offload can drop commands in the drive’s submission queue. The data goes from SSD to NIC via PCIe peer-to-peer DMA, coming out as RDMA writes. FADU’s explicitly named use cases for this include GPU file reads and checkpoints, DPU-served HTTP content delivery, and a JBoF (just a bunch of flash) file server.</p><p>FADU’s numbers for its internal test come from a client and server Intel Core i5-13600K, a CPU with 20 threads that clocks up to 5.1 GHz with 128GB of DDR5, and a Mellanox ConnectX-6 100 Gbps NIC. Synthetic testing is compared using fio with the baseline F2FS and Ext4 on block firmware versus the FS-SSD via a kernel client and an SPDK user-space client. The sequential test is 1 thread, queue depth (QD) 128, with 128KB blocks, while random is at 4 threads, the same QD128, but with 4KB blocks.</p><p>These block sizes are typical for client workloads in our experience, so they are a fair measure, with detailed preliminary results in the table below. Against SPDK on the raw drive, figures FADU provided to us, FS-SSD’s SPDK client keeps about 93% on 4K random read and 53% on 4K random write, via fio, with around a 7% and 47% cost for the in-drive file layer. Though the raw drive’s 13,240 MB/s random write is above the ~12 GB/s FADU gives as the NAND’s program bandwidth, those runs likely weren’t at steady state. The Filebench result is particularly interesting, as the SPDK random gains apply to a plain block device too.</p><p>This should be the workload most favorable for the technology, as the Filebench file server is metadata-heavy with relatively high host overhead: FADU says it uses 10,000 files in a multi-level tree, mean directory width 20, and a mean file size of 128KB with 50 client threads.</p><div ><table><caption>FADU’s preliminary fio results: Gen6 controller in a PCIe 5.0 host, direct I/O</caption><thead><tr><th class="firstcol " ><p>Setup</p></th><th  ><p>Sequential read (MB/s)</p></th><th  ><p>Sequential write (MB/s)</p></th><th  ><p>4KB random read (MB/s)</p></th><th  ><p>4KB random write (MB/s)</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>F2FS</p></td><td  ><p>14,120</p></td><td  ><p>4,510</p></td><td  ><p>7,850</p></td><td  ><p>3,980</p></td></tr><tr><td class="firstcol " ><p>Ext4</p></td><td  ><p>14,130</p></td><td  ><p>7,050</p></td><td  ><p>9,120</p></td><td  ><p>4,560</p></td></tr><tr><td class="firstcol " ><p>FS-SSD (kernel client)</p></td><td  ><p>14,140</p></td><td  ><p>11,940</p></td><td  ><p>8,070</p></td><td  ><p>4,320</p></td></tr><tr><td class="firstcol " ><p>FS-SSD (SPDK client)</p></td><td  ><p>Not reported</p></td><td  ><p>Not reported</p></td><td  ><p>13,000</p></td><td  ><p>7,030</p></td></tr><tr><td class="firstcol " ><p>Raw block (SPDK)</p></td><td  ><p>Not reported</p></td><td  ><p>Not reported</p></td><td  ><p>13,990</p></td><td  ><p>13,240</p></td></tr></tbody></table></div><p>One important thing to recall is that although this test is on a <a href="https://www.tomshardware.com/tech-industry/the-current-state-of-pcie-6-0-ssds-and-controllers-marvell-phison-and-smi-prepare-controllers-as-drives-finally-come-to-market-following-years-of-delays">Gen6 </a>drive, the test results are on a PCIe 5.0 host. Therefore, every sequential read config tops out at around 14.1 GB/s, which is roughly the link limit. The testbed also uses a consumer-class desktop CPU, while the one-core NFS server is a showcase example. The kernel client ends up 11.5% behind Ext4 on random read and 5.3% behind on random write, again by our math.</p><p>One example of a tough question more generally is: how do concurrent GPU, DPU, and host writers coordinate? How does a kernel mount’s page cache stay coherent with GPU access? NFSv3 storage devices, by the pNFS Flexible File Layout spec (RFC 8435), are “always treated as loosely coupled,” which means “Mandatory byte-range locking cannot be supported.”</p><p>Credentials riding in the NVMe command may also face security issues, as loosely coupled devices get only “a level of security equivalent to AUTH_SYS,” NFS’s basic scheme, in which the server trusts whatever user and group IDs the client sends. FADU’s published material also does not describe features such as snapshots, encryption handling, quotas, ACLs, crash handling, and so forth, related to this. Statelessness is what makes it work, but it doesn’t solve all problems, so the trust factor moves elsewhere.</p><p>In reply to our questioning, FADU said that there is no cross-host locking, as NFS’s lock manager isn’t implemented. The drive applies each write whole, in order. If two writes hit the same spot, the second one wins; this is “last-writer-wins.” Coordination depends on the clients: kernel-mounted ones get the usual advisory locks among themselves, GPU and SPDK clients coordinate in the application, and at scale a pNFS metadata server handles the coordination. Caching is also handled like ordinary NFS, with kernel clients, or is not done at all, for GPU and SPDK clients. This is stock NFSv3 behavior.</p><p>The drive has to check user/group IDs against the file permissions itself. Because those IDs ride inside the command, a client could fake them, but FADU says it closes that issue with capability-based authentication. Each command carries a capability issued by an outside authentication server, plus a code computed with a key shared only between that server and the drive. This is not detailed in its published material. FADU adds that support for the Kerberos authentication protocol is planned. For networked drives, the standard NVMe-oF protections are used, with only pre-approved hosts admitted. A drive trusts a single host’s kernel like any NVMe drive.</p><p>However, as FADU described to us, ACLs and other features are missing in the current release, aside from encryption, which in this case is handled by the FC6161’s standard TCG Opal self-encrypting drive support. Power loss is also typically handled with capacitors flushing metadata, the write buffer, and a log of in-flight operations. On restart, the drive replays the unfinished operations, and every write is durable if the drive acknowledges it.</p><p>FADU’s longer sights are on producing a JBoF storage solution with no server in front: FS-SSDs, a PCIe switch, and a DPU or NIC for handling NVMe-oF in hardware. FADU’s model puts this at about 42% lower power, about 57% lower cost per 16-drive server, and higher rack density with significantly lower five-year total cost of ownership. The drives would slot in behind existing pNFS metadata servers and could scale. FADU’s controllers already ship with drives such as the SN861, which Western Digital launched and SanDisk now sells, but this would expand the options and solutions it can provide.</p><div ><table><caption>FADU’s FS-SSD roadmap</caption><thead><tr><th class="firstcol " ><p>Milestone</p></th><th  ><p>Timing (per FADU)</p></th><th  ><p>Details</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Gen6-host results</p></td><td  ><p>By Q1 2027</p></td><td  ><p>FS-SSD tested in a PCIe 6.0 host; hardware write-amplification figures to come with them</p></td></tr><tr><td class="firstcol " ><p>Partner samples</p></td><td  ><p>Q3 2027</p></td><td  ><p>OEM/ODM talks are at an early stage; active collaboration expected once samples ship</p></td></tr><tr><td class="firstcol " ><p>Hostless JBoF</p></td><td  ><p>By end of 2027</p></td><td  ><p>Reference design built on Xsight Labs E1 DPUs, delivered to OEM/ODM partners</p></td></tr><tr><td class="firstcol " ><p>GPU access</p></td><td  ><p>Now, via GPUDirect Storage</p></td><td  ><p>Works with GDS through the kernel client; FADU intends to join the Accelerated IO SIG (xio-sig), where Nvidia open-sourced cuFile; no direct engagement with Nvidia</p></td></tr><tr><td class="firstcol " ><p>Standards</p></td><td  ><p>Nothing submitted yet</p></td><td  ><p>FADU plans to start a file-semantic command set discussion at NVM Express and OCP Storage; samples and client code come first</p></td></tr></tbody></table></div><p>Next year should bring real numbers and samples, with the JBoF reference design also expected. There may be an NVM Express or OCP Storage proposal down the line for a file-semantic command set to further cement and mature the approach. FADU has a good idea, is building it on an existing framework, and has a testbed with results and a hardware path forward on its Gen6 controller, with a trust model in mind. A Gen6-host run is now needed to support the company’s claims.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ TP-Link proposes $1 billion US Wi-Fi router manufacturing alliance with Key Tronic ]]></title>
                                                                                                <dc:content><![CDATA[ <p>TP-Link is fully determined to get in the good graces of the U.S. government to break free from a total ban on newly developed routers manufactured outside of the country. Although TP-Link is incorporated in the U.S. with a headquarters in Irvine, California, the company has been under pressure from the full weight of the U.S. government because of its historical Chinese roots. However, in an effort to appeal to the Trump administration’s zeal for creating jobs and touting American manufacturing, TP-Link this week announced that it has signed a letter of intent to assemble routers in the U.S. <a href="https://www.tp-link.com/us/press/news/22580/">through a crucial partnership with Key Tronic</a>.</p><h2 id="how-did-we-get-here">How did we get here?</h2><p>The ongoing saga started back in late 2024, when the House Select Committee on the Chinese Communist Party requested that the <a href="https://www.tomshardware.com/networking/routers/tp-link-investigated-by-us-government-over-national-security-concerns-investigation-probes-tp-link-routers-used-in-recent-cyberattacks">Department of Commerce investigate TP-Link</a>. At the time, the committee warned that “TP-Link products are also found on U.S. military bases, with the Army and Air Force Exchange and the Navy Exchange selling these devices to members of the military and their families,” and claimed TP-Link routers were sending user data back to China.</p><p>The select committee implored the Department of Commerce to use its broad powers to investigate TP-Link and take any necessary action to mitigate threats to national security. The campaign against TP-Link ramped up in mid-2025, with 17 members of Congress petitioning the Commerce Department to ban the sale of the company’s hardware in the U.S. due to alleged deep ties to the CCP. In <a href="https://www.tomshardware.com/networking/routers/tp-link-routers-face-potential-u-s-ban-over-alleged-china-related-national-security-concerns-company-vigorously-disputes-department-of-commerces-findings">late 2025</a>, 23 U.S. House members and Senators again asked the Commerce Department to investigate TP-Link’s actions, including adding the company to the Covered List, investigating any threats to national security, and informing the public about potential risks associated with using the company’s devices.</p><p>The situation reached a boiling point in March 2026, when TP-Link and other router manufacturers were added to the Federal Communications Commission (FCC) Covered List, forbidding the sale of foreign-made routers in the U.S. that hadn’t already been approved. At the time, the FCC noted, “Recently, malicious state and non-state sponsored cyber attackers have increasingly leveraged the vulnerabilities in small and home office routers produced abroad to carry out direct attacks against American civilians in their homes. From disrupting network connectivity to enabling local networking espionage and intellectual property theft, foreign-produced routers present unacceptable risks to Americans.”</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1920px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="7nbfGJHEoYDUnQHzfCLKzS" name="PR_Archer_8_Ultra_Lifestyle_20260831203428q" alt="TP-Link Router Render" src="https://cdn.mos.cms.futurecdn.net/7nbfGJHEoYDUnQHzfCLKzS-1920-80.jpg" mos="" align="middle" fullscreen="" width="1920" height="1080" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="caption-text">TP-Link Archer 8 Ultra </span><span class="credit" itemprop="copyrightHolder">(Image credit: TP-Link)</span></figcaption></figure><p>Since that time, several things have changed, with none of them being positive developments for TP-Link. Over the past six months, <a href="https://www.tomshardware.com/networking/routers/amazon-eero-and-leo-routers-gain-fcc-conditional-approval-for-us-sales-eero-products-can-skirt-router-ban-for-the-next-18-months-firm-joins-netgear-on-approval-list">Amazon</a>, <a href="https://www.tomshardware.com/networking/routers/asus-routers-gain-fccs-conditional-approval-for-sale-in-the-us-as-tp-link-remains-locked-out-asuss-wi-fi-8-ambitions-remain-intact#xenforo-comments-3900280">Asus</a>, and <a href="https://www.tomshardware.com/networking/routers/netgear-secures-conditional-approval-from-the-fcc-following-router-ban-company-can-continue-importing-foreign-made-routers-through-october-2027">Netgear</a> have all gained conditional approvals from the FCC to sell new wireless products in the U.S., while TP-Link has been stonewalled (conditional approvals are valid for 18 months from the date of issuance). Considering that TP-Link has been the world’s largest provider of Wi-Fi routers and associated devices for over a decade, losing access to the U.S. market for its new line of Wi-Fi 8 devices is a significant blow. This was readily apparent in late September when TP-Link opened preorders for its <a href="https://www.tomshardware.com/networking/routers/tp-link-opens-global-preorders-for-its-first-wi-fi-8-router-but-ban-keeps-it-out-of-us-market-us-left-off-launch-list-as-fcc-freeze-refuses-to-thaw">Archer 8 Ultra Wi-Fi 8 router</a>. The company announced that the router would be available in the United Kingdom, Germany, and Australia. However, the U.S. was conspicuously absent from that list because of the ongoing ban.</p><p>Adding insult to injury, earlier this month, multiple states (including <a href="https://www.myfloridalegal.com/sites/default/files/258584264-complaint.pdf">Florida</a>, <a href="https://dojmt.gov/attorney-general-knudsen-files-lawsuit-against-china-founded-tech-company-tp-link/">Montana</a>, <a href="https://www.iowaattorneygeneral.gov/newsroom/attorney-general-brenna-bird-sues-tp-link-systems-for-violating-iowa-consumer-fraud-act-and-allowing">Iowa</a>, and <a href="https://ago.nebraska.gov/sites/default/files/doc/Complaint_11.pdf">Nebraska</a>) sued TP-Link Systems, alleging that the company has made false and misleading claims about the security capabilities of its products and continued ties to China.</p><p>“TP-Link’s conduct has placed Montana consumers at risk of harm by making their home and small-office networks vulnerable to exploitation and intrusion by foreign actors with a proven track record of nefarious conduct. That risk is ongoing: many consumers keep their routers for years, and many of the compromised TP-Link models in Montana homes do not support automatic firmware updates and no longer receive security support at all,” stated Montana Attorney General Austin Knudsen. </p><p>Knudsen further claimed that while TP-Link touts that it employs 300 people in the U.S., the bulk of its workforce (numbering over 11,000 employees) conducts development and manufacturing operations in China. “Further, TP-Link claims to manufacture its U.S.-market products in Vietnam, but less than one percent of the components by value are sourced in Vietnam, and the rest are imported from or through China,” Knudsen added.</p><p>“Iowans’ sensitive data and our national security is at risk because of TP-Link and their connection to the communist Chinese government,” stated Iowa Attorney General Bird. “TP-Link tells Iowans its routers are safe, our personal data is secure, and that they have no ties to China. They are not telling the truth. It’s time to hold China and China-backed companies accountable.” </p><h2 id="tp-link-barks-back">TP-Link barks back</h2><p>Up until this point, TP-Link has vehemently denied any ties to the Chinese government and has refuted similar allegations from one of its primary competitors, <a href="https://www.tomshardware.com/networking/routers/netgear-countersues-tp-link-alleging-its-american-company-rebrand-is-false-advertising">Netgear</a>. However, on October 13, the company issued its <a href="https://www.fcc.gov/ecfs/document/26110081141/1">sharpest rebuttal yet in a letter to the FCC</a>. TP-Link touted its “America-first” strategy and chastised the state lawsuits brought earlier this month. </p><p>“A group of 21 state-level officials recently raised a scattershot barrage of baseless and previously debunked allegations relating to TP-Link’s confidential, non-public application,” TP-Link exclaimed. Rather than seeking to advance legitimate public policy objectives, the letter appears aimed primarily at encouraging federal officials to apply a heightened standard to TP-Link’s pending application that would be unattainable to TP-Link or any other router manufacturer in the world.” </p><p>TP-Link stated that its innovation in the router space earned its position as the market leader in the U.S., and that the constant barrage of legal fights and war of words is undermining fair competition and is an actual threat to national security and critical infrastructure given the broad proliferation of its products.</p><p>To that end, TP-Link reiterated that it has officially filed for conditional approval from the FCC, so that it can specifically sell two next-generation Wi-Fi 8 routers in the U.S. The company noted, “As the United States positions itself as a leader in wireless technology, a Conditional Approval would guarantee American consumers access to the same cutting-edge wireless equipment that will soon be available in other parts of the world."</p><h2 id="tp-link-proposes-1-billion-investment-in-the-u-s">TP-Link proposes $1 billion investment in the U.S. </h2><p>With its back against the ropes, TP-Link is appealing to the U.S. government in the best way it knows how: with money and jobs. If there’s anything that the U.S. government likes, it’s investment in manufacturing on its own soil and high-paying jobs to bolster local economies. To help push the U.S. government along in approving a conditional approval to sell Wi-Fi 8 hardware, TP-Link has signed a letter of intent with Key Tronic to manufacture routers in the U.S. instead of importing them from Vietnam. More specifically, the routers would be assembled in Corinth, Mississippi.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1920px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="xijyu9bT9sqrgoeJHNXn8a" name="PR_Deco_8_Ultra_1920x1080_20260831203545z" alt="TP-Link Deco Router lifestyle shot" src="https://cdn.mos.cms.futurecdn.net/xijyu9bT9sqrgoeJHNXn8a-1920-80.jpg" mos="" align="middle" fullscreen="" width="1920" height="1080" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="caption-text">TP Link Deco 8 Ultra </span><span class="credit" itemprop="copyrightHolder">(Image credit: TP-Link)</span></figcaption></figure><p>“This LOI with Key Tronic demonstrates our firm commitment to expanding domestic advanced manufacturing, and establishing America as our base for delivering secure, reliable connectivity solutions to households and businesses around the world,” said TP-Link Systems Inc. CEO Jeffrey Chao. “By choosing to assemble cutting-edge products here at home, we are thrilled to support U.S. high-tech job growth and empower a new generation of American innovation and innovators.”</p><p>TP-Link hopes that the investment will cut through most of the red tape that is currently preventing it from selling its latest products stateside. However, it remains to be seen if these latest sweeteners will be enough to sway President Trump and the agencies under his purview. As the company notes, it was the first to deliver Wi-Fi 7 routers in the U.S., and it wants to continue that leadership position by offering the first Wi-Fi 8 routers. The hardware is ready to go, as evidenced by the opening of preorders; it’s just currently being held up by roadblocks at the federal level, along with added pressure from state governments.</p><p>Until that time, Wi-Fi 8 products like the <a href="https://www.tomshardware.com/networking/routers/tp-link-announces-its-first-consumer-wi-fi-8-routers-archer-8-ultra-preorder-commences-september-30-in-select-regions">Archer 8 Ultra</a> and Deco 8 Ultra will remain off-limits to Americans, which have given TP-Link a reported <a href="https://www.tomshardware.com/networking/routers/tp-link-investigated-by-us-government-over-national-security-concerns-investigation-probes-tp-link-routers-used-in-recent-cyberattacks">65 percent share</a> of the U.S. consumer router market.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/networking/routers/tp-link-proposes-usd1-billion-us-wi-fi-router-manufacturing-alliance-with-key-tronic-potential-deal-seeks-to-end-ongoing-saga-blocking-its-wi-fi-8-products-from-american-homes</link>
                                                                            <description>
                            <![CDATA[ TP-Link is at the mercy of the FCC regarding its ready-to-go Wi-Fi 8 routers, so it has a sweetener in mind to move things along. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">H26THHZ6EjFUQHCpVhGbqC</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/9Bq9nzHgXdzzrrkpzKYi3T-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Tue, 20 Oct 2026 16:25:55 +0000</pubDate>                                                                                                                                <updated>Tue, 20 Oct 2026 16:27:26 +0000</updated>
                                                                                                                                            <category><![CDATA[Wi-Fi Routers]]></category>
                                                    <category><![CDATA[Networking]]></category>
                                                                                                <author><![CDATA[ brandon.hill@futurenet.com (Brandon Hill) ]]></author>                    <dc:creator><![CDATA[ Brandon Hill ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/yHeufe7JcvuJBhYPkSexNf-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Brandon has been tinkering with PCs since childhood and received his first &quot;real&quot; PC, an IBM Aptiva 310, in the mid-1990s. He next went on to build his first custom PC with an Intel Celeron 300A processor overclocked to 450MHz on an Abit BH6 motherboard. Brandon has written about PC and Mac tech since the late 1990s, first at AnandTech before moving to DailyTech and later to Hot Hardware. When Brandon is not consuming copious amounts of tech news, he can be found enjoying the NC mountains or the beach with his wife and two sons.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/9Bq9nzHgXdzzrrkpzKYi3T-1920-80.jpg">
                                                            <media:credit><![CDATA[TP-Link]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[TP-Link Router Render]]></media:description>                                                            <media:text><![CDATA[TP-Link Router Render]]></media:text>
                                <media:title type="plain"><![CDATA[TP-Link Router Render]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/9Bq9nzHgXdzzrrkpzKYi3T-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>TP-Link is fully determined to get in the good graces of the U.S. government to break free from a total ban on newly developed routers manufactured outside of the country. Although TP-Link is incorporated in the U.S. with a headquarters in Irvine, California, the company has been under pressure from the full weight of the U.S. government because of its historical Chinese roots. However, in an effort to appeal to the Trump administration’s zeal for creating jobs and touting American manufacturing, TP-Link this week announced that it has signed a letter of intent to assemble routers in the U.S. <a href="https://www.tp-link.com/us/press/news/22580/">through a crucial partnership with Key Tronic</a>.</p><h2 id="how-did-we-get-here">How did we get here?</h2><p>The ongoing saga started back in late 2024, when the House Select Committee on the Chinese Communist Party requested that the <a href="https://www.tomshardware.com/networking/routers/tp-link-investigated-by-us-government-over-national-security-concerns-investigation-probes-tp-link-routers-used-in-recent-cyberattacks">Department of Commerce investigate TP-Link</a>. At the time, the committee warned that “TP-Link products are also found on U.S. military bases, with the Army and Air Force Exchange and the Navy Exchange selling these devices to members of the military and their families,” and claimed TP-Link routers were sending user data back to China.</p><p>The select committee implored the Department of Commerce to use its broad powers to investigate TP-Link and take any necessary action to mitigate threats to national security. The campaign against TP-Link ramped up in mid-2025, with 17 members of Congress petitioning the Commerce Department to ban the sale of the company’s hardware in the U.S. due to alleged deep ties to the CCP. In <a href="https://www.tomshardware.com/networking/routers/tp-link-routers-face-potential-u-s-ban-over-alleged-china-related-national-security-concerns-company-vigorously-disputes-department-of-commerces-findings">late 2025</a>, 23 U.S. House members and Senators again asked the Commerce Department to investigate TP-Link’s actions, including adding the company to the Covered List, investigating any threats to national security, and informing the public about potential risks associated with using the company’s devices.</p><p>The situation reached a boiling point in March 2026, when TP-Link and other router manufacturers were added to the Federal Communications Commission (FCC) Covered List, forbidding the sale of foreign-made routers in the U.S. that hadn’t already been approved. At the time, the FCC noted, “Recently, malicious state and non-state sponsored cyber attackers have increasingly leveraged the vulnerabilities in small and home office routers produced abroad to carry out direct attacks against American civilians in their homes. From disrupting network connectivity to enabling local networking espionage and intellectual property theft, foreign-produced routers present unacceptable risks to Americans.”</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1920px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="7nbfGJHEoYDUnQHzfCLKzS" name="PR_Archer_8_Ultra_Lifestyle_20260831203428q" alt="TP-Link Router Render" src="https://cdn.mos.cms.futurecdn.net/7nbfGJHEoYDUnQHzfCLKzS-1920-80.jpg" mos="" align="middle" fullscreen="" width="1920" height="1080" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="caption-text">TP-Link Archer 8 Ultra </span><span class="credit" itemprop="copyrightHolder">(Image credit: TP-Link)</span></figcaption></figure><p>Since that time, several things have changed, with none of them being positive developments for TP-Link. Over the past six months, <a href="https://www.tomshardware.com/networking/routers/amazon-eero-and-leo-routers-gain-fcc-conditional-approval-for-us-sales-eero-products-can-skirt-router-ban-for-the-next-18-months-firm-joins-netgear-on-approval-list">Amazon</a>, <a href="https://www.tomshardware.com/networking/routers/asus-routers-gain-fccs-conditional-approval-for-sale-in-the-us-as-tp-link-remains-locked-out-asuss-wi-fi-8-ambitions-remain-intact#xenforo-comments-3900280">Asus</a>, and <a href="https://www.tomshardware.com/networking/routers/netgear-secures-conditional-approval-from-the-fcc-following-router-ban-company-can-continue-importing-foreign-made-routers-through-october-2027">Netgear</a> have all gained conditional approvals from the FCC to sell new wireless products in the U.S., while TP-Link has been stonewalled (conditional approvals are valid for 18 months from the date of issuance). Considering that TP-Link has been the world’s largest provider of Wi-Fi routers and associated devices for over a decade, losing access to the U.S. market for its new line of Wi-Fi 8 devices is a significant blow. This was readily apparent in late September when TP-Link opened preorders for its <a href="https://www.tomshardware.com/networking/routers/tp-link-opens-global-preorders-for-its-first-wi-fi-8-router-but-ban-keeps-it-out-of-us-market-us-left-off-launch-list-as-fcc-freeze-refuses-to-thaw">Archer 8 Ultra Wi-Fi 8 router</a>. The company announced that the router would be available in the United Kingdom, Germany, and Australia. However, the U.S. was conspicuously absent from that list because of the ongoing ban.</p><p>Adding insult to injury, earlier this month, multiple states (including <a href="https://www.myfloridalegal.com/sites/default/files/258584264-complaint.pdf">Florida</a>, <a href="https://dojmt.gov/attorney-general-knudsen-files-lawsuit-against-china-founded-tech-company-tp-link/">Montana</a>, <a href="https://www.iowaattorneygeneral.gov/newsroom/attorney-general-brenna-bird-sues-tp-link-systems-for-violating-iowa-consumer-fraud-act-and-allowing">Iowa</a>, and <a href="https://ago.nebraska.gov/sites/default/files/doc/Complaint_11.pdf">Nebraska</a>) sued TP-Link Systems, alleging that the company has made false and misleading claims about the security capabilities of its products and continued ties to China.</p><p>“TP-Link’s conduct has placed Montana consumers at risk of harm by making their home and small-office networks vulnerable to exploitation and intrusion by foreign actors with a proven track record of nefarious conduct. That risk is ongoing: many consumers keep their routers for years, and many of the compromised TP-Link models in Montana homes do not support automatic firmware updates and no longer receive security support at all,” stated Montana Attorney General Austin Knudsen. </p><p>Knudsen further claimed that while TP-Link touts that it employs 300 people in the U.S., the bulk of its workforce (numbering over 11,000 employees) conducts development and manufacturing operations in China. “Further, TP-Link claims to manufacture its U.S.-market products in Vietnam, but less than one percent of the components by value are sourced in Vietnam, and the rest are imported from or through China,” Knudsen added.</p><p>“Iowans’ sensitive data and our national security is at risk because of TP-Link and their connection to the communist Chinese government,” stated Iowa Attorney General Bird. “TP-Link tells Iowans its routers are safe, our personal data is secure, and that they have no ties to China. They are not telling the truth. It’s time to hold China and China-backed companies accountable.” </p><h2 id="tp-link-barks-back">TP-Link barks back</h2><p>Up until this point, TP-Link has vehemently denied any ties to the Chinese government and has refuted similar allegations from one of its primary competitors, <a href="https://www.tomshardware.com/networking/routers/netgear-countersues-tp-link-alleging-its-american-company-rebrand-is-false-advertising">Netgear</a>. However, on October 13, the company issued its <a href="https://www.fcc.gov/ecfs/document/26110081141/1">sharpest rebuttal yet in a letter to the FCC</a>. TP-Link touted its “America-first” strategy and chastised the state lawsuits brought earlier this month. </p><p>“A group of 21 state-level officials recently raised a scattershot barrage of baseless and previously debunked allegations relating to TP-Link’s confidential, non-public application,” TP-Link exclaimed. Rather than seeking to advance legitimate public policy objectives, the letter appears aimed primarily at encouraging federal officials to apply a heightened standard to TP-Link’s pending application that would be unattainable to TP-Link or any other router manufacturer in the world.” </p><p>TP-Link stated that its innovation in the router space earned its position as the market leader in the U.S., and that the constant barrage of legal fights and war of words is undermining fair competition and is an actual threat to national security and critical infrastructure given the broad proliferation of its products.</p><p>To that end, TP-Link reiterated that it has officially filed for conditional approval from the FCC, so that it can specifically sell two next-generation Wi-Fi 8 routers in the U.S. The company noted, “As the United States positions itself as a leader in wireless technology, a Conditional Approval would guarantee American consumers access to the same cutting-edge wireless equipment that will soon be available in other parts of the world."</p><h2 id="tp-link-proposes-1-billion-investment-in-the-u-s">TP-Link proposes $1 billion investment in the U.S. </h2><p>With its back against the ropes, TP-Link is appealing to the U.S. government in the best way it knows how: with money and jobs. If there’s anything that the U.S. government likes, it’s investment in manufacturing on its own soil and high-paying jobs to bolster local economies. To help push the U.S. government along in approving a conditional approval to sell Wi-Fi 8 hardware, TP-Link has signed a letter of intent with Key Tronic to manufacture routers in the U.S. instead of importing them from Vietnam. More specifically, the routers would be assembled in Corinth, Mississippi.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1920px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="xijyu9bT9sqrgoeJHNXn8a" name="PR_Deco_8_Ultra_1920x1080_20260831203545z" alt="TP-Link Deco Router lifestyle shot" src="https://cdn.mos.cms.futurecdn.net/xijyu9bT9sqrgoeJHNXn8a-1920-80.jpg" mos="" align="middle" fullscreen="" width="1920" height="1080" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="caption-text">TP Link Deco 8 Ultra </span><span class="credit" itemprop="copyrightHolder">(Image credit: TP-Link)</span></figcaption></figure><p>“This LOI with Key Tronic demonstrates our firm commitment to expanding domestic advanced manufacturing, and establishing America as our base for delivering secure, reliable connectivity solutions to households and businesses around the world,” said TP-Link Systems Inc. CEO Jeffrey Chao. “By choosing to assemble cutting-edge products here at home, we are thrilled to support U.S. high-tech job growth and empower a new generation of American innovation and innovators.”</p><p>TP-Link hopes that the investment will cut through most of the red tape that is currently preventing it from selling its latest products stateside. However, it remains to be seen if these latest sweeteners will be enough to sway President Trump and the agencies under his purview. As the company notes, it was the first to deliver Wi-Fi 7 routers in the U.S., and it wants to continue that leadership position by offering the first Wi-Fi 8 routers. The hardware is ready to go, as evidenced by the opening of preorders; it’s just currently being held up by roadblocks at the federal level, along with added pressure from state governments.</p><p>Until that time, Wi-Fi 8 products like the <a href="https://www.tomshardware.com/networking/routers/tp-link-announces-its-first-consumer-wi-fi-8-routers-archer-8-ultra-preorder-commences-september-30-in-select-regions">Archer 8 Ultra</a> and Deco 8 Ultra will remain off-limits to Americans, which have given TP-Link a reported <a href="https://www.tomshardware.com/networking/routers/tp-link-investigated-by-us-government-over-national-security-concerns-investigation-probes-tp-link-routers-used-in-recent-cyberattacks">65 percent share</a> of the U.S. consumer router market.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Googlebooks are as poorly timed as Valve’s Steam Machine, and may suffer a similar fate ]]></title>
                                                                                                <dc:content><![CDATA[ <p>It’s been a busy year for laptops. Apple’s value-priced <a href="https://www.tomshardware.com/laptops/macbooks/apple-macbook-neo-a18-pro-review"><u>MacBook Neo</u></a> caused a major shakeup in the budget space, forcing Windows device makers to respond with their own <a href="https://www.tomshardware.com/laptops/dell-xps-13-2026-review"><u>premium</u></a>, <a href="https://www.tomshardware.com/laptops/acer-swift-air-14-review"><u>affordable offerings</u></a>. Microsoft nixed its <a href="https://www.tomshardware.com/software/windows/microsofts-baseline-ram-for-ai-pcs-set-at-16gb"><u>original 16GB requirement</u></a> for AI PCs, ushering in a slew of <a href="https://www.tomshardware.com/laptops/8gb-of-ram-is-back-on-laptops-companies-are-lowering-memory-offerings-to-make-affordable-notebooks-during-component-crisis"><u>new 8GB Wildcat Lake options</u></a>, and further muddied<a href="https://www.tomshardware.com/tablets/microsoft-surface/microsoft-quietly-drops-copilot-branding-from-its-new-laptops-surface-cvp-confirms-new-devices-meet-hardware-requirements-but-lack-controversial-branding"><u> the waters of its Copilot branding</u></a> after vowing to <a href="https://www.tomshardware.com/software/windows/microsoft-vows-to-make-windows-11-fly-on-8gb-ram-amid-memory-shortage-optimizations-to-reduce-os-memory-footprint-have-begun"><u>improve performance on 8GB Windows PCs</u></a>.</p><p>The driving force behind all of this is, of course, is the roughly <a href="https://www.tomshardware.com/pc-components/ram/memory-prices-climb-500-percent-in-12-months-up-to-10x-the-lowest-ever-tracked-prices-128gb-of-ddr5-now-usd3-399"><u>five-fold increase in RAM costs</u></a> in the past year, driven by AI hyperscaler demand. That’s also why <a href="https://www.tomshardware.com/video-games/console-gaming/valve-delays-steam-machine-and-says-it-is-reconsidering-pricing-critical-component-shortage-and-costs-behind-the-move"><u>Valve delayed the Steam Machine</u></a> by roughly six months, eventually forcing its launch price above $1,000. That completely negated the value proposition that made the Steam Deck a success for so long. And I think Google’s new laptops are in a similarly tough spot, for many of the same reasons.</p><p>The new <a href="https://www.tomshardware.com/laptops/googlebooks-from-dell-lenovo-and-hp-tested-nice-hardware-but-the-os-isnt-ready-yet">Googlebook laptop platform</a> – an AI-focused reimagining of its successful Chromebook platform, complete with its own 16GB RAM requirement and prices that (apart from a limited-time $899 offer from Acer) start at $999. Models from Lenovo and Asus even stretch into MacBook Air pricing territory.</p><p>But, like Valve’s gaming hardware, Google’s previous laptop success has hinged primarily on affordability. Chromebooks launched in 2011, with three models all priced around $500, and Acer again undercut its competition at $349. There have been “high-end” Chromebooks over the years, but the vast majority have remained fairly affordable.</p><h2 id="cutting-edge-android-17-cutoff">Cutting-edge Android 17 cutoff</h2><p>Android phone integration (akin to an iPhone / MacBook pairing) is one of the major new selling points of the Googlebook platform. But you need a phone running Android 17 for that integration to work. That OS iteration launched just four months ago and, so far, is only available on a handful of expensive phones. </p><p>Aside from severely limiting the number of current customers, this also ignores the fundamental realities of the Android phone market. The majority of Android devices sold every year are mainstream and budget models. Even with the same RAM and storage pressures pushing prices up, <a href="https://www.androidauthority.com/flagships-better-deals-mid-rangers-3695801/"><u>less than 30% of global phone shipments are models priced higher than $600</u></a> — and that number includes iPhones, all of which are priced above that mark, unless you’re buying from a reseller.</p><p>So in short, Google is launching a new, mostly $1,000-plus (at least for now) laptop platform, hoping to piggyback off the brand recognition and mindshare of its search engine, cheap Chromebooks, and (largely) cheap Android phones. This is at a time when the price of almost everything else has also been on the rise for years, and both Apple and Microsoft’s Windows partners are offering up the best sub-$700 laptops… arguably ever.</p><p>It’s hard not to see this as poor timing, at best. Given software and hardware lead times, Googlebooks as a concept were almost certainly pretty far along a year ago when RAM pricing started to skyrocket. But with solid indications that the cost of RAM and storage <a href="https://www.tomshardware.com/pc-components/ram/lexar-regional-manager-says-that-ram-prices-are-expected-to-double-by-the-end-of-the-year-discounts-and-stabilized-prices-result-from-distributors-getting-rid-of-old-stock-or-sourcing-products-from-other-regions"><u>could continue to climb</u></a> into next year and beyond, Google may have effectively been forced into an “if not now, when” launch window, which also seems to have been the case with the Steam Machine launch earlier this year. </p><h2 id="nvidia-s-rtx-spark-is-a-completely-different-laptop-animal">Nvidia’s RTX Spark is a completely different laptop animal</h2><p>Of course, Nvidia’s also brand-new Spark laptops are priced much higher, starting at $2,599 and going <a href="https://www.microsoft.com/en-us/store/configure/surface-laptop-ultra/8wg5vqp0x4b8"><u>up to $5,899</u></a> or even <a href="https://eshop.asus.com/us/31150-us-cnf-auto.html"><u>$6,999</u></a> for 128GB models. But those Nvidia-silicon laptops are aimed primarily at AI coders, not average Android-toting consumers who might also be interested in AI features. And, in case you somehow haven’t noticed, Nvidia has become one of the largest companies in the world by selling ever-pricier GPUs to gamers, workstation users, and the global AI hyperscaler market. Nvidia may be <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-reportedly-halts-geforce-rtx-5090-production-in-favor-of-ai-data-center-and-professional-gpus-impending-supply-drought-expected-to-drive-up-prices-rtx-5080-24gb-rumored-as-new-gaming-flagship"><u>halting RTX 5090 production</u></a> in favor of more profitable AI hardware, despite that gaming GPU continuing to sell out while being consistently priced thousands of dollars above its MSRP. Early indications point to similar demand for at least some Spark laptops, with preorders of Microsoft’s 128GB Surface Laptop Ultra seemingly sold out before the platform even officially launches. </p><p>Google’s hardware division, meanwhile, has spent the last 15 years selling its mostly budget laptop platform, while repeatedly attempting and failing to make premium Google-designed devices a mainstream success. Yes, there was the Chromebook Plus rollout in 2023, but those devices were mostly priced below $700, and multiple retailer list filters suggest the sub-$500 models are the most popular.  Even Google’s flagship Pixel phones are priced well below flagships from Samsung and Apple. </p><p>This is a strange time to go upmarket, to require 16GB of RAM, and to launch a new laptop OS and platform at all, considering <a href="https://www.tomshardware.com/tech-industry/pc-shipments-tumble-over-20-percent-in-3q26-as-chip-shortages-bite-top-three-pc-vendors-ship-11-6-million-fewer-units-year-over-year"><u>PC shipments are down 20%</u></a>. And that’s especially true if you’re Google, a company known primarily for free, ad-supported software and services, and hardware that you primarily buy because it’s what you can afford, not because it’s what you want. </p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/laptops/googlebooks-are-as-poorly-timed-as-valves-steam-machine-and-may-suffer-a-similar-fate</link>
                                                                            <description>
                            <![CDATA[ Google attempted to take its Chromebook platform upscale with Googlebooks. But the timing could not have been worse. Phone support at launch is extremely limited, and most people use Android phones and Chromebooks because they’re cheap. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">wWmGHCf9t7eBRDB7N2JJna</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/d9FqdWcJbbHePRhCjmgCPA-1920-80.png" type="image/png" length="0"></enclosure>
                                                                        <pubDate>Fri, 16 Oct 2026 12:30:00 +0000</pubDate>                                                                                                                                <updated>Fri, 16 Oct 2026 15:07:21 +0000</updated>
                                                                                                                                            <category><![CDATA[Laptops]]></category>
                                                                                                                    <dc:creator><![CDATA[ Matt Safford ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/uW75KiUF9FVG2vFdwJzeZh-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Matt began piling up computer experience as a child with his Mattel Aquarius. He built his first PC in the late 1990s and ventured into mild PC modding in the early 2000s. He’s spent 15 years covering emerging technology for Smithsonian, Popular Science, and Consumer Reports, while testing components and PCs for Computer Shopper, PCMag and Digital Trends. When not writing about tech, he’s often walking—through the streets of New York, over the sheep-dotted hills of Scotland, or just at his treadmill desk at home in front of the 50-inch HDR TV that serves as his PC monitor.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/png" url="https://cdn.mos.cms.futurecdn.net/d9FqdWcJbbHePRhCjmgCPA-1920-80.png">
                                                            <media:credit><![CDATA[Tom&amp;#39;s Hardware]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Googlebook next to a phone]]></media:description>                                                            <media:text><![CDATA[Googlebook next to a phone]]></media:text>
                                <media:title type="plain"><![CDATA[Googlebook next to a phone]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/d9FqdWcJbbHePRhCjmgCPA-1920-80.png" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>It’s been a busy year for laptops. Apple’s value-priced <a href="https://www.tomshardware.com/laptops/macbooks/apple-macbook-neo-a18-pro-review"><u>MacBook Neo</u></a> caused a major shakeup in the budget space, forcing Windows device makers to respond with their own <a href="https://www.tomshardware.com/laptops/dell-xps-13-2026-review"><u>premium</u></a>, <a href="https://www.tomshardware.com/laptops/acer-swift-air-14-review"><u>affordable offerings</u></a>. Microsoft nixed its <a href="https://www.tomshardware.com/software/windows/microsofts-baseline-ram-for-ai-pcs-set-at-16gb"><u>original 16GB requirement</u></a> for AI PCs, ushering in a slew of <a href="https://www.tomshardware.com/laptops/8gb-of-ram-is-back-on-laptops-companies-are-lowering-memory-offerings-to-make-affordable-notebooks-during-component-crisis"><u>new 8GB Wildcat Lake options</u></a>, and further muddied<a href="https://www.tomshardware.com/tablets/microsoft-surface/microsoft-quietly-drops-copilot-branding-from-its-new-laptops-surface-cvp-confirms-new-devices-meet-hardware-requirements-but-lack-controversial-branding"><u> the waters of its Copilot branding</u></a> after vowing to <a href="https://www.tomshardware.com/software/windows/microsoft-vows-to-make-windows-11-fly-on-8gb-ram-amid-memory-shortage-optimizations-to-reduce-os-memory-footprint-have-begun"><u>improve performance on 8GB Windows PCs</u></a>.</p><p>The driving force behind all of this is, of course, is the roughly <a href="https://www.tomshardware.com/pc-components/ram/memory-prices-climb-500-percent-in-12-months-up-to-10x-the-lowest-ever-tracked-prices-128gb-of-ddr5-now-usd3-399"><u>five-fold increase in RAM costs</u></a> in the past year, driven by AI hyperscaler demand. That’s also why <a href="https://www.tomshardware.com/video-games/console-gaming/valve-delays-steam-machine-and-says-it-is-reconsidering-pricing-critical-component-shortage-and-costs-behind-the-move"><u>Valve delayed the Steam Machine</u></a> by roughly six months, eventually forcing its launch price above $1,000. That completely negated the value proposition that made the Steam Deck a success for so long. And I think Google’s new laptops are in a similarly tough spot, for many of the same reasons.</p><p>The new <a href="https://www.tomshardware.com/laptops/googlebooks-from-dell-lenovo-and-hp-tested-nice-hardware-but-the-os-isnt-ready-yet">Googlebook laptop platform</a> – an AI-focused reimagining of its successful Chromebook platform, complete with its own 16GB RAM requirement and prices that (apart from a limited-time $899 offer from Acer) start at $999. Models from Lenovo and Asus even stretch into MacBook Air pricing territory.</p><p>But, like Valve’s gaming hardware, Google’s previous laptop success has hinged primarily on affordability. Chromebooks launched in 2011, with three models all priced around $500, and Acer again undercut its competition at $349. There have been “high-end” Chromebooks over the years, but the vast majority have remained fairly affordable.</p><h2 id="cutting-edge-android-17-cutoff">Cutting-edge Android 17 cutoff</h2><p>Android phone integration (akin to an iPhone / MacBook pairing) is one of the major new selling points of the Googlebook platform. But you need a phone running Android 17 for that integration to work. That OS iteration launched just four months ago and, so far, is only available on a handful of expensive phones. </p><p>Aside from severely limiting the number of current customers, this also ignores the fundamental realities of the Android phone market. The majority of Android devices sold every year are mainstream and budget models. Even with the same RAM and storage pressures pushing prices up, <a href="https://www.androidauthority.com/flagships-better-deals-mid-rangers-3695801/"><u>less than 30% of global phone shipments are models priced higher than $600</u></a> — and that number includes iPhones, all of which are priced above that mark, unless you’re buying from a reseller.</p><p>So in short, Google is launching a new, mostly $1,000-plus (at least for now) laptop platform, hoping to piggyback off the brand recognition and mindshare of its search engine, cheap Chromebooks, and (largely) cheap Android phones. This is at a time when the price of almost everything else has also been on the rise for years, and both Apple and Microsoft’s Windows partners are offering up the best sub-$700 laptops… arguably ever.</p><p>It’s hard not to see this as poor timing, at best. Given software and hardware lead times, Googlebooks as a concept were almost certainly pretty far along a year ago when RAM pricing started to skyrocket. But with solid indications that the cost of RAM and storage <a href="https://www.tomshardware.com/pc-components/ram/lexar-regional-manager-says-that-ram-prices-are-expected-to-double-by-the-end-of-the-year-discounts-and-stabilized-prices-result-from-distributors-getting-rid-of-old-stock-or-sourcing-products-from-other-regions"><u>could continue to climb</u></a> into next year and beyond, Google may have effectively been forced into an “if not now, when” launch window, which also seems to have been the case with the Steam Machine launch earlier this year. </p><h2 id="nvidia-s-rtx-spark-is-a-completely-different-laptop-animal">Nvidia’s RTX Spark is a completely different laptop animal</h2><p>Of course, Nvidia’s also brand-new Spark laptops are priced much higher, starting at $2,599 and going <a href="https://www.microsoft.com/en-us/store/configure/surface-laptop-ultra/8wg5vqp0x4b8"><u>up to $5,899</u></a> or even <a href="https://eshop.asus.com/us/31150-us-cnf-auto.html"><u>$6,999</u></a> for 128GB models. But those Nvidia-silicon laptops are aimed primarily at AI coders, not average Android-toting consumers who might also be interested in AI features. And, in case you somehow haven’t noticed, Nvidia has become one of the largest companies in the world by selling ever-pricier GPUs to gamers, workstation users, and the global AI hyperscaler market. Nvidia may be <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-reportedly-halts-geforce-rtx-5090-production-in-favor-of-ai-data-center-and-professional-gpus-impending-supply-drought-expected-to-drive-up-prices-rtx-5080-24gb-rumored-as-new-gaming-flagship"><u>halting RTX 5090 production</u></a> in favor of more profitable AI hardware, despite that gaming GPU continuing to sell out while being consistently priced thousands of dollars above its MSRP. Early indications point to similar demand for at least some Spark laptops, with preorders of Microsoft’s 128GB Surface Laptop Ultra seemingly sold out before the platform even officially launches. </p><p>Google’s hardware division, meanwhile, has spent the last 15 years selling its mostly budget laptop platform, while repeatedly attempting and failing to make premium Google-designed devices a mainstream success. Yes, there was the Chromebook Plus rollout in 2023, but those devices were mostly priced below $700, and multiple retailer list filters suggest the sub-$500 models are the most popular.  Even Google’s flagship Pixel phones are priced well below flagships from Samsung and Apple. </p><p>This is a strange time to go upmarket, to require 16GB of RAM, and to launch a new laptop OS and platform at all, considering <a href="https://www.tomshardware.com/tech-industry/pc-shipments-tumble-over-20-percent-in-3q26-as-chip-shortages-bite-top-three-pc-vendors-ship-11-6-million-fewer-units-year-over-year"><u>PC shipments are down 20%</u></a>. And that’s especially true if you’re Google, a company known primarily for free, ad-supported software and services, and hardware that you primarily buy because it’s what you can afford, not because it’s what you want. </p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Ukrainian drones hit Russian data centers, disrupting Russian apps and services  ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Over the past week, Ukraine <a href="https://www.tomshardware.com/tech-industry/data-centers/russian-disinformation-bot-traffic-nearly-eliminated-overnight-after-ukraine-drone-strikes-cripple-multiple-of-russias-google-data-centers-ukraine-intensifies-bombing-of-yandex-data-centers-in-response-to-moscows-strikes-on-kyiv-data-centers">attacked</a> some of Yandex's biggest data centers, disrupting the operations of various Russian national applications and metropolitan digital services. Russia, which has itself <a href="https://www.tomshardware.com/tech-industry/data-centers/russia-bombs-ukrainian-data-centers-in-latest-escalation-100-000-households-lose-connectivity-as-firms-migrate-data-abroad-zelensky-says-ordinary-life-is-simply-a-target" target="_blank">targeted Ukrainian data centers in recent weeks</a>. </p><p>Worse yet for Russia, its <a href="https://www.tomshardware.com/tech-industry/data-centers/russian-disinformation-bot-traffic-nearly-eliminated-overnight-after-ukraine-drone-strikes-cripple-multiple-of-russias-google-data-centers-ukraine-intensifies-bombing-of-yandex-data-centers-in-response-to-moscows-strikes-on-kyiv-data-centers" target="_blank">online disinformation activities appear disrupted</a> by these attacks. Although not as catastrophic as some initial reports claimed, a new report from <a href="https://www.tomshardware.com/service-providers/network-providers/malicious-russian-disinformation-bot-traffic-barely-flinches-after-ukraine-strikes-yandex-data-centers-and-auto-migration-kicks-in-cloudflare-tracks-170-million-automated-attacks-shifting-to-alternate-cloud-hosts">Cloudflare </a>claims that the bot traffic merely migrated to other hosts.</p><h2 id="a-new-target">A new target</h2><p>Since Russia's full-scale invasion of Ukraine in 2022, the war has moved through several phases. From the pushback of Russian forces in 2022, to the rapid deployment of FPV drones in 2023, to the mid-range drone strike campaigns both sides have focused on since 2025. Over that time, the targets have changed. Ukrainian supporters published "bingo cards," with various refineries and warehouses listed, often superimposing fire graphics to showcase when they'd been damaged or destroyed.</p><p>There are tens of major targets across different industries and networks; there are only a handful of Yandex data centers. Although <a href="https://www.datacentermap.com/russia/" target="_blank">there is a more expansive list of data centers</a> from other providers that support the Yandex cloud network, most are small, only offering a few megawatts of compute capacity.</p><p>Considering that targeting just three of Yandex's facilities brought several key digital services such as shopping and taxi apps to their knees in major Russian cities, it may not take much for Ukraine to severely hamper the modern digital elements of Russia's economy and society at large. </p><p>While Ukraine isn't immune to similar attempts at disruption, its already-hardened networks and expansive list of allies make it far more resilient. </p><h2 id="friends-in-the-right-places">Friends in the right places</h2><p>While overreliance on digital services from American companies like Google, Meta, and Amazon has its own drawbacks, in Ukraine's case, it's been a far more resilient strategy than Russia's. Russian attacks on Ukrainian data centers have caused some disruption, but much of Ukraine's digital infrastructure runs on Amazon's AWS cloud platform. That includes <a href="https://aws.amazon.com/blogs/publicsector/ukraines-digital-transformation/" target="_blank">distributed hosting, resilient service structure, and AI integration</a>.</p><p>That means that when Russia hits Ukrainian data centers, the services can be quickly migrated to another data center, perhaps located in another country entirely. Amazon has the world to fall back on if necessary. Ukraine also has allies in many other technologically advanced nations that can provide digital infrastructure and expertise, including access to the latest AI cybersecurity models.</p><p>In contrast, Russian internet and services are far more inward-facing. It has attempted to cut its population off from the outside world by banning Western services like WhatsApp and Telegram, pushing its citizens towards government-run applications <a href="https://en.wikipedia.org/wiki/Max_(app)" target="_blank">like Max</a>. </p><p>As the Russian-version of Google, too, Yandex has its fingers in a similar number of digital pies. It runs online search, mapping, online shopping, cloud storage, ridesharing, food ordering, and more. When its systems are disrupted, services that people have come to rely on in a modern digital economy break down, and there are far fewer alternative options.</p><p>Russian allies like North Korea, Belarus, and Iran can't offer much in the way of digital infrastructure support. Even China, which has been aiding Russia throughout the conflict, can't help much here. While it is making strides in building out its compute network for AI infrastructure, it runs a relatively walled-garden form of the internet. Integrating Russian applications with that would be complex and time-consuming.</p><p>All of this suggests that if Ukraine continues to hit Russian data centers, there's a real potential that Russian citizens and its military could lose access to digital services that have become ubiquitous in 21st-century life. </p><p>The effect that would have on Russian society and its war-fighting goals remains unclear, but it would have an impact.</p><h2 id="russia-39-s-ai-capacity">Russia's AI capacity </h2><p>Losing its most capable data centers alongside sanctions that prevent it from taking advantage of the AI boom that's buoying many other global economies could severely hamper Russian efforts to modernize its economy.</p><p>While the U.S., EU, and China are building out hundreds of billions of dollars worth of data centers to power the next generation of artificial intelligence models, Russia is struggling to hold on to what it's got. With major data centers now in Ukraine's sights, including those scheduled for expansion, Russia could see its digital future wiped away in a matter of months.</p><p>With even medium-tier data centers often taking years to build, even with ready access to the hardware to power them, Russia could be on the cusp of falling decades behind its contemporaries.</p><h2 id="bots-are-down-but-not-that-much">Bots are Down, But Not That Much</h2><p>One highly publicized impact of these strikes on Yandex data centers has been a reduction in automated online activity from Russia. While that <a href="https://radar.cloudflare.com/bots/ru" target="_blank">does appear to have reduced bot traffic slightly</a>, particularly around support for far-right parties, it may not be as impactful as some initial reports claim, as Cloudflare has published a report stating that many of the bots have merely migrated to other hosts, reducing, but not eliminating, the bots themselves.</p><p>Journalist Ilya Lozovsky posted a lengthy explanation on Twitter/X about why Russian bots haven't been as disrupted as some have claimed, highlighting that bot activity hasn't fallen off as much as legitimate web crawlers and search functions at Yandex. </p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2109732026756563054"><p lang="en" dir="ltr">Last night I was struck — really, almost radicalized — watching in real time as experts at the heights of their careers boosted emotionally satisfying narratives that were pretty much entirely wrong. https://t.co/8WoMQvt4xh<a href="https://twitter.com/cantworkitout/status/2109732026756563054">October 12, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><p>But while Russian bots haven't gone dark, there are several reports of reduced online activity among far-right groups, particularly those supporting prominent far-right politicians.</p><p>Posts supporting the AFD in Germany don't appear as impacted as initial claims suggested — the chart doing the rounds has clear errors in it, and German tech outlet <a href="https://rohtext.de/news/afd-bots-offline-yandex-angriffen-was-virale" target="_blank"><em>Rohtext</em> claims</a> online discourse in general went down, rather than specific parties being impacted.</p><p>What these strikes have done, though, is highlight a clear and present weakness in Russia's modern society, military, and economy. If Ukraine can continue its strikes against Russian data centers, it may well bring key public services to their knees, as well as cripple Russia's ability to compete in modern digital services beyond the end of this war, further inhibiting the Russian economy in the future.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/data-centers/ukrainian-drones-hit-russian-data-centers-disrupting-russian-apps-and-services-attacks-highlight-the-fragility-of-the-countrys-digital-services</link>
                                                                            <description>
                            <![CDATA[ After Ukrainian long-range strike drones targeted Russian data centers, various Russian digital services and apps went down or faced severe disruption. There's even some suggestion that online bot activity may be down, but not by as much as initial reports claimed. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">wHtKgGTjireSgtPJshPjsd</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/nDHvMHdcfdeqRTeB2xJ8fD-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Fri, 16 Oct 2026 10:15:00 +0000</pubDate>                                                                                                                                <updated>Fri, 16 Oct 2026 12:15:40 +0000</updated>
                                                                                                                                            <category><![CDATA[Data Centers]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Jon Martindale ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/YeutDv8zJmhi7xH35MSt8Z-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;After building his first computers in his teens, Jon Martindale has spent the past two decades covering the latest advances in technology. From displays to PC components, blockchain to AI, and tablets to standing desk accessories, Jon has covered just about every facet of the tech space in his varied career. He has bylines at Forbes, USNews, Lifewire, DigitalTrends, PCWorld, and a range of other sites. He brings that same level of expertise and professional insight to Toms Hardware.Away from writing, Jon is an avid reader, board gamer, and fitness enthusiast. He lives in rural Gloucestershire with his wife, two children, and French Bulldog cross.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/nDHvMHdcfdeqRTeB2xJ8fD-1920-80.jpg">
                                                            <media:credit><![CDATA[Pavel Byrkin via Getty Images]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Yandex offices in Moscow.]]></media:description>                                                            <media:text><![CDATA[Yandex offices in Moscow.]]></media:text>
                                <media:title type="plain"><![CDATA[Yandex offices in Moscow.]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/nDHvMHdcfdeqRTeB2xJ8fD-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Over the past week, Ukraine <a href="https://www.tomshardware.com/tech-industry/data-centers/russian-disinformation-bot-traffic-nearly-eliminated-overnight-after-ukraine-drone-strikes-cripple-multiple-of-russias-google-data-centers-ukraine-intensifies-bombing-of-yandex-data-centers-in-response-to-moscows-strikes-on-kyiv-data-centers">attacked</a> some of Yandex's biggest data centers, disrupting the operations of various Russian national applications and metropolitan digital services. Russia, which has itself <a href="https://www.tomshardware.com/tech-industry/data-centers/russia-bombs-ukrainian-data-centers-in-latest-escalation-100-000-households-lose-connectivity-as-firms-migrate-data-abroad-zelensky-says-ordinary-life-is-simply-a-target" target="_blank">targeted Ukrainian data centers in recent weeks</a>. </p><p>Worse yet for Russia, its <a href="https://www.tomshardware.com/tech-industry/data-centers/russian-disinformation-bot-traffic-nearly-eliminated-overnight-after-ukraine-drone-strikes-cripple-multiple-of-russias-google-data-centers-ukraine-intensifies-bombing-of-yandex-data-centers-in-response-to-moscows-strikes-on-kyiv-data-centers" target="_blank">online disinformation activities appear disrupted</a> by these attacks. Although not as catastrophic as some initial reports claimed, a new report from <a href="https://www.tomshardware.com/service-providers/network-providers/malicious-russian-disinformation-bot-traffic-barely-flinches-after-ukraine-strikes-yandex-data-centers-and-auto-migration-kicks-in-cloudflare-tracks-170-million-automated-attacks-shifting-to-alternate-cloud-hosts">Cloudflare </a>claims that the bot traffic merely migrated to other hosts.</p><h2 id="a-new-target">A new target</h2><p>Since Russia's full-scale invasion of Ukraine in 2022, the war has moved through several phases. From the pushback of Russian forces in 2022, to the rapid deployment of FPV drones in 2023, to the mid-range drone strike campaigns both sides have focused on since 2025. Over that time, the targets have changed. Ukrainian supporters published "bingo cards," with various refineries and warehouses listed, often superimposing fire graphics to showcase when they'd been damaged or destroyed.</p><p>There are tens of major targets across different industries and networks; there are only a handful of Yandex data centers. Although <a href="https://www.datacentermap.com/russia/" target="_blank">there is a more expansive list of data centers</a> from other providers that support the Yandex cloud network, most are small, only offering a few megawatts of compute capacity.</p><p>Considering that targeting just three of Yandex's facilities brought several key digital services such as shopping and taxi apps to their knees in major Russian cities, it may not take much for Ukraine to severely hamper the modern digital elements of Russia's economy and society at large. </p><p>While Ukraine isn't immune to similar attempts at disruption, its already-hardened networks and expansive list of allies make it far more resilient. </p><h2 id="friends-in-the-right-places">Friends in the right places</h2><p>While overreliance on digital services from American companies like Google, Meta, and Amazon has its own drawbacks, in Ukraine's case, it's been a far more resilient strategy than Russia's. Russian attacks on Ukrainian data centers have caused some disruption, but much of Ukraine's digital infrastructure runs on Amazon's AWS cloud platform. That includes <a href="https://aws.amazon.com/blogs/publicsector/ukraines-digital-transformation/" target="_blank">distributed hosting, resilient service structure, and AI integration</a>.</p><p>That means that when Russia hits Ukrainian data centers, the services can be quickly migrated to another data center, perhaps located in another country entirely. Amazon has the world to fall back on if necessary. Ukraine also has allies in many other technologically advanced nations that can provide digital infrastructure and expertise, including access to the latest AI cybersecurity models.</p><p>In contrast, Russian internet and services are far more inward-facing. It has attempted to cut its population off from the outside world by banning Western services like WhatsApp and Telegram, pushing its citizens towards government-run applications <a href="https://en.wikipedia.org/wiki/Max_(app)" target="_blank">like Max</a>. </p><p>As the Russian-version of Google, too, Yandex has its fingers in a similar number of digital pies. It runs online search, mapping, online shopping, cloud storage, ridesharing, food ordering, and more. When its systems are disrupted, services that people have come to rely on in a modern digital economy break down, and there are far fewer alternative options.</p><p>Russian allies like North Korea, Belarus, and Iran can't offer much in the way of digital infrastructure support. Even China, which has been aiding Russia throughout the conflict, can't help much here. While it is making strides in building out its compute network for AI infrastructure, it runs a relatively walled-garden form of the internet. Integrating Russian applications with that would be complex and time-consuming.</p><p>All of this suggests that if Ukraine continues to hit Russian data centers, there's a real potential that Russian citizens and its military could lose access to digital services that have become ubiquitous in 21st-century life. </p><p>The effect that would have on Russian society and its war-fighting goals remains unclear, but it would have an impact.</p><h2 id="russia-39-s-ai-capacity">Russia's AI capacity </h2><p>Losing its most capable data centers alongside sanctions that prevent it from taking advantage of the AI boom that's buoying many other global economies could severely hamper Russian efforts to modernize its economy.</p><p>While the U.S., EU, and China are building out hundreds of billions of dollars worth of data centers to power the next generation of artificial intelligence models, Russia is struggling to hold on to what it's got. With major data centers now in Ukraine's sights, including those scheduled for expansion, Russia could see its digital future wiped away in a matter of months.</p><p>With even medium-tier data centers often taking years to build, even with ready access to the hardware to power them, Russia could be on the cusp of falling decades behind its contemporaries.</p><h2 id="bots-are-down-but-not-that-much">Bots are Down, But Not That Much</h2><p>One highly publicized impact of these strikes on Yandex data centers has been a reduction in automated online activity from Russia. While that <a href="https://radar.cloudflare.com/bots/ru" target="_blank">does appear to have reduced bot traffic slightly</a>, particularly around support for far-right parties, it may not be as impactful as some initial reports claim, as Cloudflare has published a report stating that many of the bots have merely migrated to other hosts, reducing, but not eliminating, the bots themselves.</p><p>Journalist Ilya Lozovsky posted a lengthy explanation on Twitter/X about why Russian bots haven't been as disrupted as some have claimed, highlighting that bot activity hasn't fallen off as much as legitimate web crawlers and search functions at Yandex. </p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2109732026756563054"><p lang="en" dir="ltr">Last night I was struck — really, almost radicalized — watching in real time as experts at the heights of their careers boosted emotionally satisfying narratives that were pretty much entirely wrong. https://t.co/8WoMQvt4xh<a href="https://twitter.com/cantworkitout/status/2109732026756563054">October 12, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><p>But while Russian bots haven't gone dark, there are several reports of reduced online activity among far-right groups, particularly those supporting prominent far-right politicians.</p><p>Posts supporting the AFD in Germany don't appear as impacted as initial claims suggested — the chart doing the rounds has clear errors in it, and German tech outlet <a href="https://rohtext.de/news/afd-bots-offline-yandex-angriffen-was-virale" target="_blank"><em>Rohtext</em> claims</a> online discourse in general went down, rather than specific parties being impacted.</p><p>What these strikes have done, though, is highlight a clear and present weakness in Russia's modern society, military, and economy. If Ukraine can continue its strikes against Russian data centers, it may well bring key public services to their knees, as well as cripple Russia's ability to compete in modern digital services beyond the end of this war, further inhibiting the Russian economy in the future.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Sipeed reveals Chinese-made DGX Spark and Ryzen AI Halo rival, expected to cost $3,999 ]]></title>
                                                                                                <dc:content><![CDATA[ <p>A Chinese rival to the DGX Spark and the Ryzen AI Halo, at least on paper, has emerged with Sipeed’s announcement of the MaixStation. The system, built on Axera’s new AX8860P, is in the same memory class with 128GB of LPDDR5X on a 256-bit bus at 273 GB/s. Sipeed told <em>Tom’s Hardware Premium</em> it expects the 128GB model to cost around $3,999, with launch discounts on Kickstarter, putting it level with the <a href="https://www.tomshardware.com/pc-components/gpus/embargo-mon-july-6-8am-pt-1100-edt-amd-ryzen-ai-halo-review/4">128GB Ryzen AI Halo</a> and well under the DGX Spark’s current MSRP of $6,950. Sign-ups are open now on Sipeed’s <a href="https://sipeed.com/maixstation" target="_blank"><u>MaixStation page</u></a> ahead of a Kickstarter campaign at the end of the year, with units shipping within three months after it closes. The NPU toolchain is due by the end of this month.</p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2109891722125938708"><p lang="en" dir="ltr">Coming today! Beat DGX Spark!MaixStation soon～ https://t.co/e7DvWweVnt<a href="https://twitter.com/cantworkitout/status/2109891722125938708">October 13, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><p>The AX8860P is a 16-core Arm Cortex-A55 part with an in-house NPU rated at 256 TOPS at native FP4 with no sparsity. The chip is produced in the 6nm process, Axera says. Sipeed claims the system has 25% faster prefill and 30% faster decode than the DGX Spark when running Qwen 27B (the company did not disclose which specific model, but later details that it expects a single MaixStation to run Qwen 3.8-27B) from a single node. </p><p>In its email, though, the company put its MaixStation about 10% ahead on both prefill and decode when compared to DGX Spark, with FP4 on each machine, with speculative decoding off, and said newer runs with a wider lead aren’t official yet. It teased the new product on its X account, showing a photo from the AX8860P’s announcement at an industrial fair in Shanghai on Tuesday.</p><div ><table><caption>Sipeed MaixStation vs Nvidia DGX Spark vs AMD Ryzen AI Halo</caption><thead><tr><th class="firstcol empty" ></th><th  ><p>Sipeed MaixStation (Axera AX8860P)</p></th><th  ><p>Nvidia DGX Spark (GB10)</p></th><th  ><p>AMD Ryzen AI Halo (Ryzen AI Max+ PRO 495)</p></th><th  ><p>AMD Ryzen AI Halo (Ryzen AI Max+ 395)</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Status</p></td><td  ><p>Kickstarter at the end of 2026; specs are Sipeed’s claims</p></td><td  ><p>128GB on sale; 64GB due Oct. 23 through partners</p></td><td  ><p>On sale</p></td><td  ><p>On sale</p></td></tr><tr><td class="firstcol " ><p>CPU</p></td><td  ><p>16 Arm Cortex-A55 cores</p></td><td  ><p>20 Arm cores (10 Cortex-X925, 10 Cortex-A725)</p></td><td  ><p>16 Zen 5 cores</p></td><td  ><p>16 Zen 5 cores</p></td></tr><tr><td class="firstcol " ><p>Memory</p></td><td  ><p>128GB LPDDR5X; 512GB on MaixStation Pro</p></td><td  ><p>128GB LPDDR5X; 64GB through partners</p></td><td  ><p>192GB LPDDR5X-8533</p></td><td  ><p>128GB LPDDR5X-8000</p></td></tr><tr><td class="firstcol " ><p>Memory bandwidth</p></td><td  ><p>273 GB/s</p></td><td  ><p>273 GB/s</p></td><td  ><p>273 GB/s</p></td><td  ><p>256 GB/s</p></td></tr><tr><td class="firstcol " ><p>AI compute (rated peak)</p></td><td  ><p>NPU: 256 TOPS FP4, no sparsity</p></td><td  ><p>Blackwell GPU: up to 1 PFLOP FP4 with sparsity, about 500 TFLOPS dense</p></td><td  ><p>Radeon 8065S GPU (40 CUs): up to 61 TFLOPS FP16; NPU: up to 55 TOPS</p></td><td  ><p>Radeon 8060S GPU (40 CUs); NPU: up to 50 TOPS</p></td></tr><tr><td class="firstcol " ><p>Software</p></td><td  ><p>Debian 13; NPU toolchain due at the end of October; Sipeed claims vLLM, SGLang and PyTorch support</p></td><td  ><p>DGX OS (Ubuntu 24.04 LTS) with CUDA and Nvidia’s AI software stack preinstalled</p></td><td  ><p>ROCm, validated as a “Best-Known Configuration”; Linux now, Windows “coming soon”</p></td><td  ><p>Linux or Windows 11 Pro</p></td></tr><tr><td class="firstcol " ><p>Networking</p></td><td  ><p>10G SFP+; 256 Gbps C2C link (up to 384 Gbps), up to eight nodes without a switch, per Sipeed</p></td><td  ><p>10GbE; ConnectX-7 at 200 Gbps, up to four systems (four need a switch)</p></td><td  ><p>10GbE</p></td><td  ><p>10GbE</p></td></tr><tr><td class="firstcol " ><p>Price</p></td><td  ><p>About $3,999 expected (128GB), per Sipeed</p></td><td  ><p>$6,950 (128GB); 64GB from $4,999 through partners from Oct. 23</p></td><td  ><p>$7,299.99 (Linux)</p></td><td  ><p>$3,999.99</p></td></tr></tbody></table></div><p>For comparison, the DGX Spark also comes in 128GB at 273 GB/s with <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-introduces-64gb-dgx-spark-to-throw-local-ai-fans-a-lifeline-amid-the-rampocalypse-new-gb10-config-starts-at-usd4999-for-those-who-can-work-with-less"><u>another SKU at 64GB</u></a>, the two models priced at $6,950 and from $4,999. The former has been on sale since October 2025, at its current price since Oct. 2, and the latter is expected through partners on Oct. 23. For AMD, its Ryzen AI Halo has a <a href="https://www.tomshardware.com/pc-components/cpus/amd-ryzen-ai-max-400-gorgon-halo-packs-up-to-192gb-of-unified-memory-refreshed-apu-uses-zen-5-and-rdna-3-5-and-can-clock-up-to-5-2-ghz"><u>192GB Max+ PRO 495 model</u></a> at 273 GB/s, released earlier this week at $7,299.99 for the Linux version, and <a href="https://www.tomshardware.com/desktops/mini-pcs/amd-challenges-nvidias-dgx-spark-with-usd3-999-ryzen-ai-halo-with-windows-11-support-strix-halo-desktop-undercuts-nvidia-by-usd700-packs-128gb-of-unified-memory"><u>the 128GB Max+ 395 model</u></a> with 256 GB/s at $3,999.99. In all cases, 256-bit LPDDR5X is used, so the salient differences are memory capacity, compute, and software.</p><p>Memory is also a big share of what these systems cost: the best U.S. price for <a href="https://www.tomshardware.com/pc-components/ram/memory-prices-climb-500-percent-in-12-months-up-to-10x-the-lowest-ever-tracked-prices-128gb-of-ddr5-now-usd3-399"><u>a 128GB DDR5 kit was $3,399 in August</u></a>, and <em>TrendForce </em>expects conventional DRAM contract prices to rise another 10–15% this quarter. As for its unit prices, Sipeed says it used current DRAM prices rather than stock bought last year. It also suggested buyers could supply their own SSD, which could cut the price further, clarifying that the MaixStation takes a standard M.2 2280 drive rather than the Spark’s less common 2242.</p><p>Software-wise, the Spark ships with DGX OS, which is <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-dgx-spark-review"><u>Nvidia’s build of Ubuntu 24.04 LTS</u></a>, with CUDA and Nvidia’s AI software stack preinstalled. AMD validates ROCm, drivers, models, and frameworks together as a “Best-Known Configuration.” Neither the Spark nor the PRO 495 ships with Windows yet, though AMD says Windows versions are “coming soon,” and the Max+ 395 model already offers Windows 11 Pro. Sipeed, aside from its upcoming NPU toolchain, claims support for vLLM, SGLang, and PyTorch, with Axera adding ONNX, Hugging Face (HF) model formats, and an OpenAI-compatible API. MaixStation itself ships with Debian 13 and Linux 6.18, and Windows support for the PCIe card will depend on demand, Sipeed says.</p><p>Chinese solutions already exist on AMD silicon, such as <a href="https://www.tomshardware.com/desktops/mini-pcs/early-amd-gorgon-halo-ai-mini-pc-packs-192gb-ram-for-an-eye-watering-usd7-099-super-early-bird-deal-cuts-down-price-of-gmktec-evo-x5-with-ryzen-ai-max-pro-495-by-usd425"><u>GMKtec’s 192GB Evo-X5 Pro</u></a> at $6,799 to $7,099. There’s also Minisforum’s MS-S1 Max-P495. What’s new with the MaixStation is the Chinese silicon underneath.</p><p>The Axera AX8860P, as recently launched, has, in addition to the above specs, PCIe 5.0 support, 64-channel video decode with 16-channel encode, and support for multiple form factors, including a single-slot, four-chip PCIe card capable of running up to 512GB of memory, with mini and pocket PC products planned. Like the eight-core AX8850 in Radxa’s and M5Stack’s M.2 accelerator cards, it pairs Arm Cortex-A55 cores with Axera’s NPU, now with twice the cores and up to 128GB of memory instead of 8GB. Two- and four-chip cascades reach a theoretical 512 and 1,024 TOPS, respectively, at a claimed scaling efficiency of about 80%. The CPU, against the Spark’s Cortex-X925 and Halo’s Zen 5, is small and in-order. It’s sufficient for the NPU but potentially more limited for other tasks.</p><p>The box itself measures about 11 × 11 × 3 cm and draws about 40W in typical use, powered over USB-C, according to Sipeed. Its ports include 10G SFP+, USB-C, an M.2 2280 slot for a PCIe 5.0 SSD, and an MCIO 8i connector for linking a second unit.</p><p>Sipeed says the MaixStation supports all configurations, that’s single, dual, and quad node, plus any “PCIe card mode for any card count you want.” The interconnect in use is a 256 Gbps C2C connection (up to 384 Gbps) “without [an] extra switch needed, and [up to] 8-node.” </p><p>The contrast is with the Spark’s QSFP ports, and Nvidia’s datasheet says linking four Sparks “requires [an] external switch.” The math, at 32 Gbps per lane, works out to 256 Gbps with the default eight lanes and 384 Gbps with the maximum of twelve. Remote-access test units and a few review units will follow during the Kickstarter campaign, the company told us.</p><p>Sipeed also commented on the openness of the platform, saying, “Yes, bsp sdk is open, you can use AI to port any distribution on it!” Sipeed has previously sold Axera boards, with products dating back to 2022 using the AX620A, AX650N, and AX630C.</p><p>The claims of the MaixStation being faster in some workloads are coupled with Sipeed saying “In fact you need [to] ask why DGX is so slow…” with potential financial implications, as Nvidia “just [wants to make] more money from users.” When an X user argued that Sipeed must be using extreme quantization at that bandwidth, the company suggested that the Spark’s specs may be overstated.</p><p>By our math, the decode claim is plausible with a 4-bit 27B model, which sits at about 13.5GB. Performance on comparable machines tops out near 20 tokens per second, which means around 270 GB/s. Prefill, which is compute-bound, is a harder sell, but Sipeed stated that “DGX Spark 1000TOPS is fake ‘sparse.’” That 1 PFLOP figure does rely on sparsity, but the DGX Spark’s <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-dgx-spark-review/2"><u>dense rate of about 500 TFLOPS</u></a> is still around double the AX8860P’s 256 TOPS.</p><p>Axera’s current stack has Pulsar2 as the NPU compiler and toolchain, with version 8.0 on HuggingFace last week, and the V7.0 docs already list the chip as the AX8860. There’s no operator support list for the chip yet, and the docs’ LLM chapter for older chips lists weights down to 4-bit integers but no FP4. Neither the ax-llm runtime nor AXCL, the host-side PCIe driver and runtime for Axera’s cards, lists the chip yet. Sipeed says the end-of-month release is AXDSL, a kernel-programming language, plus the Pulsar2 source code, and that SGLang and vLLM run on a custom backend built on it, so Hugging Face models run without a conversion step.</p><p>Enthusiasts are already eager to know the final specifications and pricing for the different models. On paper, as far as local LLM machines go, the MaixStation matches the Spark on memory size and claimed bandwidth, and the Max+ PRO 495 Halo on bandwidth alone. Testing performance claims will also have to wait for the toolchain and sample machines land, hopefully before the end of the year.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/desktops/mini-pcs/sipeed-reveals-chinese-made-dgx-spark-and-ryzen-ai-halo-rival-expected-to-cost-usd3-999-maixstation-pairs-axeras-ax8860p-with-128gb-of-lpddr5x-at-273-gb-s-with-promising-local-ai-performance-claims</link>
                                                                            <description>
                            <![CDATA[ Sipeed’s MaixStation pairs Axera’s AX8860P with 128GB of LPDDR5X at 273 GB/s to take on Nvidia’s DGX Spark and AMD’s Ryzen AI Halo. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">R6i4E7yVVxK5zzdNCo8WN7</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/qJ8AzHF86C7AK4DgmNfWmf-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Thu, 15 Oct 2026 14:30:00 +0000</pubDate>                                                                                                                                <updated>Thu, 15 Oct 2026 15:23:49 +0000</updated>
                                                                                                                                            <category><![CDATA[Mini PCs]]></category>
                                                    <category><![CDATA[Desktops]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/qJ8AzHF86C7AK4DgmNfWmf-1920-80.jpg">
                                                            <media:credit><![CDATA[Sipeed]]></media:credit>
                                                                                                                                                                        <media:description><![CDATA[From left: Sipeed’s MaixStation, Apple’s Mac mini and Nvidia’s DGX Spark]]></media:description>                                                            <media:text><![CDATA[Sipeed MaixStation mini PC beside an Apple Mac mini and an Nvidia DGX Spark on a desk]]></media:text>
                                <media:title type="plain"><![CDATA[Sipeed MaixStation mini PC beside an Apple Mac mini and an Nvidia DGX Spark on a desk]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/qJ8AzHF86C7AK4DgmNfWmf-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>A Chinese rival to the DGX Spark and the Ryzen AI Halo, at least on paper, has emerged with Sipeed’s announcement of the MaixStation. The system, built on Axera’s new AX8860P, is in the same memory class with 128GB of LPDDR5X on a 256-bit bus at 273 GB/s. Sipeed told <em>Tom’s Hardware Premium</em> it expects the 128GB model to cost around $3,999, with launch discounts on Kickstarter, putting it level with the <a href="https://www.tomshardware.com/pc-components/gpus/embargo-mon-july-6-8am-pt-1100-edt-amd-ryzen-ai-halo-review/4">128GB Ryzen AI Halo</a> and well under the DGX Spark’s current MSRP of $6,950. Sign-ups are open now on Sipeed’s <a href="https://sipeed.com/maixstation" target="_blank"><u>MaixStation page</u></a> ahead of a Kickstarter campaign at the end of the year, with units shipping within three months after it closes. The NPU toolchain is due by the end of this month.</p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2109891722125938708"><p lang="en" dir="ltr">Coming today! Beat DGX Spark!MaixStation soon～ https://t.co/e7DvWweVnt<a href="https://twitter.com/cantworkitout/status/2109891722125938708">October 13, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><p>The AX8860P is a 16-core Arm Cortex-A55 part with an in-house NPU rated at 256 TOPS at native FP4 with no sparsity. The chip is produced in the 6nm process, Axera says. Sipeed claims the system has 25% faster prefill and 30% faster decode than the DGX Spark when running Qwen 27B (the company did not disclose which specific model, but later details that it expects a single MaixStation to run Qwen 3.8-27B) from a single node. </p><p>In its email, though, the company put its MaixStation about 10% ahead on both prefill and decode when compared to DGX Spark, with FP4 on each machine, with speculative decoding off, and said newer runs with a wider lead aren’t official yet. It teased the new product on its X account, showing a photo from the AX8860P’s announcement at an industrial fair in Shanghai on Tuesday.</p><div ><table><caption>Sipeed MaixStation vs Nvidia DGX Spark vs AMD Ryzen AI Halo</caption><thead><tr><th class="firstcol empty" ></th><th  ><p>Sipeed MaixStation (Axera AX8860P)</p></th><th  ><p>Nvidia DGX Spark (GB10)</p></th><th  ><p>AMD Ryzen AI Halo (Ryzen AI Max+ PRO 495)</p></th><th  ><p>AMD Ryzen AI Halo (Ryzen AI Max+ 395)</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Status</p></td><td  ><p>Kickstarter at the end of 2026; specs are Sipeed’s claims</p></td><td  ><p>128GB on sale; 64GB due Oct. 23 through partners</p></td><td  ><p>On sale</p></td><td  ><p>On sale</p></td></tr><tr><td class="firstcol " ><p>CPU</p></td><td  ><p>16 Arm Cortex-A55 cores</p></td><td  ><p>20 Arm cores (10 Cortex-X925, 10 Cortex-A725)</p></td><td  ><p>16 Zen 5 cores</p></td><td  ><p>16 Zen 5 cores</p></td></tr><tr><td class="firstcol " ><p>Memory</p></td><td  ><p>128GB LPDDR5X; 512GB on MaixStation Pro</p></td><td  ><p>128GB LPDDR5X; 64GB through partners</p></td><td  ><p>192GB LPDDR5X-8533</p></td><td  ><p>128GB LPDDR5X-8000</p></td></tr><tr><td class="firstcol " ><p>Memory bandwidth</p></td><td  ><p>273 GB/s</p></td><td  ><p>273 GB/s</p></td><td  ><p>273 GB/s</p></td><td  ><p>256 GB/s</p></td></tr><tr><td class="firstcol " ><p>AI compute (rated peak)</p></td><td  ><p>NPU: 256 TOPS FP4, no sparsity</p></td><td  ><p>Blackwell GPU: up to 1 PFLOP FP4 with sparsity, about 500 TFLOPS dense</p></td><td  ><p>Radeon 8065S GPU (40 CUs): up to 61 TFLOPS FP16; NPU: up to 55 TOPS</p></td><td  ><p>Radeon 8060S GPU (40 CUs); NPU: up to 50 TOPS</p></td></tr><tr><td class="firstcol " ><p>Software</p></td><td  ><p>Debian 13; NPU toolchain due at the end of October; Sipeed claims vLLM, SGLang and PyTorch support</p></td><td  ><p>DGX OS (Ubuntu 24.04 LTS) with CUDA and Nvidia’s AI software stack preinstalled</p></td><td  ><p>ROCm, validated as a “Best-Known Configuration”; Linux now, Windows “coming soon”</p></td><td  ><p>Linux or Windows 11 Pro</p></td></tr><tr><td class="firstcol " ><p>Networking</p></td><td  ><p>10G SFP+; 256 Gbps C2C link (up to 384 Gbps), up to eight nodes without a switch, per Sipeed</p></td><td  ><p>10GbE; ConnectX-7 at 200 Gbps, up to four systems (four need a switch)</p></td><td  ><p>10GbE</p></td><td  ><p>10GbE</p></td></tr><tr><td class="firstcol " ><p>Price</p></td><td  ><p>About $3,999 expected (128GB), per Sipeed</p></td><td  ><p>$6,950 (128GB); 64GB from $4,999 through partners from Oct. 23</p></td><td  ><p>$7,299.99 (Linux)</p></td><td  ><p>$3,999.99</p></td></tr></tbody></table></div><p>For comparison, the DGX Spark also comes in 128GB at 273 GB/s with <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-introduces-64gb-dgx-spark-to-throw-local-ai-fans-a-lifeline-amid-the-rampocalypse-new-gb10-config-starts-at-usd4999-for-those-who-can-work-with-less"><u>another SKU at 64GB</u></a>, the two models priced at $6,950 and from $4,999. The former has been on sale since October 2025, at its current price since Oct. 2, and the latter is expected through partners on Oct. 23. For AMD, its Ryzen AI Halo has a <a href="https://www.tomshardware.com/pc-components/cpus/amd-ryzen-ai-max-400-gorgon-halo-packs-up-to-192gb-of-unified-memory-refreshed-apu-uses-zen-5-and-rdna-3-5-and-can-clock-up-to-5-2-ghz"><u>192GB Max+ PRO 495 model</u></a> at 273 GB/s, released earlier this week at $7,299.99 for the Linux version, and <a href="https://www.tomshardware.com/desktops/mini-pcs/amd-challenges-nvidias-dgx-spark-with-usd3-999-ryzen-ai-halo-with-windows-11-support-strix-halo-desktop-undercuts-nvidia-by-usd700-packs-128gb-of-unified-memory"><u>the 128GB Max+ 395 model</u></a> with 256 GB/s at $3,999.99. In all cases, 256-bit LPDDR5X is used, so the salient differences are memory capacity, compute, and software.</p><p>Memory is also a big share of what these systems cost: the best U.S. price for <a href="https://www.tomshardware.com/pc-components/ram/memory-prices-climb-500-percent-in-12-months-up-to-10x-the-lowest-ever-tracked-prices-128gb-of-ddr5-now-usd3-399"><u>a 128GB DDR5 kit was $3,399 in August</u></a>, and <em>TrendForce </em>expects conventional DRAM contract prices to rise another 10–15% this quarter. As for its unit prices, Sipeed says it used current DRAM prices rather than stock bought last year. It also suggested buyers could supply their own SSD, which could cut the price further, clarifying that the MaixStation takes a standard M.2 2280 drive rather than the Spark’s less common 2242.</p><p>Software-wise, the Spark ships with DGX OS, which is <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-dgx-spark-review"><u>Nvidia’s build of Ubuntu 24.04 LTS</u></a>, with CUDA and Nvidia’s AI software stack preinstalled. AMD validates ROCm, drivers, models, and frameworks together as a “Best-Known Configuration.” Neither the Spark nor the PRO 495 ships with Windows yet, though AMD says Windows versions are “coming soon,” and the Max+ 395 model already offers Windows 11 Pro. Sipeed, aside from its upcoming NPU toolchain, claims support for vLLM, SGLang, and PyTorch, with Axera adding ONNX, Hugging Face (HF) model formats, and an OpenAI-compatible API. MaixStation itself ships with Debian 13 and Linux 6.18, and Windows support for the PCIe card will depend on demand, Sipeed says.</p><p>Chinese solutions already exist on AMD silicon, such as <a href="https://www.tomshardware.com/desktops/mini-pcs/early-amd-gorgon-halo-ai-mini-pc-packs-192gb-ram-for-an-eye-watering-usd7-099-super-early-bird-deal-cuts-down-price-of-gmktec-evo-x5-with-ryzen-ai-max-pro-495-by-usd425"><u>GMKtec’s 192GB Evo-X5 Pro</u></a> at $6,799 to $7,099. There’s also Minisforum’s MS-S1 Max-P495. What’s new with the MaixStation is the Chinese silicon underneath.</p><p>The Axera AX8860P, as recently launched, has, in addition to the above specs, PCIe 5.0 support, 64-channel video decode with 16-channel encode, and support for multiple form factors, including a single-slot, four-chip PCIe card capable of running up to 512GB of memory, with mini and pocket PC products planned. Like the eight-core AX8850 in Radxa’s and M5Stack’s M.2 accelerator cards, it pairs Arm Cortex-A55 cores with Axera’s NPU, now with twice the cores and up to 128GB of memory instead of 8GB. Two- and four-chip cascades reach a theoretical 512 and 1,024 TOPS, respectively, at a claimed scaling efficiency of about 80%. The CPU, against the Spark’s Cortex-X925 and Halo’s Zen 5, is small and in-order. It’s sufficient for the NPU but potentially more limited for other tasks.</p><p>The box itself measures about 11 × 11 × 3 cm and draws about 40W in typical use, powered over USB-C, according to Sipeed. Its ports include 10G SFP+, USB-C, an M.2 2280 slot for a PCIe 5.0 SSD, and an MCIO 8i connector for linking a second unit.</p><p>Sipeed says the MaixStation supports all configurations, that’s single, dual, and quad node, plus any “PCIe card mode for any card count you want.” The interconnect in use is a 256 Gbps C2C connection (up to 384 Gbps) “without [an] extra switch needed, and [up to] 8-node.” </p><p>The contrast is with the Spark’s QSFP ports, and Nvidia’s datasheet says linking four Sparks “requires [an] external switch.” The math, at 32 Gbps per lane, works out to 256 Gbps with the default eight lanes and 384 Gbps with the maximum of twelve. Remote-access test units and a few review units will follow during the Kickstarter campaign, the company told us.</p><p>Sipeed also commented on the openness of the platform, saying, “Yes, bsp sdk is open, you can use AI to port any distribution on it!” Sipeed has previously sold Axera boards, with products dating back to 2022 using the AX620A, AX650N, and AX630C.</p><p>The claims of the MaixStation being faster in some workloads are coupled with Sipeed saying “In fact you need [to] ask why DGX is so slow…” with potential financial implications, as Nvidia “just [wants to make] more money from users.” When an X user argued that Sipeed must be using extreme quantization at that bandwidth, the company suggested that the Spark’s specs may be overstated.</p><p>By our math, the decode claim is plausible with a 4-bit 27B model, which sits at about 13.5GB. Performance on comparable machines tops out near 20 tokens per second, which means around 270 GB/s. Prefill, which is compute-bound, is a harder sell, but Sipeed stated that “DGX Spark 1000TOPS is fake ‘sparse.’” That 1 PFLOP figure does rely on sparsity, but the DGX Spark’s <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-dgx-spark-review/2"><u>dense rate of about 500 TFLOPS</u></a> is still around double the AX8860P’s 256 TOPS.</p><p>Axera’s current stack has Pulsar2 as the NPU compiler and toolchain, with version 8.0 on HuggingFace last week, and the V7.0 docs already list the chip as the AX8860. There’s no operator support list for the chip yet, and the docs’ LLM chapter for older chips lists weights down to 4-bit integers but no FP4. Neither the ax-llm runtime nor AXCL, the host-side PCIe driver and runtime for Axera’s cards, lists the chip yet. Sipeed says the end-of-month release is AXDSL, a kernel-programming language, plus the Pulsar2 source code, and that SGLang and vLLM run on a custom backend built on it, so Hugging Face models run without a conversion step.</p><p>Enthusiasts are already eager to know the final specifications and pricing for the different models. On paper, as far as local LLM machines go, the MaixStation matches the Spark on memory size and claimed bandwidth, and the Max+ PRO 495 Halo on bandwidth alone. Testing performance claims will also have to wait for the toolchain and sample machines land, hopefully before the end of the year.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ ASML reports record $13 billion quarter, vows to expand capacity as demand for lithography equipment soars ]]></title>
                                                                                                <dc:content><![CDATA[ <p>ASML this week reported record quarterly revenue of $13.06 billion (€11.66 billion) and profitability of nearly $5.79 billion (€5 billion) as demand for advanced semiconductor manufacturing equipment continues to accelerate and investments in AI infrastructure continue to increase. The company now expects demand for its lithography systems to exceed its production capacity in 2027 and 2028, despite plans to increase output substantially. Some clients are already interested in placing orders for equipment as far ahead as 2030, which makes ASML consider further capacity expansions beyond its previously announced intentions.</p><h2 id="record-financial-results">Record financial results</h2><p>ASML reported third-quarter 2026 revenue of $13.06 billion USD (€11.66 billion), up 55% year-over-year and 25% sequentially. The company's gross margin reached a record 57.5%, up from 51.6% a year earlier, while operating income rose to $5.79 billion (€5.17 billion). Net income reached $5.56 billion USD (€4.96 billion) , although this figure includes approximately $570 million (€510 million) from the re-evaluation of the company's investment in <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/independent-tests-rank-mistrals-new-trillion-parameter-large-4-the-best-ai-model-outside-the-u-s-and-china-but-chinese-open-weights-still-overcome-europes-best-efforts">Mistral AI.</a></p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1486px;"><p class="vanilla-image-block" style="padding-top:69.04%;"><img id="j4wvMp9oob5zgNEmu4DLPL" name="Screenshot 2026-10-15 at 07.10.26" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/j4wvMp9oob5zgNEmu4DLPL-1920-80.png" mos="" align="middle" fullscreen="" width="1486" height="1026" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>During the quarter, ASML sold 121 lithography systems, including 113 new and eight refurbished machines, up from 91 systems in Q2 2026. The company generated $4.59 billion USD (€4.1 billion) from 17 EUV lithography systems and $4.82 billion USD (€4.3 billion) from 85 DUV (including 45 ArFi scanners) and 19 i-line litho tools. In addition, ASML recognized revenue from one High-NA EUV system in Q2. Of total equipment revenue, 54% came from logic chipmakers and 46% from memory manufacturers. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1914px;"><p class="vanilla-image-block" style="padding-top:53.40%;"><img id="EXXfjzu7eqfyC8Wf9t3fQL" name="Screenshot 2026-10-15 at 07.10.49" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/EXXfjzu7eqfyC8Wf9t3fQL-1920-80.png" mos="" align="middle" fullscreen="" width="1914" height="1022" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>Taiwan accounted for 32% of ASML's equipment sales in Q3 2026, followed closely by South Korea at 31%. China represented another 23%, while the United States accounted for 13% and the rest of Asia contributed just 1%. Combined, Taiwan and South Korea generated 63% of ASML's equipment revenue, which reflects substantial investments in advanced logic and memory manufacturing capacity that by far exceed those in the U.S. and Japan. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1806px;"><p class="vanilla-image-block" style="padding-top:55.59%;"><img id="RgpkxFX7qraArMPFKX9HPL" name="Screenshot 2026-10-15 at 07.10.53" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/RgpkxFX7qraArMPFKX9HPL-1920-80.png" mos="" align="middle" fullscreen="" width="1806" height="1004" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>ASML’s Installed Base Management business generated $3.68 billion (€3.29 billion), approximately $450 million (€400 million) above guidance, largely due to strong demand for productivity upgrades (more on this later), which contributed to the company’s record profitability.</p><p>For the fourth quarter, ASML expects revenue of $15.9 billion to $17.0 billion (€14.2 billion to €15.2 billion) and gross margin of approximately 57%. Full-year revenue should reach $49.3 billion to $50.4 billion (€44 billion to €45 billion), up approximately 36% YoY. The company projects EUV revenue should increase by more than 45% year-over-year, while memory-related equipment sales are expected to rise by approximately 75%, which indicates growing demand from DRAM producers.</p><h2 id="demand-for-litho-tools-spans-to-2030">Demand for litho tools spans to 2030</h2><p>One of the biggest learnings from these results and accompanying statements is that semiconductor manufacturers are increasingly interested in securing lithography equipment years before they actually need it. According to ASML CEO Christophe Fouquet, customers expect demand for its products to exceed available supply through at least 2028, which encourages them to accelerate construction of new fabs as well as expand their existing capacity. </p><p>"Our leading-edge high-volume customers are sharing very good visibility as a result of the increasing number of long-term commitments they are securing with their customers," Fouquet said during the company's conference call with analysts and investors.  "This extended visibility has led them to conclude that demand for their products will outpace supply through at least 2028. As a result, customers are accelerating capacity build-out plans and preparing for further expansion beyond 2028 in order to better meet the anticipated level of demand."</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2465px;"><p class="vanilla-image-block" style="padding-top:48.76%;"><img id="GVfXyCp9tPctfscUnQmFTB" name="NXE3400_simplified_Front_SemiClosed.jpg" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/GVfXyCp9tPctfscUnQmFTB-1920-80.jpg" mos="" align="middle" fullscreen="" width="2465" height="1202" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>ASML is already receiving orders that are to be delivered in 2028, while some customers have expressed interest in equipment for 2029 and even 2030. </p><p>"As customers shift focus to their needs beyond 2027, our order momentum continues to build," Fouquet added. "We are increasingly receiving orders for 2028, and customers are also expressing interest in placing orders for 2029 and even 2030."</p><p>To address this demand, ASML plans to manufacture approximately 65 Low-NA EUV systems in 2026, up from 44 scanners shipped in 2025. Production of immersion DUV systems should remain approximately at last year's level, even though proveiously the company expected shipments of DUV tools to decline. </p><p>The company previously announced plans to increase production capacity for both Low-NA EUV and immersion DUV systems by 30% in 2027, followed by another potential 30% increase in 2028. Based on its 2026 EUV production target, this implies capacity for approximately 85 Low-NA EUV systems in 2027 and roughly 110 in 2028, assuming both increases materialize.</p><p>However, ASML now believes even this level of output will be insufficient and is investigating ways to expand its 2028 capacity beyond the previously planned 30% increase. However, the company has not disclosed revised production targets. </p><p>"Looking further ahead, we shared our intent last quarter to increase our capacity in 2027 by 30% for both Low NA EUV and DUV immersion systems, as well as to investigate adding a further 30% in 2028," explained Roger Dassen, Chief Financial Officer at ASML. "However, given the dynamics of this quarter, we now anticipate that our current planned levels of output in 2027 and 2028 will not fully meet the demand of our customers."</p><h2 id="high-na-euv-adoption">High-NA EUV adoption</h2><p>Meanwhile, High-NA EUV adoption is progressing. Intel has <a href="https://www.tomshardware.com/tech-industry/semiconductors/intel-surpasses-one-million-high-na-euv-wafers-processed-outpaces-the-rest-of-the-industry-combined-company-also-trailblazing-giant-6-12-photomasks-to-speed-production-and-lower-costs1">already exposed more than one million wafers using High-NA systems</a> and plans to start using High-NA EUV with its 14A process technology in 2028. Samsung plans to introduce High-NA EUV into high-volume manufacturing (HVM) in 2028 for DRAM and for logic later, while TSMC plans to use High-NA EUV scanners by 2030 – 2031. ASML says availability across its installed High-NA fleet is approaching 90%, which suggests that the tools are becoming mature enough for HVM deployment.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1280px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="pSy7xJedzqveFGvQJgoiTj" name="asml-lithography-fab-high-na-euv-tool-semiconductor-hero.jpg" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/pSy7xJedzqveFGvQJgoiTj-1920-80.jpg" mos="" align="middle" fullscreen="" width="1280" height="720" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>Interestingly, Samsung's planned 2028 introduction may create additional demand for High-NA equipment from memory manufacturers, as they will need to compete against their larger rival. However, ASML has not disclosed specific shipment targets for these tools.</p><h2 id="new-tools-more-capacity">New tools, more capacity</h2><p>ASML also intends to increase semiconductor production capacity by introducing more productive lithography systems. Its upcoming <a href="https://www.tomshardware.com/tech-industry/semiconductors/asml-lithograpy-roadmap-examined-from-duv-to-hyper-na">Twinscan NXE:3800F</a>, scheduled for introduction in early 2027, will process up to 260 wafers per hour, compared with 230 wafers per hour for the existing NXE:3800E. While this represents 'only' approximately 13% higher throughput, this translates to 21,600 extra EUV exposures per month without corresponding increases in cleanroom space. Three such tools per fab enable 90 additional EUV exposures per hour, 2,160 per day, or 64,800 per month. The company also plans to introduce a new critical-layer immersion DUV system capable of processing 330 wafers per hour.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:47.77%;"><img id="ZCWao36oNchaMoxsHRv9Gi" name="Screenshot 2026-09-07 at 21.37.23" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/ZCWao36oNchaMoxsHRv9Gi-1920-80.png" mos="" align="middle" fullscreen="" width="2560" height="1223" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>Keep in mind that the actual fab capacity improvements depend on the number of lithography exposures required per wafer on a particular node, as well as the capabilities of other manufacturing equipment.</p><figure role="gallery"><figure><img src="https://cdn.mos.cms.futurecdn.net/DgWHgFkQjbAA4MpEPAQcUY-1920-80.png" alt="ASML" /><figcaption><small role="credit">ASML</small></figcaption></figure><figure><img src="https://cdn.mos.cms.futurecdn.net/aCtWrRvfjaVdKGL6hYp4XY-1920-80.png" alt="ASML" /><figcaption><small role="credit">ASML</small></figcaption></figure></figure><p>Another important element of ASML's strategy is its growing portfolio of productivity upgrade packages for existing EUV and DUV systems. Such upgrades enable chipmakers to increase wafer throughput without purchasing additional scanners, which can provide additional capacity considerably faster than procuring new tools or constructing new fabs.</p><h2 id="value-based-pricing-is-a-go">Value-based pricing is a go</h2><p>One of the controversial things that ASML discussed back in early Q2 was its <a href="https://www.tomshardware.com/tech-industry/semiconductors/asmls-planned-low-na-euv-machine-price-hikes-reportedly-frustrate-tsmc-lithography-machine-maker-comes-knocking-to-make-bank-on-tsmcs-profitable-fabs-potentially-costing-the-taiwanese-chipmaker-billions">value-based pricing, which many thought was a blanket price hike</a>. This is not and CFO Roger Dassen this week confirmed that negotiations with major customers have been completed, although the financial benefits for ASML will not become visible immediately.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1280px;"><p class="vanilla-image-block" style="padding-top:56.33%;"><img id="dpptX8fNMiGo3jBGfgMb9j" name="asml-lithography-fab-high-na-euv-tool-semiconductor-3-hero.jpg" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/dpptX8fNMiGo3jBGfgMb9j-1920-80.jpg" mos="" align="middle" fullscreen="" width="1280" height="721" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>The company expects the first effects of higher pricing to appear during 2027, with the full impact becoming apparent in 2028 and beyond, which essentially means that the company is not setting new prices on its backlog, but generally setting new prices on new tools. </p><p>"We completed those conversations," said Dassen. "So that has been done with the key customers. The effect of that is not really visible in 2026. It will start to become visible partly in the course of 2027. The full impact of that you will start to see in 2028 and beyond." </p><p>This strategy could further improve ASML's already substantial gross margins, particularly as customers have no alternatives for advanced lithography equipment. </p><h2 id="china-remains-important">China remains important</h2><p>Despite export restrictions that prevent Chinese chipmakers from acquiring EUV scanners and advanced DUV systems, China remains a substantial market for ASML. The company expects China to account for approximately 20% of its total revenue in 2026, which implies annual sales of roughly €8.8 billion to €9 billion based on current guidance.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:64.06%;"><img id="sF5kjc768gySpL2e9YSSqA" name="Engineer-checking-assembly-instructions_48554.jpg" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/sF5kjc768gySpL2e9YSSqA-1920-80.jpg" mos="" align="middle" fullscreen="" width="2560" height="1640" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>Although China's share of ASML's business has <a href="https://www.tomshardware.com/tech-industry/semiconductors/china-stockpiled-343-immersion-duv-tools-for-advanced-chipmaking-report-claims-270-asml-scanners-can-produce-7nm-processors-without-sanctioned-euv-tools">declined from its levels in 2023 - 2025</a>, the country's semiconductor manufacturers continue to invest in mature node capacity and equipment that remains available under existing export regulations. Meanwhile, it is impossible to predict what is going to happen if the U.S. lawmakers approve the MATCH Act. </p><p>Yet, ASML management emphasizes that its key growth opportunities lie with advanced logic and memory manufacturers, which are accelerating their investment in advanced lithography tools, including sophisticated DUV systems, Low-NA EUV machines, and High-NA EUV scanners.</p><h2 id="the-future-of-asml">The future of ASML</h2><p>Thanks to gaining momentum in AI infrastructure investments, ASML reported record quarterly revenue of $13.06 billion (€11.7 billion) as demand for advanced lithography equipment continued to accelerate. The company expects demand to persist for years as its customers are interested in placing orders for tools to be delivered as far ahead as 2030. To meet growing demand, the company plans to expand production capacity beyond its previous targets for 2028. ASML also intends to introduce new EUV and DUV systems with higher productivity, offer productivity upgrades, and raise equipment prices to capitalize on unprecedented demand.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/semiconductors/asml-reports-record-usd13-billion-quarter-vows-to-expand-capacity-as-demand-for-lithography-equipment-soars-customers-eager-to-secure-orders-in-2030-as-demand-is-expected-to-outstrip-2027-and-2028-supply</link>
                                                                            <description>
                            <![CDATA[ ASML vows to expand production capacity amid growing demand, record earnings and profitability. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">yAXgSfgAVKeMvwTxHotH4e</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/PNvtnHuwa5J8MoZGgvyZc4-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Thu, 15 Oct 2026 12:00:00 +0000</pubDate>                                                                                                                                                                                                                                <category><![CDATA[Semiconductors]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                    <category><![CDATA[Manufacturing]]></category>
                                                                                                <author><![CDATA[ ashilov@gmail.com (Anton Shilov) ]]></author>                    <dc:creator><![CDATA[ Anton Shilov ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/uMZ5kNphxA2Ut6whdLaSQV-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Anton Shilov has been in the PC industry since 1990s playing games, building PCs, and writing stories about pretty much everything that relates to PCs, Macs, smartphones, tablets, and even fab equipment. Over his career, he has worked at a variety of high-ranking websites, including AnandTech, EE Times, TechRadar, X-bit Labs, and now Tom&#039;s Hardware. He is also a regular features contributor to Tom&#039;s Hardware Premium, writing about the latest developments in the semiconductor industry and related tech news and roadmaps. When Anton is not reading or writing about something high-tech, he is probably watching a good movie, playing a video game, or spending time with his family.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/PNvtnHuwa5J8MoZGgvyZc4-1920-80.jpg">
                                                            <media:credit><![CDATA[Getty Images / Freek Van Den Bergh]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[ASML Headquarters]]></media:description>                                                            <media:text><![CDATA[ASML Headquarters]]></media:text>
                                <media:title type="plain"><![CDATA[ASML Headquarters]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/PNvtnHuwa5J8MoZGgvyZc4-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>ASML this week reported record quarterly revenue of $13.06 billion (€11.66 billion) and profitability of nearly $5.79 billion (€5 billion) as demand for advanced semiconductor manufacturing equipment continues to accelerate and investments in AI infrastructure continue to increase. The company now expects demand for its lithography systems to exceed its production capacity in 2027 and 2028, despite plans to increase output substantially. Some clients are already interested in placing orders for equipment as far ahead as 2030, which makes ASML consider further capacity expansions beyond its previously announced intentions.</p><h2 id="record-financial-results">Record financial results</h2><p>ASML reported third-quarter 2026 revenue of $13.06 billion USD (€11.66 billion), up 55% year-over-year and 25% sequentially. The company's gross margin reached a record 57.5%, up from 51.6% a year earlier, while operating income rose to $5.79 billion (€5.17 billion). Net income reached $5.56 billion USD (€4.96 billion) , although this figure includes approximately $570 million (€510 million) from the re-evaluation of the company's investment in <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/independent-tests-rank-mistrals-new-trillion-parameter-large-4-the-best-ai-model-outside-the-u-s-and-china-but-chinese-open-weights-still-overcome-europes-best-efforts">Mistral AI.</a></p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1486px;"><p class="vanilla-image-block" style="padding-top:69.04%;"><img id="j4wvMp9oob5zgNEmu4DLPL" name="Screenshot 2026-10-15 at 07.10.26" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/j4wvMp9oob5zgNEmu4DLPL-1920-80.png" mos="" align="middle" fullscreen="" width="1486" height="1026" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>During the quarter, ASML sold 121 lithography systems, including 113 new and eight refurbished machines, up from 91 systems in Q2 2026. The company generated $4.59 billion USD (€4.1 billion) from 17 EUV lithography systems and $4.82 billion USD (€4.3 billion) from 85 DUV (including 45 ArFi scanners) and 19 i-line litho tools. In addition, ASML recognized revenue from one High-NA EUV system in Q2. Of total equipment revenue, 54% came from logic chipmakers and 46% from memory manufacturers. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1914px;"><p class="vanilla-image-block" style="padding-top:53.40%;"><img id="EXXfjzu7eqfyC8Wf9t3fQL" name="Screenshot 2026-10-15 at 07.10.49" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/EXXfjzu7eqfyC8Wf9t3fQL-1920-80.png" mos="" align="middle" fullscreen="" width="1914" height="1022" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>Taiwan accounted for 32% of ASML's equipment sales in Q3 2026, followed closely by South Korea at 31%. China represented another 23%, while the United States accounted for 13% and the rest of Asia contributed just 1%. Combined, Taiwan and South Korea generated 63% of ASML's equipment revenue, which reflects substantial investments in advanced logic and memory manufacturing capacity that by far exceed those in the U.S. and Japan. </p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1806px;"><p class="vanilla-image-block" style="padding-top:55.59%;"><img id="RgpkxFX7qraArMPFKX9HPL" name="Screenshot 2026-10-15 at 07.10.53" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/RgpkxFX7qraArMPFKX9HPL-1920-80.png" mos="" align="middle" fullscreen="" width="1806" height="1004" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>ASML’s Installed Base Management business generated $3.68 billion (€3.29 billion), approximately $450 million (€400 million) above guidance, largely due to strong demand for productivity upgrades (more on this later), which contributed to the company’s record profitability.</p><p>For the fourth quarter, ASML expects revenue of $15.9 billion to $17.0 billion (€14.2 billion to €15.2 billion) and gross margin of approximately 57%. Full-year revenue should reach $49.3 billion to $50.4 billion (€44 billion to €45 billion), up approximately 36% YoY. The company projects EUV revenue should increase by more than 45% year-over-year, while memory-related equipment sales are expected to rise by approximately 75%, which indicates growing demand from DRAM producers.</p><h2 id="demand-for-litho-tools-spans-to-2030">Demand for litho tools spans to 2030</h2><p>One of the biggest learnings from these results and accompanying statements is that semiconductor manufacturers are increasingly interested in securing lithography equipment years before they actually need it. According to ASML CEO Christophe Fouquet, customers expect demand for its products to exceed available supply through at least 2028, which encourages them to accelerate construction of new fabs as well as expand their existing capacity. </p><p>"Our leading-edge high-volume customers are sharing very good visibility as a result of the increasing number of long-term commitments they are securing with their customers," Fouquet said during the company's conference call with analysts and investors.  "This extended visibility has led them to conclude that demand for their products will outpace supply through at least 2028. As a result, customers are accelerating capacity build-out plans and preparing for further expansion beyond 2028 in order to better meet the anticipated level of demand."</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2465px;"><p class="vanilla-image-block" style="padding-top:48.76%;"><img id="GVfXyCp9tPctfscUnQmFTB" name="NXE3400_simplified_Front_SemiClosed.jpg" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/GVfXyCp9tPctfscUnQmFTB-1920-80.jpg" mos="" align="middle" fullscreen="" width="2465" height="1202" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>ASML is already receiving orders that are to be delivered in 2028, while some customers have expressed interest in equipment for 2029 and even 2030. </p><p>"As customers shift focus to their needs beyond 2027, our order momentum continues to build," Fouquet added. "We are increasingly receiving orders for 2028, and customers are also expressing interest in placing orders for 2029 and even 2030."</p><p>To address this demand, ASML plans to manufacture approximately 65 Low-NA EUV systems in 2026, up from 44 scanners shipped in 2025. Production of immersion DUV systems should remain approximately at last year's level, even though proveiously the company expected shipments of DUV tools to decline. </p><p>The company previously announced plans to increase production capacity for both Low-NA EUV and immersion DUV systems by 30% in 2027, followed by another potential 30% increase in 2028. Based on its 2026 EUV production target, this implies capacity for approximately 85 Low-NA EUV systems in 2027 and roughly 110 in 2028, assuming both increases materialize.</p><p>However, ASML now believes even this level of output will be insufficient and is investigating ways to expand its 2028 capacity beyond the previously planned 30% increase. However, the company has not disclosed revised production targets. </p><p>"Looking further ahead, we shared our intent last quarter to increase our capacity in 2027 by 30% for both Low NA EUV and DUV immersion systems, as well as to investigate adding a further 30% in 2028," explained Roger Dassen, Chief Financial Officer at ASML. "However, given the dynamics of this quarter, we now anticipate that our current planned levels of output in 2027 and 2028 will not fully meet the demand of our customers."</p><h2 id="high-na-euv-adoption">High-NA EUV adoption</h2><p>Meanwhile, High-NA EUV adoption is progressing. Intel has <a href="https://www.tomshardware.com/tech-industry/semiconductors/intel-surpasses-one-million-high-na-euv-wafers-processed-outpaces-the-rest-of-the-industry-combined-company-also-trailblazing-giant-6-12-photomasks-to-speed-production-and-lower-costs1">already exposed more than one million wafers using High-NA systems</a> and plans to start using High-NA EUV with its 14A process technology in 2028. Samsung plans to introduce High-NA EUV into high-volume manufacturing (HVM) in 2028 for DRAM and for logic later, while TSMC plans to use High-NA EUV scanners by 2030 – 2031. ASML says availability across its installed High-NA fleet is approaching 90%, which suggests that the tools are becoming mature enough for HVM deployment.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1280px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="pSy7xJedzqveFGvQJgoiTj" name="asml-lithography-fab-high-na-euv-tool-semiconductor-hero.jpg" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/pSy7xJedzqveFGvQJgoiTj-1920-80.jpg" mos="" align="middle" fullscreen="" width="1280" height="720" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>Interestingly, Samsung's planned 2028 introduction may create additional demand for High-NA equipment from memory manufacturers, as they will need to compete against their larger rival. However, ASML has not disclosed specific shipment targets for these tools.</p><h2 id="new-tools-more-capacity">New tools, more capacity</h2><p>ASML also intends to increase semiconductor production capacity by introducing more productive lithography systems. Its upcoming <a href="https://www.tomshardware.com/tech-industry/semiconductors/asml-lithograpy-roadmap-examined-from-duv-to-hyper-na">Twinscan NXE:3800F</a>, scheduled for introduction in early 2027, will process up to 260 wafers per hour, compared with 230 wafers per hour for the existing NXE:3800E. While this represents 'only' approximately 13% higher throughput, this translates to 21,600 extra EUV exposures per month without corresponding increases in cleanroom space. Three such tools per fab enable 90 additional EUV exposures per hour, 2,160 per day, or 64,800 per month. The company also plans to introduce a new critical-layer immersion DUV system capable of processing 330 wafers per hour.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:47.77%;"><img id="ZCWao36oNchaMoxsHRv9Gi" name="Screenshot 2026-09-07 at 21.37.23" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/ZCWao36oNchaMoxsHRv9Gi-1920-80.png" mos="" align="middle" fullscreen="" width="2560" height="1223" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>Keep in mind that the actual fab capacity improvements depend on the number of lithography exposures required per wafer on a particular node, as well as the capabilities of other manufacturing equipment.</p><figure role="gallery"><figure><img src="https://cdn.mos.cms.futurecdn.net/DgWHgFkQjbAA4MpEPAQcUY-1920-80.png" alt="ASML" /><figcaption><small role="credit">ASML</small></figcaption></figure><figure><img src="https://cdn.mos.cms.futurecdn.net/aCtWrRvfjaVdKGL6hYp4XY-1920-80.png" alt="ASML" /><figcaption><small role="credit">ASML</small></figcaption></figure></figure><p>Another important element of ASML's strategy is its growing portfolio of productivity upgrade packages for existing EUV and DUV systems. Such upgrades enable chipmakers to increase wafer throughput without purchasing additional scanners, which can provide additional capacity considerably faster than procuring new tools or constructing new fabs.</p><h2 id="value-based-pricing-is-a-go">Value-based pricing is a go</h2><p>One of the controversial things that ASML discussed back in early Q2 was its <a href="https://www.tomshardware.com/tech-industry/semiconductors/asmls-planned-low-na-euv-machine-price-hikes-reportedly-frustrate-tsmc-lithography-machine-maker-comes-knocking-to-make-bank-on-tsmcs-profitable-fabs-potentially-costing-the-taiwanese-chipmaker-billions">value-based pricing, which many thought was a blanket price hike</a>. This is not and CFO Roger Dassen this week confirmed that negotiations with major customers have been completed, although the financial benefits for ASML will not become visible immediately.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1280px;"><p class="vanilla-image-block" style="padding-top:56.33%;"><img id="dpptX8fNMiGo3jBGfgMb9j" name="asml-lithography-fab-high-na-euv-tool-semiconductor-3-hero.jpg" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/dpptX8fNMiGo3jBGfgMb9j-1920-80.jpg" mos="" align="middle" fullscreen="" width="1280" height="721" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>The company expects the first effects of higher pricing to appear during 2027, with the full impact becoming apparent in 2028 and beyond, which essentially means that the company is not setting new prices on its backlog, but generally setting new prices on new tools. </p><p>"We completed those conversations," said Dassen. "So that has been done with the key customers. The effect of that is not really visible in 2026. It will start to become visible partly in the course of 2027. The full impact of that you will start to see in 2028 and beyond." </p><p>This strategy could further improve ASML's already substantial gross margins, particularly as customers have no alternatives for advanced lithography equipment. </p><h2 id="china-remains-important">China remains important</h2><p>Despite export restrictions that prevent Chinese chipmakers from acquiring EUV scanners and advanced DUV systems, China remains a substantial market for ASML. The company expects China to account for approximately 20% of its total revenue in 2026, which implies annual sales of roughly €8.8 billion to €9 billion based on current guidance.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:2560px;"><p class="vanilla-image-block" style="padding-top:64.06%;"><img id="sF5kjc768gySpL2e9YSSqA" name="Engineer-checking-assembly-instructions_48554.jpg" alt="ASML" src="https://cdn.mos.cms.futurecdn.net/sF5kjc768gySpL2e9YSSqA-1920-80.jpg" mos="" align="middle" fullscreen="" width="2560" height="1640" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: ASML)</span></figcaption></figure><p>Although China's share of ASML's business has <a href="https://www.tomshardware.com/tech-industry/semiconductors/china-stockpiled-343-immersion-duv-tools-for-advanced-chipmaking-report-claims-270-asml-scanners-can-produce-7nm-processors-without-sanctioned-euv-tools">declined from its levels in 2023 - 2025</a>, the country's semiconductor manufacturers continue to invest in mature node capacity and equipment that remains available under existing export regulations. Meanwhile, it is impossible to predict what is going to happen if the U.S. lawmakers approve the MATCH Act. </p><p>Yet, ASML management emphasizes that its key growth opportunities lie with advanced logic and memory manufacturers, which are accelerating their investment in advanced lithography tools, including sophisticated DUV systems, Low-NA EUV machines, and High-NA EUV scanners.</p><h2 id="the-future-of-asml">The future of ASML</h2><p>Thanks to gaining momentum in AI infrastructure investments, ASML reported record quarterly revenue of $13.06 billion (€11.7 billion) as demand for advanced lithography equipment continued to accelerate. The company expects demand to persist for years as its customers are interested in placing orders for tools to be delivered as far ahead as 2030. To meet growing demand, the company plans to expand production capacity beyond its previous targets for 2028. ASML also intends to introduce new EUV and DUV systems with higher productivity, offer productivity upgrades, and raise equipment prices to capitalize on unprecedented demand.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ SK hynix to place custom tiny CPU cores in its SSD controllers to ease Arm load as drive speeds rocket  ]]></title>
                                                                                                <dc:content><![CDATA[ <p>SK hynix has detailed plans to reduce reliance on power-hungry CPU cores with an in-house core named a Hynix Transaction Processing Unit (H-TPU). Not to be confused with an AI accelerator, an H-TPU would assist with repetitive work handled by an SSD, which usually demands more cores on a controller as drives get faster. These new cores are placed in clusters and would handle routine I/O with a four-stage, in-order pipeline, while management is left to the traditional Arm CPU. The new, tiny cores are claimed to draw only about 2.5% of a general-purpose CPU core’s power, with just around 4% of its area. Now, the company plans to apply H-TPU across its portfolio of products, <a href="https://www.thelec.kr/news/articleView.html?idxno=63561"><em>TheElec</em></a> reports. </p><p>The company’s VP of system-on-chip (SoC) design, Koo In-jae, spoke at ISOCC 2026, the International SoC Design Conference, earlier this week, saying that SSD controller CPUs are at their limit on power and area. In the talk, titled “From SSD Controllers to AI Memory SoCs,” he said that PCIe bandwidth doubles every generation, and controller CPU clocks and control capability reportedly can’t keep up. SK hynix reportedly plans to apply H-TPU across its NAND products, including next-gen enterprise SSD controllers. That’s where the <a href="https://www.tomshardware.com/tech-industry/nvidia-and-kioxia-target-100-million-iops-ssd-in-2027-33-times-more-than-existing-drives-for-exclusive-use-in-ai-servers"><u>unending chase for IOPS</u></a> makes the problem most evident.</p><p>Koo’s comments underline the issue of scaling, which is evident in the “more cores” solutions found in competitor products. <a href="https://www.tomshardware.com/tech-industry/the-current-state-of-pcie-6-0-ssds-and-controllers-marvell-phison-and-smi-prepare-controllers-as-drives-finally-come-to-market-following-years-of-delays">Marvell’s PCIe 6.0 Bravera</a> SC6 uses 12 Cortex-R82 cores in two six-core clusters with three Cortex-M7 cores and a secure Cortex-M3. Compare this to its PCIe 5.0 predecessor, the <a href="https://www.tomshardware.com/news/marvell-announced-pcie-gen5-ssd-controllers"><u>Bravera SC5</u></a>, which lists 10 Cortex-R8 and Cortex-M7 cores plus a Cortex-M3 for security. Core counts have climbed with PCIe speeds out of necessity.</p><p>SK hynix’s answer is a core of its own design: the H-TPU, kept small and programmable. It comes with a dedicated compiler and other development tools for specialized code. It’s stated to use a four-stage pipeline with in-order, or sequential, execution. It is not superscalar, does not use out-of-order units, and should not be confused with a tensor processor for AI. It is designed to handle read and write command processing, manage metadata, and handle buffer control, all repetitive tasks.</p><p>Storage control requires more than 100 detailed functions, which is not ideal for a monolithic CPU. Instead, SK hynix broke those functions up and spread them across “dozens of small TPU clusters,” according to <a href="https://zdnet.co.kr/view/?no=20261012144007" target="_blank"><em>ZDNet Korea</em></a> (translated). Each H-TPU therefore has a specialized job. The total number of H-TPUs is more than 100; the function count above is a separate figure. However we currently have no details on how many cores each cluster holds. </p><p>The main CPU still handles error recovery and exception handling, plus storage-policy management, so the H-TPU complements the CPU: offloading, not replacement. The design idles the cores a given I/O doesn’t need, which reportedly drop into a minimum-power mode. Neither report names the CPU core behind SK hynix’s 2.5% power and 4% area figures.</p><p>Splitting the workload this way is not a new idea. Phison’s CoXProcessor, which pairs Cortex-R5 cores with hardened macros to offload repetitive firmware at wire speed, has been in products for some time. The company’s <a href="https://www.tomshardware.com/features/phison-ps5018-e18"><u>E18 SSD controller</u></a> uses two lower-clocked Cortex-R5 cores to supplement the three main ones. On the enterprise side, Phison’s X1 used two Cortex-R5 cores plus up to 32 “micro cores,” each dedicated to one or two tasks. H-TPU takes this approach to a larger scale.</p><p>While the talk itself is light on detail, we discovered multiple SK hynix patents that describe a similar split between small cores and a CPU. U.S. patent 11,675,537, first filed in Korea in November 2020, details a CPU-run control plane and an I/O plane of many small processors, sorted into four groups by job. A command passes through all four in order: host for ingestion, cache, map, and NVM to the NAND. The processors “configure a pipeline for the data I/O operation” this way, letting the CPU handle any exceptions. That pipeline runs across many cores and is unrelated to the four-stage pipeline inside each H-TPU.</p><p>In the patent’s example, each small core is around “200,000 logic gates” rather than the “seven million logic gates” of the CPU, or about 2.9% of the CPU’s gate count. These are not fixed elements, as “firmware may be driven in each of the TPUs,” making them more adaptable than the hardware circuits the patent calls “difficult to modify.” The patent spells out TPU as “Tensor Processing Units,” but, again, this should not be confused with AI accelerators. H-TPU shares that programmable approach, but the next step SK hynix described moves some routine jobs into fixed hardware.</p><p>This next step is hardware automation (HWA), with command processing, address lookup, and buffer copies moved to dedicated circuits. The CPU and H-TPUs still handle exceptions and key policy decisions. Moving to this framework is meant to keep completion times predictable, which is necessary with I/O, especially for today’s high-IOPS enterprise SSDs.</p><p>SK hynix has filed on fixed-hardware paths before, with a 2014 filing by its San Jose-based unit and, later, with U.S. patent 12,216,914, first filed in Korea in 2022. The latter describes a “write automation circuit” and says “a task without intelligent control can be handled or processed through the hardware layer,” while firmware keeps jobs such as garbage collection. The former went further, sending basic I/O commands down a path “completely implemented in hardware,” with anything it couldn’t handle being sent upstream.</p><p>The H-TPU itself is reportedly being used in “Storage Next,” Nvidia’s name for the joint <a href="https://www.tomshardware.com/pc-components/ssds/sk-hynix-unveils-ai-nand-strategy-including-gargantuan-petabyte-class-qlc-ssds-ultra-fast-hbf-and-100m-iops-ssds-also-in-the-pipeline"><u>AIN P</u></a> project with SK hynix. Storage Next targets 512-byte or more limited access with NAND, which traditionally aims at higher operating sizes. SK hynix also plans to apply it to <a href="https://www.tomshardware.com/pc-components/ssds/sandisk-and-sk-hynix-unveil-hbf-spec-up-to-16-hi-nand-stacks-3-tb-s-bandwidth-ucie"><u>high-bandwidth flash (HBF)</u></a>, which the outlet said is due next year. HBF stacks NAND dies for AI inference, targeting high-bandwidth memory (HBM) levels of bandwidth but with far more capacity, with up to 512GB per stack versus 64GB for HBM4.</p><p>For HBF, AI inference reads outnumber writes around 300,000 to 1, with the AI compiler and runtime passing pre-analyzed dataflow to the HBF SoC. Proper buffer management and prefetching based on that dataflow are meant to hide the slow NAND latency behind compute. Size is important in HBF stacks, so a traditional multi-core Arm design is likely less practical.</p><p>H-TPU has yet to be proven in an actual enterprise SSD controller product, and no date has been given. The move from general-purpose CPU cores to specialized single-job cores and circuits matches the type of workload SSDs handle, and the demands on storage are only increasing as AI servers become memory-constrained.</p><p>SK hynix’s statements and patents support an evolution to H-TPU and beyond to HWA, with H-TPU itself headed for traditional NAND and next year’s HBF. SSD controller manufacturers must be willing to adapt, as simple scaling is suboptimal even at current PCIe levels, let alone the speeds promised for <a href="https://www.tomshardware.com/tech-industry/pci-sig-reveals-pcie-8-0-0-5v-spec-first-draft-includes-1-tb-s-bandwidth-and-new-connector-technology"><u>PCIe 7.0 and 8.0</u></a>, the latter headed for final ratification in 2028.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/pc-components/ssds/sk-hynix-to-place-custom-tiny-cpu-cores-in-its-ssd-controllers-to-ease-arm-load-as-drive-speeds-rocket-each-h-tpu-reportedly-draws-2-5-percent-of-a-cpu-cores-power</link>
                                                                            <description>
                            <![CDATA[ SK hynix detailed H-TPU, tiny in-house cores that offload routine I/O from the Arm CPUs in its next enterprise SSD controllers. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">RduraeW4HEph2WYfHEGNJd</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/vz5pTVmQQQr3XcmP2xscZg-1920-80.png" type="image/png" length="0"></enclosure>
                                                                        <pubDate>Wed, 14 Oct 2026 14:00:00 +0000</pubDate>                                                                                                                                <updated>Wed, 14 Oct 2026 14:21:38 +0000</updated>
                                                                                                                                            <category><![CDATA[SSDs]]></category>
                                                    <category><![CDATA[PC Components]]></category>
                                                    <category><![CDATA[Storage]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/png" url="https://cdn.mos.cms.futurecdn.net/vz5pTVmQQQr3XcmP2xscZg-1920-80.png">
                                                            <media:credit><![CDATA[SK hynix]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[SK hynix PEB210 enterprise SSDs on display at an exhibition booth]]></media:description>                                                            <media:text><![CDATA[SK hynix PEB210 enterprise SSDs on display at an exhibition booth]]></media:text>
                                <media:title type="plain"><![CDATA[SK hynix PEB210 enterprise SSDs on display at an exhibition booth]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/vz5pTVmQQQr3XcmP2xscZg-1920-80.png" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>SK hynix has detailed plans to reduce reliance on power-hungry CPU cores with an in-house core named a Hynix Transaction Processing Unit (H-TPU). Not to be confused with an AI accelerator, an H-TPU would assist with repetitive work handled by an SSD, which usually demands more cores on a controller as drives get faster. These new cores are placed in clusters and would handle routine I/O with a four-stage, in-order pipeline, while management is left to the traditional Arm CPU. The new, tiny cores are claimed to draw only about 2.5% of a general-purpose CPU core’s power, with just around 4% of its area. Now, the company plans to apply H-TPU across its portfolio of products, <a href="https://www.thelec.kr/news/articleView.html?idxno=63561"><em>TheElec</em></a> reports. </p><p>The company’s VP of system-on-chip (SoC) design, Koo In-jae, spoke at ISOCC 2026, the International SoC Design Conference, earlier this week, saying that SSD controller CPUs are at their limit on power and area. In the talk, titled “From SSD Controllers to AI Memory SoCs,” he said that PCIe bandwidth doubles every generation, and controller CPU clocks and control capability reportedly can’t keep up. SK hynix reportedly plans to apply H-TPU across its NAND products, including next-gen enterprise SSD controllers. That’s where the <a href="https://www.tomshardware.com/tech-industry/nvidia-and-kioxia-target-100-million-iops-ssd-in-2027-33-times-more-than-existing-drives-for-exclusive-use-in-ai-servers"><u>unending chase for IOPS</u></a> makes the problem most evident.</p><p>Koo’s comments underline the issue of scaling, which is evident in the “more cores” solutions found in competitor products. <a href="https://www.tomshardware.com/tech-industry/the-current-state-of-pcie-6-0-ssds-and-controllers-marvell-phison-and-smi-prepare-controllers-as-drives-finally-come-to-market-following-years-of-delays">Marvell’s PCIe 6.0 Bravera</a> SC6 uses 12 Cortex-R82 cores in two six-core clusters with three Cortex-M7 cores and a secure Cortex-M3. Compare this to its PCIe 5.0 predecessor, the <a href="https://www.tomshardware.com/news/marvell-announced-pcie-gen5-ssd-controllers"><u>Bravera SC5</u></a>, which lists 10 Cortex-R8 and Cortex-M7 cores plus a Cortex-M3 for security. Core counts have climbed with PCIe speeds out of necessity.</p><p>SK hynix’s answer is a core of its own design: the H-TPU, kept small and programmable. It comes with a dedicated compiler and other development tools for specialized code. It’s stated to use a four-stage pipeline with in-order, or sequential, execution. It is not superscalar, does not use out-of-order units, and should not be confused with a tensor processor for AI. It is designed to handle read and write command processing, manage metadata, and handle buffer control, all repetitive tasks.</p><p>Storage control requires more than 100 detailed functions, which is not ideal for a monolithic CPU. Instead, SK hynix broke those functions up and spread them across “dozens of small TPU clusters,” according to <a href="https://zdnet.co.kr/view/?no=20261012144007" target="_blank"><em>ZDNet Korea</em></a> (translated). Each H-TPU therefore has a specialized job. The total number of H-TPUs is more than 100; the function count above is a separate figure. However we currently have no details on how many cores each cluster holds. </p><p>The main CPU still handles error recovery and exception handling, plus storage-policy management, so the H-TPU complements the CPU: offloading, not replacement. The design idles the cores a given I/O doesn’t need, which reportedly drop into a minimum-power mode. Neither report names the CPU core behind SK hynix’s 2.5% power and 4% area figures.</p><p>Splitting the workload this way is not a new idea. Phison’s CoXProcessor, which pairs Cortex-R5 cores with hardened macros to offload repetitive firmware at wire speed, has been in products for some time. The company’s <a href="https://www.tomshardware.com/features/phison-ps5018-e18"><u>E18 SSD controller</u></a> uses two lower-clocked Cortex-R5 cores to supplement the three main ones. On the enterprise side, Phison’s X1 used two Cortex-R5 cores plus up to 32 “micro cores,” each dedicated to one or two tasks. H-TPU takes this approach to a larger scale.</p><p>While the talk itself is light on detail, we discovered multiple SK hynix patents that describe a similar split between small cores and a CPU. U.S. patent 11,675,537, first filed in Korea in November 2020, details a CPU-run control plane and an I/O plane of many small processors, sorted into four groups by job. A command passes through all four in order: host for ingestion, cache, map, and NVM to the NAND. The processors “configure a pipeline for the data I/O operation” this way, letting the CPU handle any exceptions. That pipeline runs across many cores and is unrelated to the four-stage pipeline inside each H-TPU.</p><p>In the patent’s example, each small core is around “200,000 logic gates” rather than the “seven million logic gates” of the CPU, or about 2.9% of the CPU’s gate count. These are not fixed elements, as “firmware may be driven in each of the TPUs,” making them more adaptable than the hardware circuits the patent calls “difficult to modify.” The patent spells out TPU as “Tensor Processing Units,” but, again, this should not be confused with AI accelerators. H-TPU shares that programmable approach, but the next step SK hynix described moves some routine jobs into fixed hardware.</p><p>This next step is hardware automation (HWA), with command processing, address lookup, and buffer copies moved to dedicated circuits. The CPU and H-TPUs still handle exceptions and key policy decisions. Moving to this framework is meant to keep completion times predictable, which is necessary with I/O, especially for today’s high-IOPS enterprise SSDs.</p><p>SK hynix has filed on fixed-hardware paths before, with a 2014 filing by its San Jose-based unit and, later, with U.S. patent 12,216,914, first filed in Korea in 2022. The latter describes a “write automation circuit” and says “a task without intelligent control can be handled or processed through the hardware layer,” while firmware keeps jobs such as garbage collection. The former went further, sending basic I/O commands down a path “completely implemented in hardware,” with anything it couldn’t handle being sent upstream.</p><p>The H-TPU itself is reportedly being used in “Storage Next,” Nvidia’s name for the joint <a href="https://www.tomshardware.com/pc-components/ssds/sk-hynix-unveils-ai-nand-strategy-including-gargantuan-petabyte-class-qlc-ssds-ultra-fast-hbf-and-100m-iops-ssds-also-in-the-pipeline"><u>AIN P</u></a> project with SK hynix. Storage Next targets 512-byte or more limited access with NAND, which traditionally aims at higher operating sizes. SK hynix also plans to apply it to <a href="https://www.tomshardware.com/pc-components/ssds/sandisk-and-sk-hynix-unveil-hbf-spec-up-to-16-hi-nand-stacks-3-tb-s-bandwidth-ucie"><u>high-bandwidth flash (HBF)</u></a>, which the outlet said is due next year. HBF stacks NAND dies for AI inference, targeting high-bandwidth memory (HBM) levels of bandwidth but with far more capacity, with up to 512GB per stack versus 64GB for HBM4.</p><p>For HBF, AI inference reads outnumber writes around 300,000 to 1, with the AI compiler and runtime passing pre-analyzed dataflow to the HBF SoC. Proper buffer management and prefetching based on that dataflow are meant to hide the slow NAND latency behind compute. Size is important in HBF stacks, so a traditional multi-core Arm design is likely less practical.</p><p>H-TPU has yet to be proven in an actual enterprise SSD controller product, and no date has been given. The move from general-purpose CPU cores to specialized single-job cores and circuits matches the type of workload SSDs handle, and the demands on storage are only increasing as AI servers become memory-constrained.</p><p>SK hynix’s statements and patents support an evolution to H-TPU and beyond to HWA, with H-TPU itself headed for traditional NAND and next year’s HBF. SSD controller manufacturers must be willing to adapt, as simple scaling is suboptimal even at current PCIe levels, let alone the speeds promised for <a href="https://www.tomshardware.com/tech-industry/pci-sig-reveals-pcie-8-0-0-5v-spec-first-draft-includes-1-tb-s-bandwidth-and-new-connector-technology"><u>PCIe 7.0 and 8.0</u></a>, the latter headed for final ratification in 2028.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ AI causes uproar in mathematics as experts warn against sharing research with LLMs after OpenAI shares 722 manuscripts and proofs ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Last week, OpenAI <a href="https://openai.com/index/sharing-ai-progress-in-mathematics/" target="_blank">shared over 400 results</a> for longstanding mathematical formulas that have troubled humans for decades; now researchers say it could take years for humans to validate and prove all the AI-driven work. But while analysts try to check the work of the unnamed OpenAI internal frontier model that found these alleged solutions, mathematicians argue this kind of AI-powered research and release threatens the field itself.</p><p>There's even the claim that by leaping from a prompt to a solution without a clear working in between, there is little benefit to these solutions in the first place. Mathematicians and the field itself may not be able to build on the results, as they don't know how they were achieved, nor has the development of such solutions produced new fields of mathematical understanding. </p><p>To make matters worse, there are suggestions that the growing confluence of mathematical and cybersecurity capabilities of frontier AI could bring into question the sanctity of cryptography itself. Alongside raising serious concerns for national security and modern digital systems, it could also put trillions of dollars worth of cryptocurrency at imminent risk.</p><h2 id="over-722-manuscripts-dropped">Over 722 manuscripts dropped</h2><p>The official count for <a href="https://github.com/openai/math" target="_blank">OpenAI's maths solution release was 722 manuscripts,</a> but with only 300 of the results formalized — digitizing human-readable formulas so they can be computer-verified — it's very difficult to know what's legitimate research. Some of the researchers <em>The Verge </em>spoke to found nonsense explanations within the reports, formulas that don't make sense, and some research that has already been completed by others.</p><p>The burden of proving the veracity of OpenAI and its model's claims is very much on the mathematical community, too. OpenAI didn't release the name of the model it used to generate these results, nor the prompts that generated them — despite those recommendations from the <a href="https://agmai.org/#openai-release" target="_blank">Advisory Group on Mathematicians and Artificial Intelligence</a> (AGMAI). </p><p>OpenAI has <a href="https://github.com/openai/math/blob/main/history.md" target="_blank">at least published changelogs for its GitHub repository</a> of the results, making it clearer where mistakes were made. But with several retractions and corrections already, it seems clear that it didn't check over the results in any great detail before releasing them. </p><p>It's this brashness, the pace of releases, and OpenAI's seeming disinterest in remotely slowing down its progress in the field that have so many mathematicians up in arms. </p><h2 id="a-looming-threat-becomes-immediate">A looming threat becomes immediate</h2><p>Mathematics isn't the first field to suffer a major upheaval because of AI. Writers first felt the looming threat with the release of ChatGPT in 2022, and both visual and musical artists faced the same concerns shortly after. Artificial intelligence is impacting almost every industry in some way, but while there are some potential benefits for its use in the field of mathematics, researchers within the field are concerned about the impact this may have on the collaborative nature of the field and how it might discourage future mathematicians from exploring it.</p><p>Speaking to <a href="https://www.businessinsider.com/openais-math-dump-rattles-math-researchers-saw-programs-wiped-out-2026-10" target="_blank"><em>Business Insider</em></a>, mathematics professor at NYU, Tristan Buckmaster, said that OpenAI's release had upended the world of many early-career mathematicians, suggesting it had wiped out their whole research programs. Alongside their own research being invalidated by some of OpenAI's solutions, he suggested that some were now fearful that, by using AI tools, researchers had inadvertently contributed to the results, raising the potential of them not using advanced tools in the future and being less likely to talk about their research publicly.</p><p>This could cause huge problems for mathematics as a whole, argued mathematics professor at Northwestern University, Bryna Kra. Highlighting how collaborative maths is, with researchers sharing ideas, research data, and formulas, she is now extremely concerned that OpenAI has made mathematicians more secretive, reducing their ability to work together towards solutions and build on each other's research.</p><p>Professor of mathematics at the University of California, Terence Tao <a href="https://mathstodon.xyz/@tao/117395267721642920" target="_blank">said on Mastodon</a> that "many fewer seminars, workshops, collaborations, or other activities are being generated from these results compared to traditional breakthroughs." He argued mathematics may need to pivot to value "progress more holistically," by elevating the role of community building over raw solutions.</p><p>Researchers have described an atmosphere of extreme anxiety, that whatever they're working on could be invalidated overnight. That can undermine job applications, grant proposals, and those building the early components of their academic careers. Understanding and developing core mathematical capabilities at a high level can take years. AI dumps eradicating years of effort in such a haphazard fashion risk making it seem a near-pointless endeavour.</p><p>With swirling rumors of yet more releases from OpenAI, and potentially companion releases from other AI labs chasing the same headlines, the morale in many mathematical fields is extremely low.</p><h2 id="cryptography">Cryptography </h2><p>With how much mathematics underpins key components of our digital world, it's no surprise that major changes in the pace of its development could have a domino effect elsewhere. One proposed potential minefield is the fertile ground of cryptography. </p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2107992246360649879"><p lang="en" dir="ltr">@kmad @matthew_pines I think we might lose public key cryptography.<a href="https://twitter.com/cantworkitout/status/2107992246360649879">October 8, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><p>Associate professor of computer science at Johns Hopkins Whiting School of Engineering, Matthew Green, said on social media that he believes the congruence of mathematical developments and cybersecurity capabilities could result in AI that can defeat encryption as we know it.</p><p>While that raises potential concerns for all manner of fields that rely on encryption to protect the security of sensitive data, it could open up the cryptocurrency industry to wholesale theft, potentially impacting trillions of dollars of digital tokens.</p><p>Not everyone agrees with Green that this is an inevitability. British cryptographer — and one of the few people accused of being Bitcoin developer Satoshi Nakamoto — Adam Back, described these claims as nonsense and that there was little suggestion that just because AI can do certain mathematics now, that they'll be able to defeat encryption at some unnamed date in the future.</p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2108076121895969011"><p lang="en" dir="ltr">@matthew_d_green @kmad @matthew_pines Dubious. Do you think P=NP? Some problems are fundamentally higher complexity, and accelerating mathematically capability doesn't change that any more than gravity. Plus for what bitcoin cares about, even if all algebraic were hardness problems lost, we have hash-based SHRINCS.<a href="https://twitter.com/cantworkitout/status/2108076121895969011">October 8, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><p>Ethereum creator Vitalik Buterin urged calm too, <a href="https://x.com/VitalikButerin/status/2107976296320106851" target="_blank">stating on Twitter/X</a> that nobody should rush to move their tokens to a new wallet or panic over wallet security. However, he did argue that AI had the potential to one day be a far more dangerous threat to the sanctity of cryptocurrency than quantum computing. </p><p>While it doesn't appear that cryptography is a field AI is set to invalidate just yet, it's clear that it poses a potential threat to mathematical researchers and the core underlying 'way of doing things.' Much like it has just about everywhere else. </p><p>Researchers are adjusting to it, but the dust isn't likely to settle anytime soon, and in the meantime, OpenAI and its contemporaries may just move on to a new field to focus on as it continues to chase relevance and profitability.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/artificial-intelligence/ai-causes-uproar-in-mathematics-as-experts-warn-against-sharing-research-with-llms-after-openai-shares-722-manuscripts-and-proofs-professor-argues-that-the-field-must-focus-on-building-community-over-raw-solutions</link>
                                                                            <description>
                            <![CDATA[ Last week, OpenAI shared over 400 results for longstanding mathematical formulas that have troubled humans for decades; now researchers say it could take years for humans to validate and prove all the AI-driven work ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">fbLEPsBwKMMc4JF4HPwPsA</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/XL6skwap7n9uMaZN8gr3TY-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Tue, 13 Oct 2026 12:30:00 +0000</pubDate>                                                                                                                                <updated>Tue, 13 Oct 2026 20:56:27 +0000</updated>
                                                                                                                                            <category><![CDATA[Artificial Intelligence]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Jon Martindale ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/YeutDv8zJmhi7xH35MSt8Z-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;After building his first computers in his teens, Jon Martindale has spent the past two decades covering the latest advances in technology. From displays to PC components, blockchain to AI, and tablets to standing desk accessories, Jon has covered just about every facet of the tech space in his varied career. He has bylines at Forbes, USNews, Lifewire, DigitalTrends, PCWorld, and a range of other sites. He brings that same level of expertise and professional insight to Toms Hardware.Away from writing, Jon is an avid reader, board gamer, and fitness enthusiast. He lives in rural Gloucestershire with his wife, two children, and French Bulldog cross.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/XL6skwap7n9uMaZN8gr3TY-1920-80.jpg">
                                                            <media:credit><![CDATA[Peter M. Fisher via Getty Images]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Tired mathematician leaning on their arm.]]></media:description>                                                            <media:text><![CDATA[Tired mathematician leaning on their arm.]]></media:text>
                                <media:title type="plain"><![CDATA[Tired mathematician leaning on their arm.]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/XL6skwap7n9uMaZN8gr3TY-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Last week, OpenAI <a href="https://openai.com/index/sharing-ai-progress-in-mathematics/" target="_blank">shared over 400 results</a> for longstanding mathematical formulas that have troubled humans for decades; now researchers say it could take years for humans to validate and prove all the AI-driven work. But while analysts try to check the work of the unnamed OpenAI internal frontier model that found these alleged solutions, mathematicians argue this kind of AI-powered research and release threatens the field itself.</p><p>There's even the claim that by leaping from a prompt to a solution without a clear working in between, there is little benefit to these solutions in the first place. Mathematicians and the field itself may not be able to build on the results, as they don't know how they were achieved, nor has the development of such solutions produced new fields of mathematical understanding. </p><p>To make matters worse, there are suggestions that the growing confluence of mathematical and cybersecurity capabilities of frontier AI could bring into question the sanctity of cryptography itself. Alongside raising serious concerns for national security and modern digital systems, it could also put trillions of dollars worth of cryptocurrency at imminent risk.</p><h2 id="over-722-manuscripts-dropped">Over 722 manuscripts dropped</h2><p>The official count for <a href="https://github.com/openai/math" target="_blank">OpenAI's maths solution release was 722 manuscripts,</a> but with only 300 of the results formalized — digitizing human-readable formulas so they can be computer-verified — it's very difficult to know what's legitimate research. Some of the researchers <em>The Verge </em>spoke to found nonsense explanations within the reports, formulas that don't make sense, and some research that has already been completed by others.</p><p>The burden of proving the veracity of OpenAI and its model's claims is very much on the mathematical community, too. OpenAI didn't release the name of the model it used to generate these results, nor the prompts that generated them — despite those recommendations from the <a href="https://agmai.org/#openai-release" target="_blank">Advisory Group on Mathematicians and Artificial Intelligence</a> (AGMAI). </p><p>OpenAI has <a href="https://github.com/openai/math/blob/main/history.md" target="_blank">at least published changelogs for its GitHub repository</a> of the results, making it clearer where mistakes were made. But with several retractions and corrections already, it seems clear that it didn't check over the results in any great detail before releasing them. </p><p>It's this brashness, the pace of releases, and OpenAI's seeming disinterest in remotely slowing down its progress in the field that have so many mathematicians up in arms. </p><h2 id="a-looming-threat-becomes-immediate">A looming threat becomes immediate</h2><p>Mathematics isn't the first field to suffer a major upheaval because of AI. Writers first felt the looming threat with the release of ChatGPT in 2022, and both visual and musical artists faced the same concerns shortly after. Artificial intelligence is impacting almost every industry in some way, but while there are some potential benefits for its use in the field of mathematics, researchers within the field are concerned about the impact this may have on the collaborative nature of the field and how it might discourage future mathematicians from exploring it.</p><p>Speaking to <a href="https://www.businessinsider.com/openais-math-dump-rattles-math-researchers-saw-programs-wiped-out-2026-10" target="_blank"><em>Business Insider</em></a>, mathematics professor at NYU, Tristan Buckmaster, said that OpenAI's release had upended the world of many early-career mathematicians, suggesting it had wiped out their whole research programs. Alongside their own research being invalidated by some of OpenAI's solutions, he suggested that some were now fearful that, by using AI tools, researchers had inadvertently contributed to the results, raising the potential of them not using advanced tools in the future and being less likely to talk about their research publicly.</p><p>This could cause huge problems for mathematics as a whole, argued mathematics professor at Northwestern University, Bryna Kra. Highlighting how collaborative maths is, with researchers sharing ideas, research data, and formulas, she is now extremely concerned that OpenAI has made mathematicians more secretive, reducing their ability to work together towards solutions and build on each other's research.</p><p>Professor of mathematics at the University of California, Terence Tao <a href="https://mathstodon.xyz/@tao/117395267721642920" target="_blank">said on Mastodon</a> that "many fewer seminars, workshops, collaborations, or other activities are being generated from these results compared to traditional breakthroughs." He argued mathematics may need to pivot to value "progress more holistically," by elevating the role of community building over raw solutions.</p><p>Researchers have described an atmosphere of extreme anxiety, that whatever they're working on could be invalidated overnight. That can undermine job applications, grant proposals, and those building the early components of their academic careers. Understanding and developing core mathematical capabilities at a high level can take years. AI dumps eradicating years of effort in such a haphazard fashion risk making it seem a near-pointless endeavour.</p><p>With swirling rumors of yet more releases from OpenAI, and potentially companion releases from other AI labs chasing the same headlines, the morale in many mathematical fields is extremely low.</p><h2 id="cryptography">Cryptography </h2><p>With how much mathematics underpins key components of our digital world, it's no surprise that major changes in the pace of its development could have a domino effect elsewhere. One proposed potential minefield is the fertile ground of cryptography. </p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2107992246360649879"><p lang="en" dir="ltr">@kmad @matthew_pines I think we might lose public key cryptography.<a href="https://twitter.com/cantworkitout/status/2107992246360649879">October 8, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><p>Associate professor of computer science at Johns Hopkins Whiting School of Engineering, Matthew Green, said on social media that he believes the congruence of mathematical developments and cybersecurity capabilities could result in AI that can defeat encryption as we know it.</p><p>While that raises potential concerns for all manner of fields that rely on encryption to protect the security of sensitive data, it could open up the cryptocurrency industry to wholesale theft, potentially impacting trillions of dollars of digital tokens.</p><p>Not everyone agrees with Green that this is an inevitability. British cryptographer — and one of the few people accused of being Bitcoin developer Satoshi Nakamoto — Adam Back, described these claims as nonsense and that there was little suggestion that just because AI can do certain mathematics now, that they'll be able to defeat encryption at some unnamed date in the future.</p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2108076121895969011"><p lang="en" dir="ltr">@matthew_d_green @kmad @matthew_pines Dubious. Do you think P=NP? Some problems are fundamentally higher complexity, and accelerating mathematically capability doesn't change that any more than gravity. Plus for what bitcoin cares about, even if all algebraic were hardness problems lost, we have hash-based SHRINCS.<a href="https://twitter.com/cantworkitout/status/2108076121895969011">October 8, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><p>Ethereum creator Vitalik Buterin urged calm too, <a href="https://x.com/VitalikButerin/status/2107976296320106851" target="_blank">stating on Twitter/X</a> that nobody should rush to move their tokens to a new wallet or panic over wallet security. However, he did argue that AI had the potential to one day be a far more dangerous threat to the sanctity of cryptocurrency than quantum computing. </p><p>While it doesn't appear that cryptography is a field AI is set to invalidate just yet, it's clear that it poses a potential threat to mathematical researchers and the core underlying 'way of doing things.' Much like it has just about everywhere else. </p><p>Researchers are adjusting to it, but the dust isn't likely to settle anytime soon, and in the meantime, OpenAI and its contemporaries may just move on to a new field to focus on as it continues to chase relevance and profitability.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Jev-like decision models are flooding the AI market, after researcher claims TypeSafe copied their research ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Following <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/typesafe-ais-jev-offers-an-alternative-to-llms-that-claims-to-be-193x-faster-and-445x-cheaper-system-one-type-model-is-bespoke-for-probabilistic-decision-making">the release</a> of TypeSafe's Jev, many AI players have launched competitors, and many similar open-source projects from the likes of Microsoft, OpenAI, and many more have since emerged. Yet Nandakishor Mukkunnoth, of ConvAI Innovations, said that <a href="https://dev.to/nandakishor_m_6cc0adfde9f/i-built-non-autoregressive-decision-models-a-year-ago-then-a-frontier-lab-called-it-a-18me">he'd come up with the base notions</a> for Jev's architecture over a year before. So who copied who?</p><p>As a quick recap, Jev and its clones are APIs driven by <em>decision models</em>, for use in your own code. You feed them input data like a customer's online shopping transactions, alongside questions. For example: "is the user looking for a return?" for a yes/no answer, and "Which item are they referring to?" with a multiple-choice answer. A Jev-like AI model will answer each of the questions or choices with a probability that can then be used in your own "if" statements. While the concept by itself isn't strictly new, this practical application is indeed novel.</p><p>The release made a good portion of the internet abuzz, as the seemingly simple concept is extremely useful and neatly sidesteps the usual problems with trying to coax an LLM into giving a direct answer. There's no token-by-token text generation, no conversations, no re-feeding of conversation history, no feeding it gobs of information for context, and no conversational hallucination — just direct data in the form of probabilities. It's really fast too, answering questions in parallel, and exceedingly cheap compared to standard LLMs.</p><p>Along with Jev, TypeSafe also released a brief description of its System One architecture. Only days later, Nandakishor Mukkunnoth, founder of ConvAI Innovations, explained <a href="https://dev.to/nandakishor_m_6cc0adfde9f/i-built-non-autoregressive-decision-models-a-year-ago-then-a-frontier-lab-called-it-a-18me">in a blog post</a> that he'd actually "designed the Jev architecture one year back." He referred to his arXiv paper <a href="https://arxiv.org/abs/2503.23303">on SalesRLAgent</a>, describing how he created a model for probability estimation, published on March 30, 2025. Along with the paper, he published his <a href="https://huggingface.co/DeepMostInnovations/sales-conversion-model-reinf-learning">model weights</a> and <a href="https://huggingface.co/datasets/DeepMostInnovations/saas-sales-conversations">dataset</a>. He later followed up with <a href="https://arxiv.org/abs/2510.01237">a September paper</a> that more closely aligns with how Jev operates.</p><p>In his write-up, Mukkunnoth expressed his frustration, describing how he got overlooked "because [his idea] was built for a vertical use case," meaning SalesRLAgent was made to fix one specific problem, while TypeSafe made Jev a general-purpose decision engine. Notably, the company released Jev as a complete, hosted product, with its own servers, API, documentation, developer kits, and naturally, pricing.</p><p>Nevertheless, he said he opted to not remain bitter about the development, and ultimately released the <a href="https://github.com/NandhaKishorM/laya">open-source Laya</a> roughly the same day. Laya is essentially a Jev-style model and interface that you can run on your own hardware, complete with multi-language support. Likely adding to the confusion, both TypeSafe and Mukkunnoth called their models "System 1/One" types, likely as a reference to Daniel Kahneman's book, <a href="https://en.wikipedia.org/wiki/Thinking,_Fast_and_Slow"><em>Thinking, Fast and Slow</em></a>, which describes System 1 as a faster, emotional mode of thought, whereas System 2 refers to slower, more logical thought.</p><p>Regardless, there was no stopping the hype generated by this new use case for AI, so players in this space did what they do best: copied the concept and rolled out their own services. In the space of only a few weeks after Jev was released, the world saw Microsoft Decision, OpenAI Decisions API, and Fastino Glide. Both Cloudflare Clef and Perplexity Decider arrived with open weights, meaning you can run them on your own gear, assuming it's up to the task. Over on the open-source aisle, there's Amazon's <a href="https://strandsagents.com/blog/introducing-strands-decider/">Strands Decider 2B</a>, Kev, Von, Jeff, jevlike, SemIf, OpenJev, NanoJev, Ollaya, and many others.</p><p>Why so many? Well, because most of these projects, including the commercial ones, share the same root: taking an existing lightweight model, usually always a variant of  Alibaba's open-source Qwen, and adding the extra layers that shape the input, output, and decision-making procedure. Mukkunnoth's Laya is based on encoder-only models from the Bert family, rather than an LLM like Qwen. And yet it's not too hard to imagine that many folks might have cribbed at least some notes from his work, given how quickly the alternatives materialized.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/artificial-intelligence/jev-like-decision-models-are-flooding-the-ai-market-after-researcher-claims-typesafe-copied-their-research-microsoft-openai-among-companies-releasing-lightweight-system-one-ai-models</link>
                                                                            <description>
                            <![CDATA[ Jev and the flood of AI decision models — TypeSafe allegedly copied researcher's architecture but got copied itself ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">ZoLckuro9uDHCsN6ykihvZ</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/JLe7QVUdAmgC2wZ83yU7he-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Mon, 12 Oct 2026 16:00:00 +0000</pubDate>                                                                                                                                <updated>Tue, 13 Oct 2026 20:56:27 +0000</updated>
                                                                                                                                            <category><![CDATA[Artificial Intelligence]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                <author><![CDATA[ editors@tomshardware.com (Bruno Ferreira) ]]></author>                    <dc:creator><![CDATA[ Bruno Ferreira ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/ZQiPPaXaAuQ4VrVEYnnR7G-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Bruno Ferreira&#039;s journey kicked off with the venerable ZX Spectrum, a cassette player, and his hopes and dreams. He quickly realized he had more fun figuring out how computers work than he did actually using the things. Kicking off a developer career with C and Assembly before moving to scripting languages, he&#039;s worn many hats, including both database architect and systems administration. As a teen, Bruno co-founded a web development outfit where he was for 17 years before moving on to spend nearly a decade at The Tech Report as a writer, editor, and (of course) developer. In this decade, he&#039;s been at Asus, MLCommons, and HotHardware, among others. When not fiddling with computers and games, his love for music and production sends him off to live shows and festivals. Occasionally, he pretends he can play the guitar and bass.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/JLe7QVUdAmgC2wZ83yU7he-1920-80.jpg">
                                                            <media:credit><![CDATA[Getty Images]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[A mind with cogs inside]]></media:description>                                                            <media:text><![CDATA[A mind with cogs inside]]></media:text>
                                <media:title type="plain"><![CDATA[A mind with cogs inside]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/JLe7QVUdAmgC2wZ83yU7he-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Following <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/typesafe-ais-jev-offers-an-alternative-to-llms-that-claims-to-be-193x-faster-and-445x-cheaper-system-one-type-model-is-bespoke-for-probabilistic-decision-making">the release</a> of TypeSafe's Jev, many AI players have launched competitors, and many similar open-source projects from the likes of Microsoft, OpenAI, and many more have since emerged. Yet Nandakishor Mukkunnoth, of ConvAI Innovations, said that <a href="https://dev.to/nandakishor_m_6cc0adfde9f/i-built-non-autoregressive-decision-models-a-year-ago-then-a-frontier-lab-called-it-a-18me">he'd come up with the base notions</a> for Jev's architecture over a year before. So who copied who?</p><p>As a quick recap, Jev and its clones are APIs driven by <em>decision models</em>, for use in your own code. You feed them input data like a customer's online shopping transactions, alongside questions. For example: "is the user looking for a return?" for a yes/no answer, and "Which item are they referring to?" with a multiple-choice answer. A Jev-like AI model will answer each of the questions or choices with a probability that can then be used in your own "if" statements. While the concept by itself isn't strictly new, this practical application is indeed novel.</p><p>The release made a good portion of the internet abuzz, as the seemingly simple concept is extremely useful and neatly sidesteps the usual problems with trying to coax an LLM into giving a direct answer. There's no token-by-token text generation, no conversations, no re-feeding of conversation history, no feeding it gobs of information for context, and no conversational hallucination — just direct data in the form of probabilities. It's really fast too, answering questions in parallel, and exceedingly cheap compared to standard LLMs.</p><p>Along with Jev, TypeSafe also released a brief description of its System One architecture. Only days later, Nandakishor Mukkunnoth, founder of ConvAI Innovations, explained <a href="https://dev.to/nandakishor_m_6cc0adfde9f/i-built-non-autoregressive-decision-models-a-year-ago-then-a-frontier-lab-called-it-a-18me">in a blog post</a> that he'd actually "designed the Jev architecture one year back." He referred to his arXiv paper <a href="https://arxiv.org/abs/2503.23303">on SalesRLAgent</a>, describing how he created a model for probability estimation, published on March 30, 2025. Along with the paper, he published his <a href="https://huggingface.co/DeepMostInnovations/sales-conversion-model-reinf-learning">model weights</a> and <a href="https://huggingface.co/datasets/DeepMostInnovations/saas-sales-conversations">dataset</a>. He later followed up with <a href="https://arxiv.org/abs/2510.01237">a September paper</a> that more closely aligns with how Jev operates.</p><p>In his write-up, Mukkunnoth expressed his frustration, describing how he got overlooked "because [his idea] was built for a vertical use case," meaning SalesRLAgent was made to fix one specific problem, while TypeSafe made Jev a general-purpose decision engine. Notably, the company released Jev as a complete, hosted product, with its own servers, API, documentation, developer kits, and naturally, pricing.</p><p>Nevertheless, he said he opted to not remain bitter about the development, and ultimately released the <a href="https://github.com/NandhaKishorM/laya">open-source Laya</a> roughly the same day. Laya is essentially a Jev-style model and interface that you can run on your own hardware, complete with multi-language support. Likely adding to the confusion, both TypeSafe and Mukkunnoth called their models "System 1/One" types, likely as a reference to Daniel Kahneman's book, <a href="https://en.wikipedia.org/wiki/Thinking,_Fast_and_Slow"><em>Thinking, Fast and Slow</em></a>, which describes System 1 as a faster, emotional mode of thought, whereas System 2 refers to slower, more logical thought.</p><p>Regardless, there was no stopping the hype generated by this new use case for AI, so players in this space did what they do best: copied the concept and rolled out their own services. In the space of only a few weeks after Jev was released, the world saw Microsoft Decision, OpenAI Decisions API, and Fastino Glide. Both Cloudflare Clef and Perplexity Decider arrived with open weights, meaning you can run them on your own gear, assuming it's up to the task. Over on the open-source aisle, there's Amazon's <a href="https://strandsagents.com/blog/introducing-strands-decider/">Strands Decider 2B</a>, Kev, Von, Jeff, jevlike, SemIf, OpenJev, NanoJev, Ollaya, and many others.</p><p>Why so many? Well, because most of these projects, including the commercial ones, share the same root: taking an existing lightweight model, usually always a variant of  Alibaba's open-source Qwen, and adding the extra layers that shape the input, output, and decision-making procedure. Mukkunnoth's Laya is based on encoder-only models from the Bert family, rather than an LLM like Qwen. And yet it's not too hard to imagine that many folks might have cribbed at least some notes from his work, given how quickly the alternatives materialized.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Mistral’s new Large 4 trails some Chinese open models in independent tests ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Mistral launched its trillion-parameter Large 4 (ML4) model this week, saying it “pushes the frontier of open-weight performance.” Artificial Analysis (AA), an independent AI benchmarking firm, ranked Mistral's Large 4 as the best model available from outside the U.S. and China in its index. The mixture-of-experts model, trained on Nvidia’s <a href="https://www.tomshardware.com/pc-components/gpus/nvidias-next-gen-ai-gpu-revealed-blackwell-b200-gpu-delivers-up-to-20-petaflops-of-compute-and-massive-improvements-over-hopper-h100"><u>Grace Blackwell</u></a> GPUs, scored its personal best in AA’s CyberGym-E2E-AA end-to-end cybersecurity measure, although several Chinese open models beat it overall.</p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2107467221421420919"><p lang="en" dir="ltr">Mistral has released Mistral Large 4, scoring 38 on the Artificial Analysis Intelligence Index; France is back to having the most intelligent model from outside the US and China@MistralAI has released Mistral Large 4 in Research Public Preview, with plans to release the weights of the 1T parameter (49B active) model at the end of October. It achieves 38 on the Artificial Analysis Intelligence Index, comparable to GPT-6 Luna (max, 38) and DeepSeek V4.1 Flash (max, 39). It also achieves 50% on the Artificial Analysis Cyber Index, level with GLM-5.3-Flash and ahead of models such as Kimi K3 and DeepSeek V4.1 Flash (max).Key benchmarking results for Mistral Large 4 Preview:➤ Most intelligent model from outside the US and China: Mistral Large 4 Preview scores 38 on the Intelligence Index, comparable to DeepSeek V4.1 Flash (max, 39) and GPT-6 Luna (max, 38). This makes it the most intelligent model from outside the US and China, ahead of countries such as South Korea and the United Arab Emirates➤ Level with GLM-5.3-Flash on cyber defense capability: Mistral Large 4 Preview scores 50 on the Artificial Analysis Cyber Index, level with GLM-5.3-Flash (50) and behind MiMo-V2.6-Pro (56). Once its weights are released, it will rank among the top three open weights models on the Cyber Index. Its strongest result is on CyberGym-E2E-AA, where it scores 82%, ahead of MiMo-V2.6-Pro (79%) and GPT-6 Luna (max, 78%)➤ Over 4x the Cost per Task of similar-intelligence open weights models: Mistral Large 4 Preview costs $1.13 per Intelligence Index task with standard pricing of $1.36/$4.18 per 1M input/output tokens, with $0.14 per 1M cached input tokens. For the first two weeks, Mistral Large 4 Preview will be served at a 50% launch discount, bringing its Cost per Task down to $0.57. This is still more costly than GLM-5.3-Flash ($0.25) and DeepSeek V4.1 Flash (max, $0.27)➤ Strong document and image reasoning: Mistral Large 4 Preview scores 19% on GDP.pdf, on par with MiMo-V2.6-Pro (19%) and behind Kimi K3 (22%). This is an 18-point improvement from Mistral Large 3, partly driven by improvements in their API, which now accepts 100 images per request, up from 8 for previous Mistral modelsKey model details:➤ Context Window: 512k tokens➤ Multimodality: Text and image input, with text output➤ Pricing: $1.36/$4.18 per 1M input/output tokens ($0.14 per 1M cached input tokens), with 50% off for the first two weeks ($0.68/$2.09)➤ Availability: Research Public Preview on Mistral's API, with open weights planned for the end of October<a href="https://twitter.com/cantworkitout/status/2107467221421420919">October 6, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><p>ML4 scored 38 on the Artificial Analysis Intelligence Index v4.3.2, which AA says makes it the “most intelligent model from outside the US and China." At least five <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/chinas-open-weight-ai-models-are-now-just-4-months-behind-frontier-us-offerings-mozilla-report-claims-models-still-lag-in-some-benchmarks-but-are-drastically-cheaper-to-use"><u>Chinese open models</u></a> scored higher on the index: DeepSeek V4.1 Flash (max) at 39, GLM-5.3-Flash at 42, <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3"><u>Kimi K3</u></a> (max) at 44, GLM-5.3 (max) at 45, and MiMo-V2.6-Pro at 46. While MiMo-V2.6-Pro is comparable in total size to ML4, GLM-5.3-Flash is about one-third the size at 320B parameters total and 18B versus ML4’s 49B–52B active. Unlike the rest, ML4 isn’t open yet: Mistral says its weights are due “by the end of the month.”</p><div ><table><caption>Mistral Large 4 vs. Chinese open models on Artificial Analysis' Intelligence Index (Oct. 6)</caption><thead><tr><th class="firstcol " ><p>Model</p></th><th  ><p>Index score</p></th><th  ><p>Parameters, total / active</p></th><th  ><p>Cost per index task</p></th><th  ><p>Weights</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Xiaomi MiMo-V2.6-Pro</p></td><td  ><p>46</p></td><td  ><p>1.0T / 42B</p></td><td  ><p>$0.13</p></td><td  ><p>Open</p></td></tr><tr><td class="firstcol " ><p>Z.ai GLM-5.3 (max)</p></td><td  ><p>45</p></td><td  ><p>753B / 40B</p></td><td  ><p>$2.01</p></td><td  ><p>Open</p></td></tr><tr><td class="firstcol " ><p>Moonshot Kimi K3 (max)</p></td><td  ><p>44</p></td><td  ><p>2.8T / 104B</p></td><td  ><p>$2.00</p></td><td  ><p>Open</p></td></tr><tr><td class="firstcol " ><p>Z.ai GLM-5.3-Flash</p></td><td  ><p>42</p></td><td  ><p>320B / 18B</p></td><td  ><p>$0.25</p></td><td  ><p>Open</p></td></tr><tr><td class="firstcol " ><p>DeepSeek V4.1 Flash (max)</p></td><td  ><p>39</p></td><td  ><p>552B / 16B</p></td><td  ><p>$0.27</p></td><td  ><p>Open</p></td></tr><tr><td class="firstcol " ><p>Mistral Large 4 Preview</p></td><td  ><p>38</p></td><td  ><p>1.05T / 49B–52B</p></td><td  ><p>$1.13 ($0.57 at launch pricing)</p></td><td  ><p>Due by end of October</p></td></tr><tr><td class="firstcol " ><p>DeepSeek V4 Pro 0813 (max)</p></td><td  ><p>36</p></td><td  ><p>1.6T / 49B</p></td><td  ><p>$0.67</p></td><td  ><p>Open</p></td></tr></tbody></table></div><p><a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/frontier-ai-faces-pricing-reckoning-as-token-volume-explodes-25-fold-mid-tier-models-deliver-90-percent-of-flagship-capability-at-one-sixth-the-cost"><u>Cost is a different story</u></a>. ML4 comes in at $1.13 per AA task at list price, or $0.57 at the 50% discounted launch price. This compares to $0.13 for MiMo-V2.6-Pro and $0.25 for GLM-5.3-Flash. Although comparatively expensive, ML4 scored what AA calls its “strongest result” and Mistral calls “the highest of any model” with an 81.7% score on CyberGym-E2E-AA, compared with GLM-5.3-Flash’s 74%, from a model whose post-training run Mistral claims is “still in flight.”</p><p>While the score is impressive, it is only one of three tests in the Cyber Index; ML4 only scores 16% in DeepsecBench-AA and 51% in CWE-Bench-AA. However, AA’s projection is that, once the weights ship, it “will rank among the top three open weights models on the Cyber Index.” This leaves the door open for ML4 to apply to narrower tasks.</p><div ><table><caption>Mistral Large 4 on Artificial Analysis' Cyber Index and its three tests (Oct. 6)</caption><thead><tr><th class="firstcol " ><p>Model</p></th><th  ><p>Cyber Index</p></th><th  ><p>CWE-Bench-AA</p></th><th  ><p>DeepsecBench-AA</p></th><th  ><p>CyberGym-E2E-AA</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Grok 4.7 (xhigh)</p></td><td  ><p>56</p></td><td  ><p>68%</p></td><td  ><p>27%</p></td><td  ><p>74%</p></td></tr><tr><td class="firstcol " ><p>Xiaomi MiMo-V2.6-Pro</p></td><td  ><p>56</p></td><td  ><p>63%</p></td><td  ><p>26%</p></td><td  ><p>79%</p></td></tr><tr><td class="firstcol " ><p>OpenAI GPT-6 Luna (max)</p></td><td  ><p>53</p></td><td  ><p>57%</p></td><td  ><p>24%</p></td><td  ><p>78%</p></td></tr><tr><td class="firstcol " ><p>Z.ai GLM-5.3-Flash</p></td><td  ><p>50</p></td><td  ><p>56%</p></td><td  ><p>20%</p></td><td  ><p>74%</p></td></tr><tr><td class="firstcol " ><p>Mistral Large 4 Preview</p></td><td  ><p>50</p></td><td  ><p>51%</p></td><td  ><p>16%</p></td><td  ><p>82%</p></td></tr><tr><td class="firstcol " ><p>OpenAI GPT-6 Astra (max; declined some tasks)</p></td><td  ><p>33</p></td><td  ><p>63%</p></td><td  ><p>37%</p></td><td  ><p>0%</p></td></tr><tr><td class="firstcol " ><p>Anthropic Claude Opus 5.5 (max with fallback; declined some tasks)</p></td><td  ><p>29</p></td><td  ><p>58%</p></td><td  ><p>27%</p></td><td  ><p>1%</p></td></tr></tbody></table></div><p>Mistral bets on this, saying ML4 is “state-of-the-art among open models” on enterprise workloads such as “cybersecurity, finance and law,” and that in visual grounding, where its 1.6B-parameter vision encoder comes in, it can go as far as “surpassing even frontier closed models.” On AA’s Cyber Index, Mistral says, ML4 ranks “among the top five models globally,” and it solves 93% of Cybench, a set of 40 challenges drawn from security competitions.</p><p>CEO Arthur Mensch went further in Abu Dhabi, saying the model is “above the Chinese models on certain aspects, including cyber,” <em>Reuters</em> reported. Mistral’s other figures include DeepSWE v1.1 at 61.7%, Coding Agent Index at 49.8%, AutomationBench at 59.9%, and Dense 200 at 42%; AA’s own chart puts ML4 at 27% on Terminal-Bench 4.0.</p><p>ML4 does, in fact, beat OpenAI’s closed GPT-6 Luna (max) on CyberGym-E2E-AA, but Luna matches ML4’s 38 on AA’s index at about one-sixteenth of the cost per task ($0.07). Mistral says frontier models such as Claude Opus 5.5 and GPT-6 Astra “score near zero on the same test because they refuse to perform the task,” and AA’s chart marks both as declining some tasks on safety grounds.</p><p>Outside of that test, Mistral says ML4 also beats DeepSeek V4 Pro 0813 and Qwen3.8 Max on the Coding Agent Index; per AA, it is cheaper than the latter. Overall, AA’s index-vs.-cost chart puts ML4 below the Pareto line and just outside the “Most attractive quadrant.”</p><p>The cost comes from needing 200 million output tokens to run AA’s index, against a median of 81 million. As for speed, ML4 generated 116.1 tokens per second, with 1.46 seconds to the first token. This puts ML4 at about 33% faster than its tier median and 2.6x faster to the first token. AA’s speed is measured on Mistral’s own API, which Mistral says runs on the same hardware the model used for training.</p><p>ML4 was “trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own datacenters in Europe,” the company said. This is up from 3,000 Nvidia H200s used to train Large 3, released last year, which was smaller at 675B total parameters with 41B active. The training occurred over “roughly two months,” as reported by <em>VentureBeat</em>.</p><p>Mistral raised a €3 billion Series D in what the company calls “the largest equity round ever raised by a European technology company” in its post. Mistral Compute has planned for <a href="https://www.tomshardware.com/tech-industry/semiconductors/asml-makes-usd1-5billion-investment-in-mistral-ai-asml-becomes-the-largest-shareholder-for-the-french-ai-start-up"><u>18,000 Grace Blackwell chips</u></a>, as announced with Nvidia last year. Mistral says it is “significantly scaling up our compute capacity in our own European datacenters,” and that ML4 will become the base for “a new generation of specialized and optimized Mistral models.” The model’s open weights, once released, will be able to run on a private cloud or on-premises.</p><p>The weights, architecture details, and additional benchmarks are expected before the end of October. Artificial Analysis lists the preview as proprietary until the weights arrive. Once it does, the model may be re-scored, and Mistral’s claims will be held to task. For the first two weeks following release, ML4 can be used via API at a discounted rate.</p><p>While European models still have a lot to prove, based on previous performance and ML4’s showing with an 8-point gap versus a similar-size Chinese model that’s also a lot less expensive, Mistral has demonstrated resolve and produced a model with, at least, niche application, as the basis to eventually close that gap.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/artificial-intelligence/independent-tests-rank-mistrals-new-trillion-parameter-large-4-the-best-ai-model-outside-the-u-s-and-china-but-chinese-open-weights-still-overcome-europes-best-efforts</link>
                                                                            <description>
                            <![CDATA[ Artificial Analysis scores Mistral’s new Large 4 at 38, behind open models from Xiaomi, Z.ai, Moonshot, and DeepSeek. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">MKR67xiYqfkGJRjySBZMnC</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/aPGFkWcuerwAS8Pm6AWY4a-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Fri, 09 Oct 2026 11:00:00 +0000</pubDate>                                                                                                                                <updated>Sat, 10 Oct 2026 13:10:47 +0000</updated>
                                                                                                                                            <category><![CDATA[Artificial Intelligence]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/aPGFkWcuerwAS8Pm6AWY4a-1920-80.jpg">
                                                            <media:credit><![CDATA[Getty Images / Bloomberg]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Mistral AI on a desktop]]></media:description>                                                            <media:text><![CDATA[Mistral AI on a desktop]]></media:text>
                                <media:title type="plain"><![CDATA[Mistral AI on a desktop]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/aPGFkWcuerwAS8Pm6AWY4a-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Mistral launched its trillion-parameter Large 4 (ML4) model this week, saying it “pushes the frontier of open-weight performance.” Artificial Analysis (AA), an independent AI benchmarking firm, ranked Mistral's Large 4 as the best model available from outside the U.S. and China in its index. The mixture-of-experts model, trained on Nvidia’s <a href="https://www.tomshardware.com/pc-components/gpus/nvidias-next-gen-ai-gpu-revealed-blackwell-b200-gpu-delivers-up-to-20-petaflops-of-compute-and-massive-improvements-over-hopper-h100"><u>Grace Blackwell</u></a> GPUs, scored its personal best in AA’s CyberGym-E2E-AA end-to-end cybersecurity measure, although several Chinese open models beat it overall.</p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2107467221421420919"><p lang="en" dir="ltr">Mistral has released Mistral Large 4, scoring 38 on the Artificial Analysis Intelligence Index; France is back to having the most intelligent model from outside the US and China@MistralAI has released Mistral Large 4 in Research Public Preview, with plans to release the weights of the 1T parameter (49B active) model at the end of October. It achieves 38 on the Artificial Analysis Intelligence Index, comparable to GPT-6 Luna (max, 38) and DeepSeek V4.1 Flash (max, 39). It also achieves 50% on the Artificial Analysis Cyber Index, level with GLM-5.3-Flash and ahead of models such as Kimi K3 and DeepSeek V4.1 Flash (max).Key benchmarking results for Mistral Large 4 Preview:➤ Most intelligent model from outside the US and China: Mistral Large 4 Preview scores 38 on the Intelligence Index, comparable to DeepSeek V4.1 Flash (max, 39) and GPT-6 Luna (max, 38). This makes it the most intelligent model from outside the US and China, ahead of countries such as South Korea and the United Arab Emirates➤ Level with GLM-5.3-Flash on cyber defense capability: Mistral Large 4 Preview scores 50 on the Artificial Analysis Cyber Index, level with GLM-5.3-Flash (50) and behind MiMo-V2.6-Pro (56). Once its weights are released, it will rank among the top three open weights models on the Cyber Index. Its strongest result is on CyberGym-E2E-AA, where it scores 82%, ahead of MiMo-V2.6-Pro (79%) and GPT-6 Luna (max, 78%)➤ Over 4x the Cost per Task of similar-intelligence open weights models: Mistral Large 4 Preview costs $1.13 per Intelligence Index task with standard pricing of $1.36/$4.18 per 1M input/output tokens, with $0.14 per 1M cached input tokens. For the first two weeks, Mistral Large 4 Preview will be served at a 50% launch discount, bringing its Cost per Task down to $0.57. This is still more costly than GLM-5.3-Flash ($0.25) and DeepSeek V4.1 Flash (max, $0.27)➤ Strong document and image reasoning: Mistral Large 4 Preview scores 19% on GDP.pdf, on par with MiMo-V2.6-Pro (19%) and behind Kimi K3 (22%). This is an 18-point improvement from Mistral Large 3, partly driven by improvements in their API, which now accepts 100 images per request, up from 8 for previous Mistral modelsKey model details:➤ Context Window: 512k tokens➤ Multimodality: Text and image input, with text output➤ Pricing: $1.36/$4.18 per 1M input/output tokens ($0.14 per 1M cached input tokens), with 50% off for the first two weeks ($0.68/$2.09)➤ Availability: Research Public Preview on Mistral's API, with open weights planned for the end of October<a href="https://twitter.com/cantworkitout/status/2107467221421420919">October 6, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><p>ML4 scored 38 on the Artificial Analysis Intelligence Index v4.3.2, which AA says makes it the “most intelligent model from outside the US and China." At least five <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/chinas-open-weight-ai-models-are-now-just-4-months-behind-frontier-us-offerings-mozilla-report-claims-models-still-lag-in-some-benchmarks-but-are-drastically-cheaper-to-use"><u>Chinese open models</u></a> scored higher on the index: DeepSeek V4.1 Flash (max) at 39, GLM-5.3-Flash at 42, <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3"><u>Kimi K3</u></a> (max) at 44, GLM-5.3 (max) at 45, and MiMo-V2.6-Pro at 46. While MiMo-V2.6-Pro is comparable in total size to ML4, GLM-5.3-Flash is about one-third the size at 320B parameters total and 18B versus ML4’s 49B–52B active. Unlike the rest, ML4 isn’t open yet: Mistral says its weights are due “by the end of the month.”</p><div ><table><caption>Mistral Large 4 vs. Chinese open models on Artificial Analysis' Intelligence Index (Oct. 6)</caption><thead><tr><th class="firstcol " ><p>Model</p></th><th  ><p>Index score</p></th><th  ><p>Parameters, total / active</p></th><th  ><p>Cost per index task</p></th><th  ><p>Weights</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Xiaomi MiMo-V2.6-Pro</p></td><td  ><p>46</p></td><td  ><p>1.0T / 42B</p></td><td  ><p>$0.13</p></td><td  ><p>Open</p></td></tr><tr><td class="firstcol " ><p>Z.ai GLM-5.3 (max)</p></td><td  ><p>45</p></td><td  ><p>753B / 40B</p></td><td  ><p>$2.01</p></td><td  ><p>Open</p></td></tr><tr><td class="firstcol " ><p>Moonshot Kimi K3 (max)</p></td><td  ><p>44</p></td><td  ><p>2.8T / 104B</p></td><td  ><p>$2.00</p></td><td  ><p>Open</p></td></tr><tr><td class="firstcol " ><p>Z.ai GLM-5.3-Flash</p></td><td  ><p>42</p></td><td  ><p>320B / 18B</p></td><td  ><p>$0.25</p></td><td  ><p>Open</p></td></tr><tr><td class="firstcol " ><p>DeepSeek V4.1 Flash (max)</p></td><td  ><p>39</p></td><td  ><p>552B / 16B</p></td><td  ><p>$0.27</p></td><td  ><p>Open</p></td></tr><tr><td class="firstcol " ><p>Mistral Large 4 Preview</p></td><td  ><p>38</p></td><td  ><p>1.05T / 49B–52B</p></td><td  ><p>$1.13 ($0.57 at launch pricing)</p></td><td  ><p>Due by end of October</p></td></tr><tr><td class="firstcol " ><p>DeepSeek V4 Pro 0813 (max)</p></td><td  ><p>36</p></td><td  ><p>1.6T / 49B</p></td><td  ><p>$0.67</p></td><td  ><p>Open</p></td></tr></tbody></table></div><p><a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/frontier-ai-faces-pricing-reckoning-as-token-volume-explodes-25-fold-mid-tier-models-deliver-90-percent-of-flagship-capability-at-one-sixth-the-cost"><u>Cost is a different story</u></a>. ML4 comes in at $1.13 per AA task at list price, or $0.57 at the 50% discounted launch price. This compares to $0.13 for MiMo-V2.6-Pro and $0.25 for GLM-5.3-Flash. Although comparatively expensive, ML4 scored what AA calls its “strongest result” and Mistral calls “the highest of any model” with an 81.7% score on CyberGym-E2E-AA, compared with GLM-5.3-Flash’s 74%, from a model whose post-training run Mistral claims is “still in flight.”</p><p>While the score is impressive, it is only one of three tests in the Cyber Index; ML4 only scores 16% in DeepsecBench-AA and 51% in CWE-Bench-AA. However, AA’s projection is that, once the weights ship, it “will rank among the top three open weights models on the Cyber Index.” This leaves the door open for ML4 to apply to narrower tasks.</p><div ><table><caption>Mistral Large 4 on Artificial Analysis' Cyber Index and its three tests (Oct. 6)</caption><thead><tr><th class="firstcol " ><p>Model</p></th><th  ><p>Cyber Index</p></th><th  ><p>CWE-Bench-AA</p></th><th  ><p>DeepsecBench-AA</p></th><th  ><p>CyberGym-E2E-AA</p></th></tr></thead><tbody><tr><td class="firstcol " ><p>Grok 4.7 (xhigh)</p></td><td  ><p>56</p></td><td  ><p>68%</p></td><td  ><p>27%</p></td><td  ><p>74%</p></td></tr><tr><td class="firstcol " ><p>Xiaomi MiMo-V2.6-Pro</p></td><td  ><p>56</p></td><td  ><p>63%</p></td><td  ><p>26%</p></td><td  ><p>79%</p></td></tr><tr><td class="firstcol " ><p>OpenAI GPT-6 Luna (max)</p></td><td  ><p>53</p></td><td  ><p>57%</p></td><td  ><p>24%</p></td><td  ><p>78%</p></td></tr><tr><td class="firstcol " ><p>Z.ai GLM-5.3-Flash</p></td><td  ><p>50</p></td><td  ><p>56%</p></td><td  ><p>20%</p></td><td  ><p>74%</p></td></tr><tr><td class="firstcol " ><p>Mistral Large 4 Preview</p></td><td  ><p>50</p></td><td  ><p>51%</p></td><td  ><p>16%</p></td><td  ><p>82%</p></td></tr><tr><td class="firstcol " ><p>OpenAI GPT-6 Astra (max; declined some tasks)</p></td><td  ><p>33</p></td><td  ><p>63%</p></td><td  ><p>37%</p></td><td  ><p>0%</p></td></tr><tr><td class="firstcol " ><p>Anthropic Claude Opus 5.5 (max with fallback; declined some tasks)</p></td><td  ><p>29</p></td><td  ><p>58%</p></td><td  ><p>27%</p></td><td  ><p>1%</p></td></tr></tbody></table></div><p>Mistral bets on this, saying ML4 is “state-of-the-art among open models” on enterprise workloads such as “cybersecurity, finance and law,” and that in visual grounding, where its 1.6B-parameter vision encoder comes in, it can go as far as “surpassing even frontier closed models.” On AA’s Cyber Index, Mistral says, ML4 ranks “among the top five models globally,” and it solves 93% of Cybench, a set of 40 challenges drawn from security competitions.</p><p>CEO Arthur Mensch went further in Abu Dhabi, saying the model is “above the Chinese models on certain aspects, including cyber,” <em>Reuters</em> reported. Mistral’s other figures include DeepSWE v1.1 at 61.7%, Coding Agent Index at 49.8%, AutomationBench at 59.9%, and Dense 200 at 42%; AA’s own chart puts ML4 at 27% on Terminal-Bench 4.0.</p><p>ML4 does, in fact, beat OpenAI’s closed GPT-6 Luna (max) on CyberGym-E2E-AA, but Luna matches ML4’s 38 on AA’s index at about one-sixteenth of the cost per task ($0.07). Mistral says frontier models such as Claude Opus 5.5 and GPT-6 Astra “score near zero on the same test because they refuse to perform the task,” and AA’s chart marks both as declining some tasks on safety grounds.</p><p>Outside of that test, Mistral says ML4 also beats DeepSeek V4 Pro 0813 and Qwen3.8 Max on the Coding Agent Index; per AA, it is cheaper than the latter. Overall, AA’s index-vs.-cost chart puts ML4 below the Pareto line and just outside the “Most attractive quadrant.”</p><p>The cost comes from needing 200 million output tokens to run AA’s index, against a median of 81 million. As for speed, ML4 generated 116.1 tokens per second, with 1.46 seconds to the first token. This puts ML4 at about 33% faster than its tier median and 2.6x faster to the first token. AA’s speed is measured on Mistral’s own API, which Mistral says runs on the same hardware the model used for training.</p><p>ML4 was “trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own datacenters in Europe,” the company said. This is up from 3,000 Nvidia H200s used to train Large 3, released last year, which was smaller at 675B total parameters with 41B active. The training occurred over “roughly two months,” as reported by <em>VentureBeat</em>.</p><p>Mistral raised a €3 billion Series D in what the company calls “the largest equity round ever raised by a European technology company” in its post. Mistral Compute has planned for <a href="https://www.tomshardware.com/tech-industry/semiconductors/asml-makes-usd1-5billion-investment-in-mistral-ai-asml-becomes-the-largest-shareholder-for-the-french-ai-start-up"><u>18,000 Grace Blackwell chips</u></a>, as announced with Nvidia last year. Mistral says it is “significantly scaling up our compute capacity in our own European datacenters,” and that ML4 will become the base for “a new generation of specialized and optimized Mistral models.” The model’s open weights, once released, will be able to run on a private cloud or on-premises.</p><p>The weights, architecture details, and additional benchmarks are expected before the end of October. Artificial Analysis lists the preview as proprietary until the weights arrive. Once it does, the model may be re-scored, and Mistral’s claims will be held to task. For the first two weeks following release, ML4 can be used via API at a discounted rate.</p><p>While European models still have a lot to prove, based on previous performance and ML4’s showing with an 8-point gap versus a similar-size Chinese model that’s also a lot less expensive, Mistral has demonstrated resolve and produced a model with, at least, niche application, as the basis to eventually close that gap.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Former Groq engineers sue board over $20 billion Nvidia deal ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Two former Groq engineers are suing Groq’s former board and former CEO Jonathan Ross, alleging the board handed over employees and company technology to Nvidia in a $20 billion deal. The proposed class action, filed in Delaware’s Court of Chancery by Benjamin Serebrin and Joshua Rubin and unsealed Monday, alleges the deal, with a $17 billion license fee for all backers and a separate $3 billion Nvidia stock pool for engineers who transferred, kept billions from other shareholders. Meanwhile, Groq is planning to deploy its own technology, now Nvidia’s Groq 3 LPX,  <a href="https://www.ft.com/content/93ee425d-9ac7-4543-8cc9-fef2e0670787"><u>the </u><u><em>Financial Times</em></u></a> reports.</p><p>Nvidia’s annual report (10-K) books $14.4 billion of the license as goodwill “primarily attributable to the workforce and future development of the licensed technology” plus $2.5 billion for the technology itself. The plaintiffs concede that “no Delaware decision has directly answered the question this case raises,” but the deal is already under a reported Department of Justice (DOJ) inquiry. The lawsuit’s outcome has implications for future license-plus-hire deals. Groq told <em>CNBC </em>the suit is “meritless” and said its licensing agreement with Nvidia “delivered exceptional value for Groq, our investors, and our employees.” </p><p>The deal involved both the LPU technology and <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-buys-ai-chip-startup-groqs-assets-for-usd20-billion-in-the-companys-biggest-deal-ever-transaction-includes-acquihires-of-key-groq-employees-including-ceo"><u>the people who built it</u></a>, including Ross, a former Google TPU engineer who founded Groq in 2016. The LP30 has 500MB SRAM per die and 150 TB/s bandwidth, <a href="https://www.tomshardware.com/tech-industry/semiconductors/nvidias-20-billion-groq-deal-produces-its-first-chip"><u>uses Samsung’s 4nm process technology</u></a>, and is rated by Nvidia at up to 1.23 FP8 PFLOPS. A single LPX rack has 256 LPUs with 128GB SRAM, 40 PB/s, and 315 FP8 PFLOPS in aggregate, as a decode co-processor for Vera Rubin NVL72. The LP30 appears to be the next-gen LPU Groq roadmapped in August 2023 on Samsung’s SF4X process with its first-gen Tensor Streaming architecture.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1920px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="jDVsdo5TQfePXWQypttpVg" name="Groq 3 LPX" alt="Nvidia slide showing the Groq 3 LPX rack and its specifications" src="https://cdn.mos.cms.futurecdn.net/jDVsdo5TQfePXWQypttpVg-1920-80.jpg" mos="" align="middle" fullscreen="" width="1920" height="1080" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Nvidia)</span></figcaption></figure><p><a href="https://www.tomshardware.com/tech-industry/semiconductors/nvidia-presents-groq-3-lpx-architecture-and-unveils-its-first-third-party-inference-benchmark"><u>During Hot Chips in August</u></a> of this year, Igor Arsovski, Groq’s former chief architect and now Nvidia VP of hardware, called the occasion “a pinch me moment for the Groq team that’s now integrated into the Nvidia group.” The rack was already in production at the time, Arsovski said. This supports the complaint’s factual premise, but does not answer the legal question about whether more is owed.</p><p>Nvidia’s 10-K notes indicate that “no customer contracts, existing products, or equity interests were purchased” with “$13.0 billion paid at closing and $4 billion, inclusive of imputed interest, payable within one year.” The goodwill number, as reported above, is about 85% of the $16.9 billion Nvidia booked and around 5.8x the value of the technology asset. Nvidia put significantly more value on the LPU design team and the technology’s future development than on the design itself.</p><p>Groq said when the deal was announced that it would continue “to operate as an independent company,” and it joined the Nvidia Cloud Partner program in August. Also in August, “Groq LLC” had a $350 million Series A and a $3.5 billion valuation after raising $650 million in June. Before the deal, in September 2025, the valuation was put at $6.9 billion after a $750 million round. Groq says it now operates 13 data centers and expects to scale from 54 to 200+ MW in 2027. Nvidia was slated to participate in the Series A, which Groq says will support those seeking “clusters of NVIDIA accelerated computing.”</p><p>Groq also said in August that it is “among the first adopters of NVIDIA Groq 3 LPX” with Nvidia’s Vera Rubin NVL72, deploying with Dell. Those racks are built on the LPU technology Groq licensed to Nvidia, and Groq claims to be “the only team with hands-on experience operating LPUs in production at scale.” The <em>Financial Times </em>said Groq “pivoted entirely to AI cloud computing, dropping its chip design efforts.”</p><p>The suit alleges that the Groq board was conflicted, failing to secure the best terms for every shareholder and denying some of them a vote on the deal. Specifically, it says Ross and other top Groq employees were allowed to “take a discount on those shares and be paid separately for following the technology to Nvidia.” According to the suit, Nvidia hired “nearly all” of Groq’s engineers, as many as 200. However, the final payout is alleged to have ignored potential upside and synergies with Nvidia.</p><p>The plaintiffs also say that because Nvidia didn’t buy Groq outright, the $17 billion license payment was taxable as income at Groq, and that the Series A valued the remaining Groq above the price paid to the shareholders who were bought out. They say four “conflicted funds” on Groq’s board, BlackRock, Social Capital, Infinitum, and Disruptive, which are not named as defendants in the case, engineered that gap and profited by staying with Groq after the buyout.</p><p>As reported by Axios in December 2025, though, sources close to the deal said that most shareholders would get per-share payouts “tied to the $20 billion valuation.” Around 90% of employees would join Nvidia, with vested shares paid in cash and unvested shares paid in vesting Nvidia stock. Staff who stayed with Groq would be paid for vested shares and get a package including “economic participation in the ongoing company.”</p><p>As for the DOJ, the <em>New York Times</em> reported in September that the department sent Nvidia a formal request for information with a fine being possible. Nvidia said at the time that the Groq story was “a prime example of the American system working as designed.” The deal did not go unnoticed by some senators who, in February, put it under the “reverse acqui-hire” deal label. Federal Trade Commission Chair Andrew Ferguson said in January that the agency is “beginning to examine these acqui-hires to make sure they are not an attempt to get around” its merger review process.</p><p>The litigation comes not long after Nvidia’s agreement to acquire Hugging Face (HF) a month ago, which included a purchase price and equity for HF employees who join Nvidia, a similar split. The deal is waiting on approval, but unlike the Groq deal, it is an outright purchase of the company. When the Groq deal was announced, Nvidia CEO Jensen Huang told employees in an email, obtained by <em>CNBC</em>, that while Nvidia was adding Groq staff and licensing its IP, “we are not acquiring Groq as a company.”</p><p>The legal question on the Groq deal remains open, and more details will come out as the case proceeds and the defendants respond. As for the LPX customers, Nvidia calls Nebius the first AI cloud to adopt LPX and says the racks will be online there later this year. Looking ahead, Nvidia has the LP35, with NVFP4 support, alongside the Rubin Ultra and <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-updates-data-center-roadmap-with-rosa-cpu-and-stacked-feynman-gpus-optical-nvlink-groq-lpus-with-nvfp4-and-nvlink-also-on-deck"><u>LP40 for Feynman on its roadmap</u></a>.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/semiconductors/former-groq-engineers-sue-board-over-usd20-billion-nvidia-deal-saying-it-handed-nvidia-the-lpu-and-the-team-that-built-it-plaintiffs-allege-the-board-kept-billions-from-other-shareholders</link>
                                                                            <description>
                            <![CDATA[ Two former Groq engineers sued Groq’s board in Delaware over the Nvidia licensing deal behind the Groq 3 LPU. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">Ldzua9XRChdLNzdnfNsHYL</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/49QX9DhQjJDwWiR2NPT9tD-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Thu, 08 Oct 2026 11:00:00 +0000</pubDate>                                                                                                                                <updated>Thu, 08 Oct 2026 12:21:08 +0000</updated>
                                                                                                                                            <category><![CDATA[Semiconductors]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                    <category><![CDATA[Manufacturing]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/49QX9DhQjJDwWiR2NPT9tD-1920-80.jpg">
                                                            <media:credit><![CDATA[Nvidia]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Nvidia Rubin GPU package beside the smaller Groq 3 LPU package]]></media:description>                                                            <media:text><![CDATA[Nvidia Rubin GPU package beside the smaller Groq 3 LPU package]]></media:text>
                                <media:title type="plain"><![CDATA[Nvidia Rubin GPU package beside the smaller Groq 3 LPU package]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/49QX9DhQjJDwWiR2NPT9tD-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Two former Groq engineers are suing Groq’s former board and former CEO Jonathan Ross, alleging the board handed over employees and company technology to Nvidia in a $20 billion deal. The proposed class action, filed in Delaware’s Court of Chancery by Benjamin Serebrin and Joshua Rubin and unsealed Monday, alleges the deal, with a $17 billion license fee for all backers and a separate $3 billion Nvidia stock pool for engineers who transferred, kept billions from other shareholders. Meanwhile, Groq is planning to deploy its own technology, now Nvidia’s Groq 3 LPX,  <a href="https://www.ft.com/content/93ee425d-9ac7-4543-8cc9-fef2e0670787"><u>the </u><u><em>Financial Times</em></u></a> reports.</p><p>Nvidia’s annual report (10-K) books $14.4 billion of the license as goodwill “primarily attributable to the workforce and future development of the licensed technology” plus $2.5 billion for the technology itself. The plaintiffs concede that “no Delaware decision has directly answered the question this case raises,” but the deal is already under a reported Department of Justice (DOJ) inquiry. The lawsuit’s outcome has implications for future license-plus-hire deals. Groq told <em>CNBC </em>the suit is “meritless” and said its licensing agreement with Nvidia “delivered exceptional value for Groq, our investors, and our employees.” </p><p>The deal involved both the LPU technology and <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-buys-ai-chip-startup-groqs-assets-for-usd20-billion-in-the-companys-biggest-deal-ever-transaction-includes-acquihires-of-key-groq-employees-including-ceo"><u>the people who built it</u></a>, including Ross, a former Google TPU engineer who founded Groq in 2016. The LP30 has 500MB SRAM per die and 150 TB/s bandwidth, <a href="https://www.tomshardware.com/tech-industry/semiconductors/nvidias-20-billion-groq-deal-produces-its-first-chip"><u>uses Samsung’s 4nm process technology</u></a>, and is rated by Nvidia at up to 1.23 FP8 PFLOPS. A single LPX rack has 256 LPUs with 128GB SRAM, 40 PB/s, and 315 FP8 PFLOPS in aggregate, as a decode co-processor for Vera Rubin NVL72. The LP30 appears to be the next-gen LPU Groq roadmapped in August 2023 on Samsung’s SF4X process with its first-gen Tensor Streaming architecture.</p><figure class="van-image-figure  inline-layout" data-bordeaux-image-check="" ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1920px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="jDVsdo5TQfePXWQypttpVg" name="Groq 3 LPX" alt="Nvidia slide showing the Groq 3 LPX rack and its specifications" src="https://cdn.mos.cms.futurecdn.net/jDVsdo5TQfePXWQypttpVg-1920-80.jpg" mos="" align="middle" fullscreen="" width="1920" height="1080" attribution="" endorsement="" class="inline" /></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Nvidia)</span></figcaption></figure><p><a href="https://www.tomshardware.com/tech-industry/semiconductors/nvidia-presents-groq-3-lpx-architecture-and-unveils-its-first-third-party-inference-benchmark"><u>During Hot Chips in August</u></a> of this year, Igor Arsovski, Groq’s former chief architect and now Nvidia VP of hardware, called the occasion “a pinch me moment for the Groq team that’s now integrated into the Nvidia group.” The rack was already in production at the time, Arsovski said. This supports the complaint’s factual premise, but does not answer the legal question about whether more is owed.</p><p>Nvidia’s 10-K notes indicate that “no customer contracts, existing products, or equity interests were purchased” with “$13.0 billion paid at closing and $4 billion, inclusive of imputed interest, payable within one year.” The goodwill number, as reported above, is about 85% of the $16.9 billion Nvidia booked and around 5.8x the value of the technology asset. Nvidia put significantly more value on the LPU design team and the technology’s future development than on the design itself.</p><p>Groq said when the deal was announced that it would continue “to operate as an independent company,” and it joined the Nvidia Cloud Partner program in August. Also in August, “Groq LLC” had a $350 million Series A and a $3.5 billion valuation after raising $650 million in June. Before the deal, in September 2025, the valuation was put at $6.9 billion after a $750 million round. Groq says it now operates 13 data centers and expects to scale from 54 to 200+ MW in 2027. Nvidia was slated to participate in the Series A, which Groq says will support those seeking “clusters of NVIDIA accelerated computing.”</p><p>Groq also said in August that it is “among the first adopters of NVIDIA Groq 3 LPX” with Nvidia’s Vera Rubin NVL72, deploying with Dell. Those racks are built on the LPU technology Groq licensed to Nvidia, and Groq claims to be “the only team with hands-on experience operating LPUs in production at scale.” The <em>Financial Times </em>said Groq “pivoted entirely to AI cloud computing, dropping its chip design efforts.”</p><p>The suit alleges that the Groq board was conflicted, failing to secure the best terms for every shareholder and denying some of them a vote on the deal. Specifically, it says Ross and other top Groq employees were allowed to “take a discount on those shares and be paid separately for following the technology to Nvidia.” According to the suit, Nvidia hired “nearly all” of Groq’s engineers, as many as 200. However, the final payout is alleged to have ignored potential upside and synergies with Nvidia.</p><p>The plaintiffs also say that because Nvidia didn’t buy Groq outright, the $17 billion license payment was taxable as income at Groq, and that the Series A valued the remaining Groq above the price paid to the shareholders who were bought out. They say four “conflicted funds” on Groq’s board, BlackRock, Social Capital, Infinitum, and Disruptive, which are not named as defendants in the case, engineered that gap and profited by staying with Groq after the buyout.</p><p>As reported by Axios in December 2025, though, sources close to the deal said that most shareholders would get per-share payouts “tied to the $20 billion valuation.” Around 90% of employees would join Nvidia, with vested shares paid in cash and unvested shares paid in vesting Nvidia stock. Staff who stayed with Groq would be paid for vested shares and get a package including “economic participation in the ongoing company.”</p><p>As for the DOJ, the <em>New York Times</em> reported in September that the department sent Nvidia a formal request for information with a fine being possible. Nvidia said at the time that the Groq story was “a prime example of the American system working as designed.” The deal did not go unnoticed by some senators who, in February, put it under the “reverse acqui-hire” deal label. Federal Trade Commission Chair Andrew Ferguson said in January that the agency is “beginning to examine these acqui-hires to make sure they are not an attempt to get around” its merger review process.</p><p>The litigation comes not long after Nvidia’s agreement to acquire Hugging Face (HF) a month ago, which included a purchase price and equity for HF employees who join Nvidia, a similar split. The deal is waiting on approval, but unlike the Groq deal, it is an outright purchase of the company. When the Groq deal was announced, Nvidia CEO Jensen Huang told employees in an email, obtained by <em>CNBC</em>, that while Nvidia was adding Groq staff and licensing its IP, “we are not acquiring Groq as a company.”</p><p>The legal question on the Groq deal remains open, and more details will come out as the case proceeds and the defendants respond. As for the LPX customers, Nvidia calls Nebius the first AI cloud to adopt LPX and says the racks will be online there later this year. Looking ahead, Nvidia has the LP35, with NVFP4 support, alongside the Rubin Ultra and <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-updates-data-center-roadmap-with-rosa-cpu-and-stacked-feynman-gpus-optical-nvlink-groq-lpus-with-nvfp4-and-nvlink-also-on-deck"><u>LP40 for Feynman on its roadmap</u></a>.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Qualcomm will license patents behind Huawei’s LogicFolding chip architecture, report says ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Qualcomm has reached a deal to license Huawei patents behind the LogicFolding chip-stacking method, <a href="https://www.bloomberg.com/news/articles/2026-10-05/qualcomm-licenses-patents-on-huawei-s-logicfolding-chip-tech"><u><em>Bloomberg</em></u></a> reports. The technology, which is used in the Kirin 9050 Pro, splits nearly every block across two bonded dies, and a teardown and die shots show the lower die holds mostly cache and I/O. </p><p>Qualcomm’s multiyear, cross-license deal with Huawei has a wide scope, covering 5G devices, AI services, near- and co-packaged optics and networks, a Huawei spokesperson told <em>Bloomberg</em>. Qualcomm also intends to purchase specific Huawei U.S. patents in compute, AI, networking, and other technologies, according to a release. Neither company’s announcement named any patents, LogicFolding, or any chip that uses it.</p><p><em>Bloomberg's</em> report frames the deal as a win for Huawei, validating its chipmaking acumen. The deal remains subject to approval, and Qualcomm will <a href="https://www.tomshardware.com/tech-industry/semiconductors/qualcomm-will-pay-huawei-to-license-its-patents-for-the-first-time-in-historic-turnaround-reversal-in-a-5g-and-ai-cross-licensing-deal-comes-25-years-after-huawei-first-paid-qualcomm">reportedly pay Huawei to license its patents for the first time</a>. According to the report, LogicFolding targets faster data transmission as a workaround for China’s lack of access to EUV lithography machines.</p><p>The Kirin 9050 Pro is used in the Huawei Mate XT 2 tri-fold smartphone, launched Sept. 7. Geekerwan, a Chinese tech review channel known for in-depth chip testing, had a lab slice the chip for its teardown. Kurnal, a semiconductor researcher who publishes die-shot analysis, also posted annotated shots of both dies on X. Geekerwan’s cross-section shows the two dies bonded face to face, and both show the execution units on top, with the PLLs also on the lower die.</p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2105967029861953597"><p lang="en" dir="ltr">Hisilicon Hi36E0(Kirin 9050) Dieshothttps://t.co/iKiY621Bmg https://t.co/qup0ANqYUE<a href="https://twitter.com/cantworkitout/status/2105967029861953597">October 2, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><p>LogicFolding was <a href="https://www.tomshardware.com/tech-industry/semiconductors/huawei-claims-sanctions-busting-breakthrough-with-1-4nm-class-chips-by-2031-claims-55-percent-higher-transistor-density-firm-claims-new-logicfolding-chip-architecture-can-bypass-euv-restrictions-introduces-tau-scaling-law-to-replace-moores-law"><u>first unveiled on May 25</u></a> by He Tingbo, president of Huawei’s semiconductor business, at IEEE ISCAS 2026 in Shanghai, under the company’s “Tau Scaling Law,” which uses signal delay rather than transistor size as a measure. Huawei describes both layers as active and working as one. Vertical interconnects work to create shorter signal paths, which can reduce energy consumption. <a href="https://www.tomshardware.com/tech-industry/semiconductors/peking-university-builds-3d-chip-design-tool-tailored-to-huaweis-logicfolding-architecture"><u>Peking University’s “true 3D” design tool</u></a> for the technology shows a 30% wire-length cut in its own tests.</p><p>Huawei’s papers describe a “Kirin 2026” chip whose figures match the Kirin 9050 Pro’s. In a paper last month, relayed by <em>TrendForce</em>, Huawei said the chip uses a 1.5-micrometer hybrid-bonding pitch with about 50 million interconnects. About 10–15% of those carry signals. Core routing is 20% shorter, with critical paths up to 70% shorter. Power is reduced by as much as 66% for NPU, 58% for GPU, and 41% for a CPU performance core at the same performance level. NPU performance holds 29 TOPS at 63% less frequency and at a lower voltage.</p><p>Huawei does not name a process node. On its roadmap is Kirin 2027 at a 1-micrometer pitch and more than 100 million interconnects; for around three years out, a 720-nanometer pitch with over 200 million. It is targeting a 1.4nm-class density by 2031, with LogicFolding in the Ascend 990 AI accelerator around 2030. HiSilicon, Huawei’s chip-design unit, designs the Kirin chips, and SMIC is believed to make them, according to Reuters via <em>TrendForce</em>. It isn’t stated who handles the bonding process.</p><p>As for the Kirin 9050 Pro, Geekerwan tested it in a Mate 90 Pro Max phone. Analysis shows that the two dies are bonded with copper-to-copper hybrid bonding, metal layer to metal layer. The bonded circuit layers of both dies sit between two thick layers of silicon rather than directly on the package substrate. For the interconnects, about 80,000 through-silicon vias (TSVs) run through the lower die, taking up, with their keep-out zones, about 8% of the lower die’s usable area, Geekerwan says.</p><div class="youtube-video" data-nosnippet ><div class="video-aspect-box"><iframe data-lazy-priority="high" data-lazy-src="https://www.youtube-nocookie.com/embed/PpNFmPcBYPY" allowfullscreen=""></iframe></div></div><p>The split spans both dies. In the big CPU core, the execution units are on top, with the L1 and L2 caches directly beneath the load/store units, Geekerwan says. Kurnal's die shots put the GPU and NPU logic over their caches too. The lower die also holds a shared L3, a 12MiB system cache, PLLs, and the LPDDR, PCIe, camera, display, UFS, and USB interfaces, while the top die carries the ISP, modem, and memory controllers. Cache, I/O, and PLLs produce less heat and are less sensitive to the process, so they go below, Geekerwan explains; the denser compute goes on top, on a somewhat more advanced process.</p><p>Each die is just over 120 mm², which is smaller than the previous chip’s 140-plus mm². The total silicon area for both is therefore over 240 mm², up 70% overall. Density rose from 155 to 238 MTr/mm² at launch, about 53.5% higher, according to Huawei’s launch numbers. Other notable changes include the NPU’s area, up 150% across both dies, with Geekerwan measuring 68 TOPS INT8. The configuration makes sense and the rationale follows the layout, but an independent die analysis is needed for full understanding.</p><p>Huawei’s stacking approach has its competitors. <a href="https://www.tomshardware.com/tech-industry/semiconductors/hybrid-bonding-roadmap-examined"><u>TSMC’s SoIC</u></a> roadmap set a 6-micrometer bond pitch for 2025 and 4.5 micrometers by 2029, with 3nm chip stacking in volume production last year. Intel’s Foveros Direct is said to have “copper bonding at a pitch of 9µm,” which reached high volume with the Clearwater Forest Xeon, and a second generation targets 3 micrometers. <a href="https://www.tomshardware.com/news/amd-unveils-more-ryzen-3d-packaging-and-v-cache-details-at-hot-chips"><u>AMD’s first-generation 3D V-Cache</u></a> stacks a 64MB L3 SRAM die on a complete CPU die, using direct copper bonding at a 9-micrometer pitch. Huawei’s 1.5-micrometer claim comes from its own paper and any comparison has to be taken with a grain of salt.</p><p>As to why Qualcomm might be interested, John Han, executive vice president and general manager of Qualcomm Technology Licensing, said the agreement reflects Qualcomm’s “recognition of Huawei’s continued innovation and intellectual property in 5G and other technology fields.” Bloomberg’s report says that at least some of the patents underpin LogicFolding, but no source names the patents. Dr. Ian Cutress, chief analyst at More Than Moore, questioned how new Huawei’s Tau Scaling is, writing on X that it is “a repackaging of what the industry has always done.”</p><p>Upon “receipt of the necessary regulatory approvals,” the cross-license and patent purchase deal should proceed. The deal is also subject to Federal Trade Commission review under the Hart-Scott-Rodino Act, Bloomberg reported. That U.S. premerger law sets a minimum deal size, $133.9 million this year, among other tests, and the deal’s value wasn’t disclosed.</p><p><em>Nikkei Asia</em> reported that Qualcomm will be the net payer under the deal. However Qualcomm said on October 6 that reports characterizing it as a net payor on the license agreement are incorrect and that assertions linking the agreement to LogicFold are not accurate. Qualcomm’s licensing revenue stopped including Huawei royalties in fiscal Q2 2025, after the license expired. This follows a long history of Huawei paying for Qualcomm licenses, an agreement first inked in 2001. Huawei says it has brought in more license-related income than it’s spent since 2021. Patents behind its chip-stacking method reportedly join that business under the deal. </p><p>The industry now waits to see what comes of the FTC review. Qualcomm’s fiscal Q4 results should come early next month, as last year’s did, but its fiscal year ended Sept. 27, before the deal was announced. Additional independent analysis of the 9050 Pro’s die may come from <em>TechInsights</em> or <em>SemiAnalysis</em>. Also of interest is the next iteration in Huawei’s Kirin 2027, to see if its claims are met. The deal between the two companies currently lacks details on the extent of patent coverage, but the silicon on record can help reveal the nature of Qualcomm’s interest.</p><p><em><strong>Update: 10/12/2026, 02:50 PT: Added passage reflecting Qualcomm's October 6 statement.</strong></em></p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/semiconductors/qualcomm-will-license-patents-behind-huaweis-logicfolding-chip-architecture-report-says-the-kirin-9050-pro-already-uses-it-with-a-teardown-showing-its-lower-die-is-mostly-cache-and-i-o</link>
                                                                            <description>
                            <![CDATA[ Qualcomm will license patents behind Huawei’s LogicFolding, according to one report, as a teardown shows off the Kirin 9050 Pro’s two dies. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">oPRaQKvDdJM9H3ti5u5X5P</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/LXVxdeGZ5SXqimCn6caf2h-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Wed, 07 Oct 2026 13:00:00 +0000</pubDate>                                                                                                                                <updated>Mon, 12 Oct 2026 12:25:38 +0000</updated>
                                                                                                                                            <category><![CDATA[Semiconductors]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                    <category><![CDATA[Manufacturing]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/LXVxdeGZ5SXqimCn6caf2h-1920-80.jpg">
                                                            <media:credit><![CDATA[Getty Images / CFOTO]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Huawei Logo as displayed on a building]]></media:description>                                                            <media:text><![CDATA[Huawei Logo as displayed on a building]]></media:text>
                                <media:title type="plain"><![CDATA[Huawei Logo as displayed on a building]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/LXVxdeGZ5SXqimCn6caf2h-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Qualcomm has reached a deal to license Huawei patents behind the LogicFolding chip-stacking method, <a href="https://www.bloomberg.com/news/articles/2026-10-05/qualcomm-licenses-patents-on-huawei-s-logicfolding-chip-tech"><u><em>Bloomberg</em></u></a> reports. The technology, which is used in the Kirin 9050 Pro, splits nearly every block across two bonded dies, and a teardown and die shots show the lower die holds mostly cache and I/O. </p><p>Qualcomm’s multiyear, cross-license deal with Huawei has a wide scope, covering 5G devices, AI services, near- and co-packaged optics and networks, a Huawei spokesperson told <em>Bloomberg</em>. Qualcomm also intends to purchase specific Huawei U.S. patents in compute, AI, networking, and other technologies, according to a release. Neither company’s announcement named any patents, LogicFolding, or any chip that uses it.</p><p><em>Bloomberg's</em> report frames the deal as a win for Huawei, validating its chipmaking acumen. The deal remains subject to approval, and Qualcomm will <a href="https://www.tomshardware.com/tech-industry/semiconductors/qualcomm-will-pay-huawei-to-license-its-patents-for-the-first-time-in-historic-turnaround-reversal-in-a-5g-and-ai-cross-licensing-deal-comes-25-years-after-huawei-first-paid-qualcomm">reportedly pay Huawei to license its patents for the first time</a>. According to the report, LogicFolding targets faster data transmission as a workaround for China’s lack of access to EUV lithography machines.</p><p>The Kirin 9050 Pro is used in the Huawei Mate XT 2 tri-fold smartphone, launched Sept. 7. Geekerwan, a Chinese tech review channel known for in-depth chip testing, had a lab slice the chip for its teardown. Kurnal, a semiconductor researcher who publishes die-shot analysis, also posted annotated shots of both dies on X. Geekerwan’s cross-section shows the two dies bonded face to face, and both show the execution units on top, with the PLLs also on the lower die.</p><div class="see-more see-more--clipped"><figure><blockquote class="twitter-tweet hawk-ignore" data-lang="en" cite="https://twitter.com/cantworkitout/status/2105967029861953597"><p lang="en" dir="ltr">Hisilicon Hi36E0(Kirin 9050) Dieshothttps://t.co/iKiY621Bmg https://t.co/qup0ANqYUE<a href="https://twitter.com/cantworkitout/status/2105967029861953597">October 2, 2026</a></p></blockquote></figure><div class="see-more__filter"></div></div><p>LogicFolding was <a href="https://www.tomshardware.com/tech-industry/semiconductors/huawei-claims-sanctions-busting-breakthrough-with-1-4nm-class-chips-by-2031-claims-55-percent-higher-transistor-density-firm-claims-new-logicfolding-chip-architecture-can-bypass-euv-restrictions-introduces-tau-scaling-law-to-replace-moores-law"><u>first unveiled on May 25</u></a> by He Tingbo, president of Huawei’s semiconductor business, at IEEE ISCAS 2026 in Shanghai, under the company’s “Tau Scaling Law,” which uses signal delay rather than transistor size as a measure. Huawei describes both layers as active and working as one. Vertical interconnects work to create shorter signal paths, which can reduce energy consumption. <a href="https://www.tomshardware.com/tech-industry/semiconductors/peking-university-builds-3d-chip-design-tool-tailored-to-huaweis-logicfolding-architecture"><u>Peking University’s “true 3D” design tool</u></a> for the technology shows a 30% wire-length cut in its own tests.</p><p>Huawei’s papers describe a “Kirin 2026” chip whose figures match the Kirin 9050 Pro’s. In a paper last month, relayed by <em>TrendForce</em>, Huawei said the chip uses a 1.5-micrometer hybrid-bonding pitch with about 50 million interconnects. About 10–15% of those carry signals. Core routing is 20% shorter, with critical paths up to 70% shorter. Power is reduced by as much as 66% for NPU, 58% for GPU, and 41% for a CPU performance core at the same performance level. NPU performance holds 29 TOPS at 63% less frequency and at a lower voltage.</p><p>Huawei does not name a process node. On its roadmap is Kirin 2027 at a 1-micrometer pitch and more than 100 million interconnects; for around three years out, a 720-nanometer pitch with over 200 million. It is targeting a 1.4nm-class density by 2031, with LogicFolding in the Ascend 990 AI accelerator around 2030. HiSilicon, Huawei’s chip-design unit, designs the Kirin chips, and SMIC is believed to make them, according to Reuters via <em>TrendForce</em>. It isn’t stated who handles the bonding process.</p><p>As for the Kirin 9050 Pro, Geekerwan tested it in a Mate 90 Pro Max phone. Analysis shows that the two dies are bonded with copper-to-copper hybrid bonding, metal layer to metal layer. The bonded circuit layers of both dies sit between two thick layers of silicon rather than directly on the package substrate. For the interconnects, about 80,000 through-silicon vias (TSVs) run through the lower die, taking up, with their keep-out zones, about 8% of the lower die’s usable area, Geekerwan says.</p><div class="youtube-video" data-nosnippet ><div class="video-aspect-box"><iframe data-lazy-priority="high" data-lazy-src="https://www.youtube-nocookie.com/embed/PpNFmPcBYPY" allowfullscreen=""></iframe></div></div><p>The split spans both dies. In the big CPU core, the execution units are on top, with the L1 and L2 caches directly beneath the load/store units, Geekerwan says. Kurnal's die shots put the GPU and NPU logic over their caches too. The lower die also holds a shared L3, a 12MiB system cache, PLLs, and the LPDDR, PCIe, camera, display, UFS, and USB interfaces, while the top die carries the ISP, modem, and memory controllers. Cache, I/O, and PLLs produce less heat and are less sensitive to the process, so they go below, Geekerwan explains; the denser compute goes on top, on a somewhat more advanced process.</p><p>Each die is just over 120 mm², which is smaller than the previous chip’s 140-plus mm². The total silicon area for both is therefore over 240 mm², up 70% overall. Density rose from 155 to 238 MTr/mm² at launch, about 53.5% higher, according to Huawei’s launch numbers. Other notable changes include the NPU’s area, up 150% across both dies, with Geekerwan measuring 68 TOPS INT8. The configuration makes sense and the rationale follows the layout, but an independent die analysis is needed for full understanding.</p><p>Huawei’s stacking approach has its competitors. <a href="https://www.tomshardware.com/tech-industry/semiconductors/hybrid-bonding-roadmap-examined"><u>TSMC’s SoIC</u></a> roadmap set a 6-micrometer bond pitch for 2025 and 4.5 micrometers by 2029, with 3nm chip stacking in volume production last year. Intel’s Foveros Direct is said to have “copper bonding at a pitch of 9µm,” which reached high volume with the Clearwater Forest Xeon, and a second generation targets 3 micrometers. <a href="https://www.tomshardware.com/news/amd-unveils-more-ryzen-3d-packaging-and-v-cache-details-at-hot-chips"><u>AMD’s first-generation 3D V-Cache</u></a> stacks a 64MB L3 SRAM die on a complete CPU die, using direct copper bonding at a 9-micrometer pitch. Huawei’s 1.5-micrometer claim comes from its own paper and any comparison has to be taken with a grain of salt.</p><p>As to why Qualcomm might be interested, John Han, executive vice president and general manager of Qualcomm Technology Licensing, said the agreement reflects Qualcomm’s “recognition of Huawei’s continued innovation and intellectual property in 5G and other technology fields.” Bloomberg’s report says that at least some of the patents underpin LogicFolding, but no source names the patents. Dr. Ian Cutress, chief analyst at More Than Moore, questioned how new Huawei’s Tau Scaling is, writing on X that it is “a repackaging of what the industry has always done.”</p><p>Upon “receipt of the necessary regulatory approvals,” the cross-license and patent purchase deal should proceed. The deal is also subject to Federal Trade Commission review under the Hart-Scott-Rodino Act, Bloomberg reported. That U.S. premerger law sets a minimum deal size, $133.9 million this year, among other tests, and the deal’s value wasn’t disclosed.</p><p><em>Nikkei Asia</em> reported that Qualcomm will be the net payer under the deal. However Qualcomm said on October 6 that reports characterizing it as a net payor on the license agreement are incorrect and that assertions linking the agreement to LogicFold are not accurate. Qualcomm’s licensing revenue stopped including Huawei royalties in fiscal Q2 2025, after the license expired. This follows a long history of Huawei paying for Qualcomm licenses, an agreement first inked in 2001. Huawei says it has brought in more license-related income than it’s spent since 2021. Patents behind its chip-stacking method reportedly join that business under the deal. </p><p>The industry now waits to see what comes of the FTC review. Qualcomm’s fiscal Q4 results should come early next month, as last year’s did, but its fiscal year ended Sept. 27, before the deal was announced. Additional independent analysis of the 9050 Pro’s die may come from <em>TechInsights</em> or <em>SemiAnalysis</em>. Also of interest is the next iteration in Huawei’s Kirin 2027, to see if its claims are met. The deal between the two companies currently lacks details on the extent of patent coverage, but the silicon on record can help reveal the nature of Qualcomm’s interest.</p><p><em><strong>Update: 10/12/2026, 02:50 PT: Added passage reflecting Qualcomm's October 6 statement.</strong></em></p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Gigaphoton debuts neon recycling system with claimed 50% recovery rate ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Japanese semiconductor lithography tool maker Gigaphoton has developed a novel neon gas recycling system named hTGM for its argon fluoride (ArF) excimer lasers. The method has demonstrated a 50% recycling rate, with hopes that it can be improved dramatically with adjusted configurations, <a href="https://www.businesswire.com/news/home/20260929911995/en/Gigaphoton-Develops-Gas-Recycling-System-for-ArF-Excimer-Lasers" target="_blank">Business Wire reports</a>.  This could help fill a gap in the chip supply chain caused by the Russian invasion of Ukraine. Before 2022, <a href="https://www.tomshardware.com/news/ukraine-neon-gas-production" target="_blank">Ukraine shipped up to 50% of the neon gas used</a> in the semiconductor industry, but its major producers had to stop exports, <a href="https://www.tomshardware.com/news/gas-used-to-make-semiconductors-threatened-by-russian-invasion-of-ukraine" target="_blank">raising concerns over chip supply</a>. 70% of global neon production is used in semiconductor manufacturing, particularly in advanced DUV processes.</p><p>In the aftermath, China, Taiwan, and South Korea all increased their own domestic Neon production efforts, but various chip manufacturers also integrated recycling systems at various points during the fabrication process. Gigaphoton's more efficient ArF lasers with neon recycling could further help close those loops, reducing material costs and further reducing reliance on global supply chains.</p><p>It's not the only company pushing this technology forward, either. At the start of September, Japanese environmental protection company Kanken Techno <a href="https://semicontaiwan.org/en/node/15936" target="_blank">demonstrated its own Neon gas recycling system</a> for excimer lasers at SemiCon Taiwan. It demonstrated a recycling rate of over 90%, suggesting there is considerable headroom to improve this process in the future.</p><h2 id="what-makes-neon-so-scarce">What makes Neon so scarce?</h2><p>Neon is one of a handful of noble gases, making it inert under standard conditions and easy to handle. However, the ease with which it vaporizes and its limited bonding ability mean it is relatively scarce in Earth's atmosphere, making it scarce. It is most efficiently obtained through fractional distillation of super-cooled air, making it most affordable to obtain as a byproduct of other processes. For decades, Ukrainian steel-mill air-separation units produced crude neon from large-scale air-separation units (ASUs), and then refined it in-country via a pair of key companies.</p><p>That meant that when Russia escalated its war against Ukraine into a full-scale invasion in 2022 and Ukrainian neon refiners shut down production, there was an immediate global response. </p><p>As <a href="https://www.sfa-oxford.com/rare-earths-and-minor-metals/gaseous-elements-and-noble-gases/electronic-gases/neon-10-ne/" target="_blank">SFA Oxford explains</a>, South Korea accelerated domestic production and implemented a large-scale, national neon recycling initiative focused on in-fab recovery. Taiwan installed purification systems in existing blast-furnace ASUs, developing its own domestic supply chain. By utilizing its national infrastructure, it targeted complete self-sufficiency in neon gas production within a few years. </p><p>China integrated neon gas purification into many elements of its industrial base, twinning it with steel and petrochemical production facilities. The U.S. also responded with the development of new ASUs at Gulf-coast petrochemical hubs, but those aren't expected to meaningfully improve American neon supplies until near the end of the decade.</p><p>Lithography accounts for around 70% of global neon use, particularly in more advanced DUV processes — though crucially, not in EUV lithography, which uses a tin-plasma laser instead. With global DUV-produced semiconductor supply rapidly advancing in the wake of the AI buildout, neon supply is an important component in the chip supply chain, which has rapidly evolved to become a key asset within strategic defence and global trade.</p><h2 id="working-with-what-you-39-ve-got">Working with what you've got</h2><p>For individual companies that can't directly control their supply chains, recycling has been a much more immediate focus. The first neon gas recycling systems were <a href="https://www.efcgases.com/news/efc-launches-cymer-qualified-neon-gas-recycling-system-a-game-changer-for-excimer-lasers/" target="_blank">qualified for use with Cymer excimer ArF lasers at the end of 2023</a>, and <a href="https://www.eecone.com/eecone/news/?id=202403005" target="_blank">Samsung and</a> SK Hynix became the first of the major memory makers to <a href="https://www.gasworld.com/story/industrys-first-neon-gas-recycling-technology-announced-by-sk-hynix-temc/2136766.article/" target="_blank">implement neon gas recycling in 2024</a>.</p><p>Gigaphoton previously offered neon recycling with KrF (krypton fluoride) lasers, but has now extended that recycling capability to ArF lasers as well. It claims a single hTGM for ArF unit can reduce neon gas consumption by up to 470 kilotres per year, representing substantial material cost savings, as well as reducing company reliance on a global supply chain that has been sufficiently unstable in recent years.</p><p>However, its claims of a 50% recycling rate aren't exactly groundbreaking. Cymer's excimer laser neon recovery system <a href="https://www.laserfocusworld.com/lasers-sources/article/16557989/cymer-second-generation-lithography-lasers-reduce-neon-consumption" target="_blank">demonstrated that same level of efficiency in 2015</a>, and <a href="https://ui.adsabs.harvard.edu/abs/2017SPIE10147E..1OY/abstract" target="_blank">Gigaphoton's own testing suggested a potential peak for ArF recovery of over 90% in 2017</a>.</p><p>That suggests it should be able to ramp up its efforts to a higher percentage quite swiftly. Especially since fellow Japanese firm Kanken Techno recently unveiled a neon recycling system with over 90% efficiency. Kanken works with both major ArF laser developers and is currently going through the certification process, with expected approval in the near future.</p><h2 id="a-temporary-fix">A temporary fix?</h2><p>With neon gas recycling during chip production effectively solved, it may be that the supply of neon for the chip industry has proved to be a temporary problem. But it may be a temporary solution too, as the use of neon within the industry may be set to shrink dramatically in the years to come.</p><p><a href="https://www.tomshardware.com/tech-industry/semiconductors/tsmc-to-start-using-high-na-euv-lithography-in-2030-a10-or-a11-technology-prime-candidates-for-use" target="_blank">EUV lithography</a> can build far more complex chips than DUV, and offers a future path beyond the need for neon gas usage in excimer lasers as newer electronics adopt the more capable chip designs. On top of that, alternative technologies to DUV are being developed which could produce the same DUV light to generate advanced semiconductors, but without the need for noble or toxic gases. </p><p>Solid-state laser technologies have been rapidly developing in recent years, though <a href="https://www.tomshardware.com/tech-industry/chinese-scientists-create-solid-state-duv-laser-sources-for-lithography-equipment-used-in-chip-manufacturing" target="_blank">last year's Chinese efforts</a> are still below required industrial power thresholds. <a href="https://www.tomshardware.com/tech-industry/semiconductors/chinese-startup-claims-photonic-chip-production-without-duv-lithography-says-nanoimprint-process-cuts-costs-by-90-percent-8-inch-wafers-produced-without-conventional-optical-lithography" target="_blank">Nanoimprint lithography also holds promise</a> as a DUV alternative, as well as cutting costs by a comparative 90%.</p><p>There are still huge hurdles in performance and certification, as well as integration with existing fabrication plans and supply chains. But the potential to move beyond the need for neon gas within semiconductor lithography is there. Over the coming years, the combination of laser neon recycling systems and alternative manufacturing techniques could reduce the need for international neon sources to a fraction of existing levels.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/gigaphoton-debuts-neon-recycling-system-with-claimed-50-percent-recovery-rate-systems-throw-a-lifeline-to-chipmakers-that-utilize-70-percent-of-global-neon-supply-in-duv-lithography</link>
                                                                            <description>
                            <![CDATA[ New neon gas recycling systems promise to reduce the demand for the noble gas at major chip manufacturers using DUV lithography. But as that technology is supplanted, this much-needed fix may only be needed temporarily. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">oXsA9zh9kwzb3dbsefHcZJ</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/QTsaWMMDmqTRrNAmFg4JPo-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Tue, 06 Oct 2026 13:20:00 +0000</pubDate>                                                                                                                                <updated>Wed, 07 Oct 2026 14:16:18 +0000</updated>
                                                                                                                                            <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Jon Martindale ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/YeutDv8zJmhi7xH35MSt8Z-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;After building his first computers in his teens, Jon Martindale has spent the past two decades covering the latest advances in technology. From displays to PC components, blockchain to AI, and tablets to standing desk accessories, Jon has covered just about every facet of the tech space in his varied career. He has bylines at Forbes, USNews, Lifewire, DigitalTrends, PCWorld, and a range of other sites. He brings that same level of expertise and professional insight to Toms Hardware.Away from writing, Jon is an avid reader, board gamer, and fitness enthusiast. He lives in rural Gloucestershire with his wife, two children, and French Bulldog cross.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/QTsaWMMDmqTRrNAmFg4JPo-1920-80.jpg">
                                                            <media:credit><![CDATA[Jens SCHLÜTER / AFP via Getty Images]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[A worker with a wafer polishing machine.]]></media:description>                                                            <media:text><![CDATA[A worker with a wafer polishing machine.]]></media:text>
                                <media:title type="plain"><![CDATA[A worker with a wafer polishing machine.]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/QTsaWMMDmqTRrNAmFg4JPo-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Japanese semiconductor lithography tool maker Gigaphoton has developed a novel neon gas recycling system named hTGM for its argon fluoride (ArF) excimer lasers. The method has demonstrated a 50% recycling rate, with hopes that it can be improved dramatically with adjusted configurations, <a href="https://www.businesswire.com/news/home/20260929911995/en/Gigaphoton-Develops-Gas-Recycling-System-for-ArF-Excimer-Lasers" target="_blank">Business Wire reports</a>.  This could help fill a gap in the chip supply chain caused by the Russian invasion of Ukraine. Before 2022, <a href="https://www.tomshardware.com/news/ukraine-neon-gas-production" target="_blank">Ukraine shipped up to 50% of the neon gas used</a> in the semiconductor industry, but its major producers had to stop exports, <a href="https://www.tomshardware.com/news/gas-used-to-make-semiconductors-threatened-by-russian-invasion-of-ukraine" target="_blank">raising concerns over chip supply</a>. 70% of global neon production is used in semiconductor manufacturing, particularly in advanced DUV processes.</p><p>In the aftermath, China, Taiwan, and South Korea all increased their own domestic Neon production efforts, but various chip manufacturers also integrated recycling systems at various points during the fabrication process. Gigaphoton's more efficient ArF lasers with neon recycling could further help close those loops, reducing material costs and further reducing reliance on global supply chains.</p><p>It's not the only company pushing this technology forward, either. At the start of September, Japanese environmental protection company Kanken Techno <a href="https://semicontaiwan.org/en/node/15936" target="_blank">demonstrated its own Neon gas recycling system</a> for excimer lasers at SemiCon Taiwan. It demonstrated a recycling rate of over 90%, suggesting there is considerable headroom to improve this process in the future.</p><h2 id="what-makes-neon-so-scarce">What makes Neon so scarce?</h2><p>Neon is one of a handful of noble gases, making it inert under standard conditions and easy to handle. However, the ease with which it vaporizes and its limited bonding ability mean it is relatively scarce in Earth's atmosphere, making it scarce. It is most efficiently obtained through fractional distillation of super-cooled air, making it most affordable to obtain as a byproduct of other processes. For decades, Ukrainian steel-mill air-separation units produced crude neon from large-scale air-separation units (ASUs), and then refined it in-country via a pair of key companies.</p><p>That meant that when Russia escalated its war against Ukraine into a full-scale invasion in 2022 and Ukrainian neon refiners shut down production, there was an immediate global response. </p><p>As <a href="https://www.sfa-oxford.com/rare-earths-and-minor-metals/gaseous-elements-and-noble-gases/electronic-gases/neon-10-ne/" target="_blank">SFA Oxford explains</a>, South Korea accelerated domestic production and implemented a large-scale, national neon recycling initiative focused on in-fab recovery. Taiwan installed purification systems in existing blast-furnace ASUs, developing its own domestic supply chain. By utilizing its national infrastructure, it targeted complete self-sufficiency in neon gas production within a few years. </p><p>China integrated neon gas purification into many elements of its industrial base, twinning it with steel and petrochemical production facilities. The U.S. also responded with the development of new ASUs at Gulf-coast petrochemical hubs, but those aren't expected to meaningfully improve American neon supplies until near the end of the decade.</p><p>Lithography accounts for around 70% of global neon use, particularly in more advanced DUV processes — though crucially, not in EUV lithography, which uses a tin-plasma laser instead. With global DUV-produced semiconductor supply rapidly advancing in the wake of the AI buildout, neon supply is an important component in the chip supply chain, which has rapidly evolved to become a key asset within strategic defence and global trade.</p><h2 id="working-with-what-you-39-ve-got">Working with what you've got</h2><p>For individual companies that can't directly control their supply chains, recycling has been a much more immediate focus. The first neon gas recycling systems were <a href="https://www.efcgases.com/news/efc-launches-cymer-qualified-neon-gas-recycling-system-a-game-changer-for-excimer-lasers/" target="_blank">qualified for use with Cymer excimer ArF lasers at the end of 2023</a>, and <a href="https://www.eecone.com/eecone/news/?id=202403005" target="_blank">Samsung and</a> SK Hynix became the first of the major memory makers to <a href="https://www.gasworld.com/story/industrys-first-neon-gas-recycling-technology-announced-by-sk-hynix-temc/2136766.article/" target="_blank">implement neon gas recycling in 2024</a>.</p><p>Gigaphoton previously offered neon recycling with KrF (krypton fluoride) lasers, but has now extended that recycling capability to ArF lasers as well. It claims a single hTGM for ArF unit can reduce neon gas consumption by up to 470 kilotres per year, representing substantial material cost savings, as well as reducing company reliance on a global supply chain that has been sufficiently unstable in recent years.</p><p>However, its claims of a 50% recycling rate aren't exactly groundbreaking. Cymer's excimer laser neon recovery system <a href="https://www.laserfocusworld.com/lasers-sources/article/16557989/cymer-second-generation-lithography-lasers-reduce-neon-consumption" target="_blank">demonstrated that same level of efficiency in 2015</a>, and <a href="https://ui.adsabs.harvard.edu/abs/2017SPIE10147E..1OY/abstract" target="_blank">Gigaphoton's own testing suggested a potential peak for ArF recovery of over 90% in 2017</a>.</p><p>That suggests it should be able to ramp up its efforts to a higher percentage quite swiftly. Especially since fellow Japanese firm Kanken Techno recently unveiled a neon recycling system with over 90% efficiency. Kanken works with both major ArF laser developers and is currently going through the certification process, with expected approval in the near future.</p><h2 id="a-temporary-fix">A temporary fix?</h2><p>With neon gas recycling during chip production effectively solved, it may be that the supply of neon for the chip industry has proved to be a temporary problem. But it may be a temporary solution too, as the use of neon within the industry may be set to shrink dramatically in the years to come.</p><p><a href="https://www.tomshardware.com/tech-industry/semiconductors/tsmc-to-start-using-high-na-euv-lithography-in-2030-a10-or-a11-technology-prime-candidates-for-use" target="_blank">EUV lithography</a> can build far more complex chips than DUV, and offers a future path beyond the need for neon gas usage in excimer lasers as newer electronics adopt the more capable chip designs. On top of that, alternative technologies to DUV are being developed which could produce the same DUV light to generate advanced semiconductors, but without the need for noble or toxic gases. </p><p>Solid-state laser technologies have been rapidly developing in recent years, though <a href="https://www.tomshardware.com/tech-industry/chinese-scientists-create-solid-state-duv-laser-sources-for-lithography-equipment-used-in-chip-manufacturing" target="_blank">last year's Chinese efforts</a> are still below required industrial power thresholds. <a href="https://www.tomshardware.com/tech-industry/semiconductors/chinese-startup-claims-photonic-chip-production-without-duv-lithography-says-nanoimprint-process-cuts-costs-by-90-percent-8-inch-wafers-produced-without-conventional-optical-lithography" target="_blank">Nanoimprint lithography also holds promise</a> as a DUV alternative, as well as cutting costs by a comparative 90%.</p><p>There are still huge hurdles in performance and certification, as well as integration with existing fabrication plans and supply chains. But the potential to move beyond the need for neon gas within semiconductor lithography is there. Over the coming years, the combination of laser neon recycling systems and alternative manufacturing techniques could reduce the need for international neon sources to a fraction of existing levels.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ OpenAI and Synopsys partner to build "GPT-Synopsys" for autonomous chip design ]]></title>
                                                                                                <dc:content><![CDATA[ <p>OpenAI and Synopsys have signed a multi-year agreement to jointly develop GPT-Synopsys, a specialized model optimized for semiconductor design using Synopsys's electronic design automation (EDA) tools. According to the September 30 <a href="https://investor.synopsys.com/news/news-details/2026/OpenAI-and-Synopsys-Announce-GPT-Synopsys-Frontier-Intelligence-to-Revolutionize-Chip-Design/default.aspx" target="_blank">announcement</a>, the partnership “brings together OpenAI's advanced AI capabilities with Synopsys's industry-leading EDA tools and agentic AI capabilities to revolutionize the design of semiconductors.”</p><p>Engineers would delegate to agents that run the tools and execute the required processes until the work is ready for review. The companies say this would allow design teams to evaluate more options and deliver more sophisticated chips. Under the preferred partner agreement, OpenAI is licensing Synopsys's tools to develop the model — which will run on OpenAI-hosted infrastructure — after which the companies will jointly release the product and share revenue. The announcement did not disclose a release date or a pricing structure.</p><h2 id="ai-in-chip-design-so-far">AI in chip design so far</h2><p>Engineers use EDA tools to design and verify chips before manufacturing. This multi-step process begins with writing the hardware description in register-transfer level (RTL) code and using synthesis software to convert it into logic gates. Next, physical design tools lay out the circuit elements and route the connections between them. Lastly, designers close timing, verify functionality, and ensure manufacturing compliance before tapeout for fabrication.</p><p>Each of these stages requires several iterations — repeatedly running various tools and troubleshooting issues — to balance critical trade-offs: power, performance, and silicon area (PPA). Synopsys, Cadence, and Siemens EDA have dominated this market for these tools long before the current AI boom. Given the technology's capabilities, it is only natural that <a href="https://www.tomshardware.com/tech-industry/semiconductors/silicon-is-starting-to-design-silicon-how-ai-is-being-used-in-chipmaking-from-eda-tools-to-openais-jalapeno-and-beyond" target="_blank">AI has found its way into the chip design process</a>, creating a sort of “silicon designing silicon” loop. </p><p>In March 2020, Synopsys launched DSO.ai (Design Space Optimization AI), a reinforcement learning tool that explores and learns from previous design optimizations to improve PPA. The company then launched Synopsys.ai Copilot in November 2023 — its first integration of generative AI — through a collaboration with Microsoft, integrating the Azure OpenAI service to provide natural language assistance within its engineering tools. </p><p>The company’s current direction is Agentic AI, first revealed at the Design Automation Conference in July, where it showed an autonomous verification workflow built on Nvidia's Agent Toolkit and Nemotron 3 Ultra model. On September 28, two days before the OpenAI deal, <a href="https://www.tomshardware.com/tech-industry/semiconductors/synopsys-debuts-autopilot-platform-for-developing-chips-autonomously-using-ai-new-agentengineer-platform-is-poised-for-general-availability-by-the-end-of-2026" target="_blank">Synopsys announced its Autopilot platform and AgentEngineer</a>, a portfolio of seven long-horizon agents covering verification, implementation, analog, manufacturing, meshing, combustion, and EMC analysis. The company reported more than 50 engagements, with availability planned for late 2026.</p><p>As detailed in our <a href="https://www.tomshardware.com/tech-industry/semiconductors/the-state-of-agentic-ai-in-chip-design-tools-in-2026-cadence-synopsys-and-siemens-all-pitch-autonomous-engineers" target="_blank">State of agentic AI in chip design tools</a> roadmap, the move to agentic AI is not exclusive to Synopsys; Cadence and Siemens are also pitching autonomous design agents. Meanwhile, OpenAI, in collaboration with Broadcom, unveiled Jalapeño in June, the company’s first custom inference accelerator. <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-jalapeno-design-interview-transcript-hardware-vp-richard-ho-explains-how-ai-assisted-design-may-shape-the-future-of-inference-asics" target="_blank">OpenAI said the accelerator's design and optimization process leaned heavily on its AI models</a>, allowing it to go from initial design to tape-out in just nine months. Now, GPT-Synopsys turns that internal experiment into a product.</p><h2 id="gpt-synopsys">GPT-Synopsys</h2><p>According to the announcement, GPT-Synopsys will combine OpenAI's frontier models with Synopsys’s EDA software and domain expertise, allowing the model to reason about chip design and verification and operate Synopsys tools directly. Engineers would hand the model objectives such as PPA optimization, timing, and verification closure, while the agents run the tools, interpret the results, implement changes, and iterate toward verified outcomes for an engineer to review.</p><p>In other words, the model would be capable of making engineering judgments involved in using Synopsys software. The companies describe it as learning to operate the tools like an expert engineer, interpreting their outputs and using the results to guide further changes. This is the proposed specialization that takes the model beyond connecting a general-purpose model to a set of tools.</p><p>Synopsys’s EDA engines would perform the calculations and checks, with the model interpreting results and the agent software managing execution. GPT-Synopsys will run on OpenAI-hosted infrastructure and is intended to integrate with Synopsys.ai and Autopilot, as well as work with customers’ own agent harnesses. Autopilot already provides services such as memory and governance.</p><p>Synopsys says early technology engagements are underway with leading semiconductor customers, but didn't name any. The immediate audience is professional semiconductor-design teams, with companies developing their own custom silicon also likely to use it. Although the announcement covers semiconductor design generally, OpenAI has a particular interest in improving the hardware that runs its models. “By helping them build better chips, we can build better AI and bring it to more people,” said Greg Brockman, OpenAI’s president and co-founder. Whether the service also makes advanced design more accessible to smaller teams will depend on pricing and the expertise still required to supervise it.</p><h2 id="questions-and-concerns">Questions and concerns</h2><p>The recent announcement leaves a couple of questions and concerns unaddressed. First, hosting the model on OpenAI infrastructure raises concerns about how confidential designs are handled. The companies say customer data will be excluded from model training, encrypted at rest and in transit, and governed by configurable retention, audit, and permission controls. However, the release does not identify hosting regions or default retention periods. It also provides no contractual terms for ownership of generated outputs, use of third-party licensed design IP, or indemnities. The service's terms will likely address these questions.</p><p>Another question is the practical economics for users. The industry has realized that <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/talent-over-tokens-ai-models-are-becoming-more-expensive-to-run-and-productivity-gains-are-limited-efficient-workers-might-be-the-solution-to-strained-budgets" target="_blank">handing over everything to AI doesn't always result in cost savings</a>, at least for now. Sometimes the reverse is the case.  Earlier this year, Uber’s CTO and an Nvidia executive said <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-exec-says-ai-is-more-expensive-than-actual-workers-yet-some-companies-dont-see-the-extra-costs-as-a-negative" target="_blank">AI was more expensive than human workers</a>, although many companies are reportedly fine with the extra cost. For GPT-Synopsys, advanced engineering reasoning and faster design-to-tapeout could justify the service, but model calls, EDA runs, integration, and human review all consume resources. The announcement did not include a pricing structure. It also didn't specify a launch date.</p><p>Meanwhile, Synopsys’s main rival Cadence launched its<a href="https://www.tomshardware.com/tech-industry/semiconductors/cadence-embeds-ai-across-its-eda-portfolio" target="_blank"> ChipStack AI Super Agent</a> in February for front-end design and verification, built on frontier LLMs. In April, the company announced a collaboration with Google to optimize ChipStack with Gemini on Google Cloud. At Computex, it extended ChipStack to what it calls Level-5 autonomy, powered by Nvidia's Nemotron models, with early access expected in the second half of 2026. The main difference is that while ChipStack is Cadence’s own agent software running on other companies’ general-purpose models, GPT-Synopsys will be an OpenAI model specifically trained to operate Synopsys’s tools.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-and-synopsys-partner-to-build-gpt-synopsys-for-autonomous-chip-design-specialized-ai-model-will-operate-eda-tools-allowing-engineers-to-deliver-more-sophisticated-chips-faster</link>
                                                                            <description>
                            <![CDATA[ OpenAI and Synopsys are developing GPT-Synopsys, a specialized semiconductor-design model that will directly operate Synopsys EDA tools ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">7ozoPaiHoiU85DsS253ZmV</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/L5bu9i78ZpqMnudoVp2DSk-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Tue, 06 Oct 2026 11:00:00 +0000</pubDate>                                                                                                                                <updated>Tue, 06 Oct 2026 12:11:31 +0000</updated>
                                                                                                                                            <category><![CDATA[Artificial Intelligence]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Etiido Uko ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/BBrMt7jWtSo2Dc3iKoroyD-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Etiido Uko is a mechanical engineer and senior technical writer with over nine years of experience in documentation and reporting. He is deeply passionate about all things engineering and technology, and is an expert in gadgets, manufacturing, robotics, automotive, and aerospace. His work spans content creation for industry leaders across multiple sectors, including Autodesk, Siemens, Xometry, Telus, and Coca-Cola. When he is not writing or keeping up with the latest innovations, you can find him exploring lands unknown. Check out more of his work at etiidowrites.com.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/L5bu9i78ZpqMnudoVp2DSk-1920-80.jpg">
                                                            <media:credit><![CDATA[Synopsys]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[GPT-Synopsys]]></media:description>                                                            <media:text><![CDATA[GPT-Synopsys]]></media:text>
                                <media:title type="plain"><![CDATA[GPT-Synopsys]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/L5bu9i78ZpqMnudoVp2DSk-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>OpenAI and Synopsys have signed a multi-year agreement to jointly develop GPT-Synopsys, a specialized model optimized for semiconductor design using Synopsys's electronic design automation (EDA) tools. According to the September 30 <a href="https://investor.synopsys.com/news/news-details/2026/OpenAI-and-Synopsys-Announce-GPT-Synopsys-Frontier-Intelligence-to-Revolutionize-Chip-Design/default.aspx" target="_blank">announcement</a>, the partnership “brings together OpenAI's advanced AI capabilities with Synopsys's industry-leading EDA tools and agentic AI capabilities to revolutionize the design of semiconductors.”</p><p>Engineers would delegate to agents that run the tools and execute the required processes until the work is ready for review. The companies say this would allow design teams to evaluate more options and deliver more sophisticated chips. Under the preferred partner agreement, OpenAI is licensing Synopsys's tools to develop the model — which will run on OpenAI-hosted infrastructure — after which the companies will jointly release the product and share revenue. The announcement did not disclose a release date or a pricing structure.</p><h2 id="ai-in-chip-design-so-far">AI in chip design so far</h2><p>Engineers use EDA tools to design and verify chips before manufacturing. This multi-step process begins with writing the hardware description in register-transfer level (RTL) code and using synthesis software to convert it into logic gates. Next, physical design tools lay out the circuit elements and route the connections between them. Lastly, designers close timing, verify functionality, and ensure manufacturing compliance before tapeout for fabrication.</p><p>Each of these stages requires several iterations — repeatedly running various tools and troubleshooting issues — to balance critical trade-offs: power, performance, and silicon area (PPA). Synopsys, Cadence, and Siemens EDA have dominated this market for these tools long before the current AI boom. Given the technology's capabilities, it is only natural that <a href="https://www.tomshardware.com/tech-industry/semiconductors/silicon-is-starting-to-design-silicon-how-ai-is-being-used-in-chipmaking-from-eda-tools-to-openais-jalapeno-and-beyond" target="_blank">AI has found its way into the chip design process</a>, creating a sort of “silicon designing silicon” loop. </p><p>In March 2020, Synopsys launched DSO.ai (Design Space Optimization AI), a reinforcement learning tool that explores and learns from previous design optimizations to improve PPA. The company then launched Synopsys.ai Copilot in November 2023 — its first integration of generative AI — through a collaboration with Microsoft, integrating the Azure OpenAI service to provide natural language assistance within its engineering tools. </p><p>The company’s current direction is Agentic AI, first revealed at the Design Automation Conference in July, where it showed an autonomous verification workflow built on Nvidia's Agent Toolkit and Nemotron 3 Ultra model. On September 28, two days before the OpenAI deal, <a href="https://www.tomshardware.com/tech-industry/semiconductors/synopsys-debuts-autopilot-platform-for-developing-chips-autonomously-using-ai-new-agentengineer-platform-is-poised-for-general-availability-by-the-end-of-2026" target="_blank">Synopsys announced its Autopilot platform and AgentEngineer</a>, a portfolio of seven long-horizon agents covering verification, implementation, analog, manufacturing, meshing, combustion, and EMC analysis. The company reported more than 50 engagements, with availability planned for late 2026.</p><p>As detailed in our <a href="https://www.tomshardware.com/tech-industry/semiconductors/the-state-of-agentic-ai-in-chip-design-tools-in-2026-cadence-synopsys-and-siemens-all-pitch-autonomous-engineers" target="_blank">State of agentic AI in chip design tools</a> roadmap, the move to agentic AI is not exclusive to Synopsys; Cadence and Siemens are also pitching autonomous design agents. Meanwhile, OpenAI, in collaboration with Broadcom, unveiled Jalapeño in June, the company’s first custom inference accelerator. <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-jalapeno-design-interview-transcript-hardware-vp-richard-ho-explains-how-ai-assisted-design-may-shape-the-future-of-inference-asics" target="_blank">OpenAI said the accelerator's design and optimization process leaned heavily on its AI models</a>, allowing it to go from initial design to tape-out in just nine months. Now, GPT-Synopsys turns that internal experiment into a product.</p><h2 id="gpt-synopsys">GPT-Synopsys</h2><p>According to the announcement, GPT-Synopsys will combine OpenAI's frontier models with Synopsys’s EDA software and domain expertise, allowing the model to reason about chip design and verification and operate Synopsys tools directly. Engineers would hand the model objectives such as PPA optimization, timing, and verification closure, while the agents run the tools, interpret the results, implement changes, and iterate toward verified outcomes for an engineer to review.</p><p>In other words, the model would be capable of making engineering judgments involved in using Synopsys software. The companies describe it as learning to operate the tools like an expert engineer, interpreting their outputs and using the results to guide further changes. This is the proposed specialization that takes the model beyond connecting a general-purpose model to a set of tools.</p><p>Synopsys’s EDA engines would perform the calculations and checks, with the model interpreting results and the agent software managing execution. GPT-Synopsys will run on OpenAI-hosted infrastructure and is intended to integrate with Synopsys.ai and Autopilot, as well as work with customers’ own agent harnesses. Autopilot already provides services such as memory and governance.</p><p>Synopsys says early technology engagements are underway with leading semiconductor customers, but didn't name any. The immediate audience is professional semiconductor-design teams, with companies developing their own custom silicon also likely to use it. Although the announcement covers semiconductor design generally, OpenAI has a particular interest in improving the hardware that runs its models. “By helping them build better chips, we can build better AI and bring it to more people,” said Greg Brockman, OpenAI’s president and co-founder. Whether the service also makes advanced design more accessible to smaller teams will depend on pricing and the expertise still required to supervise it.</p><h2 id="questions-and-concerns">Questions and concerns</h2><p>The recent announcement leaves a couple of questions and concerns unaddressed. First, hosting the model on OpenAI infrastructure raises concerns about how confidential designs are handled. The companies say customer data will be excluded from model training, encrypted at rest and in transit, and governed by configurable retention, audit, and permission controls. However, the release does not identify hosting regions or default retention periods. It also provides no contractual terms for ownership of generated outputs, use of third-party licensed design IP, or indemnities. The service's terms will likely address these questions.</p><p>Another question is the practical economics for users. The industry has realized that <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/talent-over-tokens-ai-models-are-becoming-more-expensive-to-run-and-productivity-gains-are-limited-efficient-workers-might-be-the-solution-to-strained-budgets" target="_blank">handing over everything to AI doesn't always result in cost savings</a>, at least for now. Sometimes the reverse is the case.  Earlier this year, Uber’s CTO and an Nvidia executive said <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-exec-says-ai-is-more-expensive-than-actual-workers-yet-some-companies-dont-see-the-extra-costs-as-a-negative" target="_blank">AI was more expensive than human workers</a>, although many companies are reportedly fine with the extra cost. For GPT-Synopsys, advanced engineering reasoning and faster design-to-tapeout could justify the service, but model calls, EDA runs, integration, and human review all consume resources. The announcement did not include a pricing structure. It also didn't specify a launch date.</p><p>Meanwhile, Synopsys’s main rival Cadence launched its<a href="https://www.tomshardware.com/tech-industry/semiconductors/cadence-embeds-ai-across-its-eda-portfolio" target="_blank"> ChipStack AI Super Agent</a> in February for front-end design and verification, built on frontier LLMs. In April, the company announced a collaboration with Google to optimize ChipStack with Gemini on Google Cloud. At Computex, it extended ChipStack to what it calls Level-5 autonomy, powered by Nvidia's Nemotron models, with early access expected in the second half of 2026. The main difference is that while ChipStack is Cadence’s own agent software running on other companies’ general-purpose models, GPT-Synopsys will be an OpenAI model specifically trained to operate Synopsys’s tools.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Tencent scores 100,000 offshore AI chip deal with Oracle for $7 billion despite climbing prices  ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Oracle has reportedly made a deal with Tencent for a five-year lease at several Oracle data centers in Southeast Asia, covering about 100,000 advanced AI chips unavailable in China. The deal is estimated to be worth about $7 billion, which would come out to about $1.60 per chip per hour over five years, with about 30% upfront, according to a report by the <a href="https://www.ft.com/content/8799b33d-f07c-4a03-82f0-bf5d3d1d29e9"><u><em>Financial Times</em></u></a><em> </em>(<em>FT</em>). The deal lands as compute rental prices continue to climb, per Chief Strategy Officer James Mitchell on Tencent’s August 12 earnings call. Neither company has commented on the report, but such leases are legal under current U.S. rules, per the <em>FT</em>.</p><p>The estimated chip-hour price is around 43% less than the roughly $2.80 per GPU-hour for Nvidia’s H100 with a one-year contract, per research firm <em>SemiAnalysis </em>in August. It’s also about 48% lower than the $3.09 per GPU-hour for a three-year contract with Nvidia’s B200, per a 2025 filing by Japanese data center operator Datasection for an unnamed customer. Market trends have pricing running the other way, with the <em>FT </em>reporting longer terms and larger upfront payments for such leases. Tencent president Martin Lau said on the same call the company could sell its existing orders at “more than 30% profit” compared to what it “paid just a few months ago.”</p><p>No chips are named. The <em>FT  </em>says that such leases could reach even Nvidia’s top processors, but that does not mean this deal involves Nvidia hardware. One-year H100 contracts on <em>SemiAnalysis’s </em>index were above the estimated rate at about $1.70 per GPU-hour even as of October 2025, with the B200 rate being even more expensive despite its longer, three-year term. Going by raw B200 server costs, which exclude buildings, power, networking, and financing, 100,000 B200s would cost $5.44 billion, or about 78% of the reported value, based on Datasection’s 2025 purchase of servers with 5,000 B200s for $272 million.</p><p>Without specifics on the hardware being rented, it is not possible to determine whether this is a bulk deal or if the leasing costs are closer to the actual hardware cost. The pricing on paper, however, better fits Hopper-class at volume. Hopper would include the H100 and H200 and, with <a href="https://www.tomshardware.com/pc-components/gpus/first-nvidia-h200-shipments-reach-bytedance-and-tencent-as-beijing-loosens-its-import-block"><u>H200 imports still limited</u></a>, could meet the criterion of the chips in the deal not being available in China.</p><p>The Oracle data centers are located outside China, with Tencent merely renting time on them. Details including the countries involved, sites, and start date remain undisclosed. Oracle has two Singapore cloud regions, with a Malaysia one announced. The <em>FT </em>says the biggest clients of Southeast Asian data centers are <a href="https://www.tomshardware.com/tech-industry/semiconductors/chinas-top-ai-firms-shift-model-training-overseas-to-access-nvidia-gpus"><u>ByteDance and Alibaba</u></a>, with Chinese firms using the capacity especially for AI training, which domestic chips cannot yet do. Tencent previously inked a $1.2 billion-plus deal for Datasection Blackwell rentals in Japan and Australia, the FT reported in December 2025.</p><p>Some Oracle clients have prepaid for GPUs or brought their own hardware since its March call, and fiscal Q4, from March to May, added $67 billion of AI contracts, mostly of those two kinds, per its June call. June through August, per its quarterly filing, the company “received $11.4 billion of prepayments from customers that included a significant financing component.” Deferred revenue rose from $15.4 billion to $30.8 billion over the quarter. No customers were named, so this can’t be tied to Tencent. The company reported 97.9% GPU utilization during its September 10 earnings call, with renewals at a 20% average premium, which was mostly on GPUs four years or older.</p><p>Tencent prepaying about 30% makes sense given Oracle’s statements earlier this year. It’s a good deal for both parties. Oracle would improve its cash flow following a quarter of negative free cash flow, and Tencent Cloud’s May price raises, per Mitchell, could help recoup the cost of the rented chips.</p><p>The legal picture, given the friction between the United States and China over AI export controls, is thorny. The Remote Access Security Act, a bipartisan bill that <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/u-s-house-passes-bill-to-stop-chinese-companies-from-accessing-export-controlled-american-ai-chips-using-offshore-rental-loophole-remote-access-security-access-act-effectively-extends-export-controls-to-the-cloud"><u>passed the House</u></a> 369–22 in January, would give the U.S. government the right to regulate remote cloud access to sensitive technology, which could cover the chips under the reported deal. The bill specifically targets cloud loopholes and applies to any foreign person rather than naming countries, to prevent adversaries from leveraging cloud networks to bypass U.S. export bans. The bill currently sits in the Senate Banking Committee.</p><p>Separately, Export Administration Regulations license requirements under the Department of Commerce’s Bureau of Industry and Security (BIS) have covered shipments of advanced AI chips, including to China-headquartered companies abroad, since November 2023. Cloud rentals were left out, with the BIS only asking for comment on them. Renting remains one loophole around <a href="https://www.tomshardware.com/tech-industry/us-closes-loophole-that-allowed-chinese-owned-subsidiaries-located-outside-china-to-buy-ai-chips-report-claims-that-hundreds-of-thousands-of-advanced-ai-chips-have-been-acquired-through-bis-blind-spot"><u>BIS guidance on restricted chips</u></a>. <em>Bloomberg </em>reported that the BIS was reviewing offshore access with legal rentals included. The deal is also not barred by Tencent being on the <a href="https://www.tomshardware.com/pc-components/dram/us-dod-adds-cxmt-catl-tencent-to-list-of-companies-suspected-of-aiding-the-chinese-military"><u>Pentagon’s 1260H list</u></a> of “Chinese military companies,” as this is not a sanction. No statement about the deal by any regulatory government official accompanied the report.</p><p>According to <em>The Information</em>, the Department of Commerce is <a href="https://www.tomshardware.com/tech-industry/policy/new-us-export-controls-reportedly-target-chinese-access-to-remote-ai-servers-trump-admins-cut-down-ai-diffusion-rule-could-be-shared-with-industry-as-soon-as-september"><u>drafting a rule</u></a> to block Chinese firms from renting compute in third countries, such as Thailand and Singapore, which would align with the reported details of this deal. U.S. President Donald Trump and Chinese President Xi Jinping met late last month in Washington, covering AI and extending a trade truce for a further two months.</p><p>However, the deal is of a type that might face trouble from both the House-passed bill and Commerce’s reported rule. But the latter cannot be enforced under current law, an attorney at Baker McKenzie told <em>The Information</em>. At the summit, the two leaders agreed to set up a channel to handle AI-related incidents, but nothing was announced on chip controls. While the U.S. has cleared some firms, including Tencent, to buy Nvidia’s H200, China has let in only about 10,000 for Tencent so far, the <em>FT </em>reported in August. The next meeting between the two, at APEC in Shenzhen in November, may clarify the situation.</p><p>The reported deal comes weeks before Oracle’s Investor Day on Oct. 28 in Las Vegas, during its Oct. 25–28 Oracle AI World event at the Venetian. That would be a chance for the company to disclose details about the deal, such as announcing what hardware is to be leased. Given the hardware math, our assumption is that it involves older or non-Nvidia hardware rather than an action born of financial stress at Oracle. The month following Oracle’s event, Tencent is expected to report its Q3 results. These events could also reveal more concrete details about the arrangement.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/data-centers/tencent-scores-100-000-offshore-ai-chip-deal-with-oracle-for-usd7-billion-despite-climbing-prices-per-hour-costs-estimated-to-be-43-percent-under-standard-h100-rental-rates</link>
                                                                            <description>
                            <![CDATA[ Tencent is reportedly renting 100,000 AI chips from Oracle data centers in Southeast Asia for about $7 billion over five years. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">z8uXFHnY8W6Y5SDx48wz2V</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/T6qHP3TtmGJ8MCdVA94feF-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Mon, 05 Oct 2026 13:20:00 +0000</pubDate>                                                                                                                                <updated>Mon, 05 Oct 2026 16:08:54 +0000</updated>
                                                                                                                                            <category><![CDATA[Data Centers]]></category>
                                                    <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Shane Downing ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/Zosi9VrDytS9FkgJiHvc69-320-70.png ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;Shane has a background in computer engineering and has worked as a freelance consultant in multiple industries. He has a strong affection for history and loves to game. He worked his way up from a Commodore 64 and has always been interested in technology and writing. He particularly enjoys breaking down complex concepts into understandable ideas. He is a lifelong East Coaster and animal lover.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/T6qHP3TtmGJ8MCdVA94feF-1920-80.jpg">
                                                            <media:credit><![CDATA[Getty Images / mtcurado]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[The Oracle logo in red letters near the top of a glass office building against a blue sky.]]></media:description>                                                            <media:text><![CDATA[The Oracle logo in red letters near the top of a glass office building against a blue sky.]]></media:text>
                                <media:title type="plain"><![CDATA[The Oracle logo in red letters near the top of a glass office building against a blue sky.]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/T6qHP3TtmGJ8MCdVA94feF-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Oracle has reportedly made a deal with Tencent for a five-year lease at several Oracle data centers in Southeast Asia, covering about 100,000 advanced AI chips unavailable in China. The deal is estimated to be worth about $7 billion, which would come out to about $1.60 per chip per hour over five years, with about 30% upfront, according to a report by the <a href="https://www.ft.com/content/8799b33d-f07c-4a03-82f0-bf5d3d1d29e9"><u><em>Financial Times</em></u></a><em> </em>(<em>FT</em>). The deal lands as compute rental prices continue to climb, per Chief Strategy Officer James Mitchell on Tencent’s August 12 earnings call. Neither company has commented on the report, but such leases are legal under current U.S. rules, per the <em>FT</em>.</p><p>The estimated chip-hour price is around 43% less than the roughly $2.80 per GPU-hour for Nvidia’s H100 with a one-year contract, per research firm <em>SemiAnalysis </em>in August. It’s also about 48% lower than the $3.09 per GPU-hour for a three-year contract with Nvidia’s B200, per a 2025 filing by Japanese data center operator Datasection for an unnamed customer. Market trends have pricing running the other way, with the <em>FT </em>reporting longer terms and larger upfront payments for such leases. Tencent president Martin Lau said on the same call the company could sell its existing orders at “more than 30% profit” compared to what it “paid just a few months ago.”</p><p>No chips are named. The <em>FT  </em>says that such leases could reach even Nvidia’s top processors, but that does not mean this deal involves Nvidia hardware. One-year H100 contracts on <em>SemiAnalysis’s </em>index were above the estimated rate at about $1.70 per GPU-hour even as of October 2025, with the B200 rate being even more expensive despite its longer, three-year term. Going by raw B200 server costs, which exclude buildings, power, networking, and financing, 100,000 B200s would cost $5.44 billion, or about 78% of the reported value, based on Datasection’s 2025 purchase of servers with 5,000 B200s for $272 million.</p><p>Without specifics on the hardware being rented, it is not possible to determine whether this is a bulk deal or if the leasing costs are closer to the actual hardware cost. The pricing on paper, however, better fits Hopper-class at volume. Hopper would include the H100 and H200 and, with <a href="https://www.tomshardware.com/pc-components/gpus/first-nvidia-h200-shipments-reach-bytedance-and-tencent-as-beijing-loosens-its-import-block"><u>H200 imports still limited</u></a>, could meet the criterion of the chips in the deal not being available in China.</p><p>The Oracle data centers are located outside China, with Tencent merely renting time on them. Details including the countries involved, sites, and start date remain undisclosed. Oracle has two Singapore cloud regions, with a Malaysia one announced. The <em>FT </em>says the biggest clients of Southeast Asian data centers are <a href="https://www.tomshardware.com/tech-industry/semiconductors/chinas-top-ai-firms-shift-model-training-overseas-to-access-nvidia-gpus"><u>ByteDance and Alibaba</u></a>, with Chinese firms using the capacity especially for AI training, which domestic chips cannot yet do. Tencent previously inked a $1.2 billion-plus deal for Datasection Blackwell rentals in Japan and Australia, the FT reported in December 2025.</p><p>Some Oracle clients have prepaid for GPUs or brought their own hardware since its March call, and fiscal Q4, from March to May, added $67 billion of AI contracts, mostly of those two kinds, per its June call. June through August, per its quarterly filing, the company “received $11.4 billion of prepayments from customers that included a significant financing component.” Deferred revenue rose from $15.4 billion to $30.8 billion over the quarter. No customers were named, so this can’t be tied to Tencent. The company reported 97.9% GPU utilization during its September 10 earnings call, with renewals at a 20% average premium, which was mostly on GPUs four years or older.</p><p>Tencent prepaying about 30% makes sense given Oracle’s statements earlier this year. It’s a good deal for both parties. Oracle would improve its cash flow following a quarter of negative free cash flow, and Tencent Cloud’s May price raises, per Mitchell, could help recoup the cost of the rented chips.</p><p>The legal picture, given the friction between the United States and China over AI export controls, is thorny. The Remote Access Security Act, a bipartisan bill that <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/u-s-house-passes-bill-to-stop-chinese-companies-from-accessing-export-controlled-american-ai-chips-using-offshore-rental-loophole-remote-access-security-access-act-effectively-extends-export-controls-to-the-cloud"><u>passed the House</u></a> 369–22 in January, would give the U.S. government the right to regulate remote cloud access to sensitive technology, which could cover the chips under the reported deal. The bill specifically targets cloud loopholes and applies to any foreign person rather than naming countries, to prevent adversaries from leveraging cloud networks to bypass U.S. export bans. The bill currently sits in the Senate Banking Committee.</p><p>Separately, Export Administration Regulations license requirements under the Department of Commerce’s Bureau of Industry and Security (BIS) have covered shipments of advanced AI chips, including to China-headquartered companies abroad, since November 2023. Cloud rentals were left out, with the BIS only asking for comment on them. Renting remains one loophole around <a href="https://www.tomshardware.com/tech-industry/us-closes-loophole-that-allowed-chinese-owned-subsidiaries-located-outside-china-to-buy-ai-chips-report-claims-that-hundreds-of-thousands-of-advanced-ai-chips-have-been-acquired-through-bis-blind-spot"><u>BIS guidance on restricted chips</u></a>. <em>Bloomberg </em>reported that the BIS was reviewing offshore access with legal rentals included. The deal is also not barred by Tencent being on the <a href="https://www.tomshardware.com/pc-components/dram/us-dod-adds-cxmt-catl-tencent-to-list-of-companies-suspected-of-aiding-the-chinese-military"><u>Pentagon’s 1260H list</u></a> of “Chinese military companies,” as this is not a sanction. No statement about the deal by any regulatory government official accompanied the report.</p><p>According to <em>The Information</em>, the Department of Commerce is <a href="https://www.tomshardware.com/tech-industry/policy/new-us-export-controls-reportedly-target-chinese-access-to-remote-ai-servers-trump-admins-cut-down-ai-diffusion-rule-could-be-shared-with-industry-as-soon-as-september"><u>drafting a rule</u></a> to block Chinese firms from renting compute in third countries, such as Thailand and Singapore, which would align with the reported details of this deal. U.S. President Donald Trump and Chinese President Xi Jinping met late last month in Washington, covering AI and extending a trade truce for a further two months.</p><p>However, the deal is of a type that might face trouble from both the House-passed bill and Commerce’s reported rule. But the latter cannot be enforced under current law, an attorney at Baker McKenzie told <em>The Information</em>. At the summit, the two leaders agreed to set up a channel to handle AI-related incidents, but nothing was announced on chip controls. While the U.S. has cleared some firms, including Tencent, to buy Nvidia’s H200, China has let in only about 10,000 for Tencent so far, the <em>FT </em>reported in August. The next meeting between the two, at APEC in Shenzhen in November, may clarify the situation.</p><p>The reported deal comes weeks before Oracle’s Investor Day on Oct. 28 in Las Vegas, during its Oct. 25–28 Oracle AI World event at the Venetian. That would be a chance for the company to disclose details about the deal, such as announcing what hardware is to be leased. Given the hardware math, our assumption is that it involves older or non-Nvidia hardware rather than an action born of financial stress at Oracle. The month following Oracle’s event, Tencent is expected to report its Q3 results. These events could also reveal more concrete details about the arrangement.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
                                <item>
                                                            <title><![CDATA[ Amazon and Synopsys ink multi-year billion-dollar deal in multi-year IP agreement to accelerate AI chip design efforts  ]]></title>
                                                                                                <dc:content><![CDATA[ <p>Amazon and chip design tool maker Synopsys are entering a multi-year deal, said to be worth over a billion dollars, in which the two companies will deepen their cooperation on AI chip design, while better optimizing their tools for one another's services. As part of the deal, Amazon will license Synopsys' IP and expand its use of Synopsys' electronic design automation (EDA) software tools for AI chip design and agentic AI technologies.</p><p>The deal goes both ways. From its side of the equation, Synopsys will work with Amazon to optimize its multiphysics solutions for <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/amazon-launches-trainium3-ai-accelerator-competing-directly-against-blackwell-ultra-in-fp8-performance-new-trn3-gen2-ultraserver-takes-vertical-scaling-notes-from-nvidias-playbook">Amazon Trainium</a> and <a href="https://www.tomshardware.com/pc-components/cpus/amazon-unveils-192-core-graviton5-cpu-with-massive-180-mb-l3-cache-in-tow-ambitious-server-silicon-challenges-high-end-amd-epyc-and-intel-xeon-in-the-cloud">Graviton </a>chips, will adopt AWS cloud computing and storage services, and will begin using Amazon Bedrock to build and deploy AI applications and agents for its own development work.</p><p>Although there's clearly some element of cooperative back-scratching with this deal, Synopsys and its competitors are <a href="https://www.tomshardware.com/tech-industry/semiconductors/synopsys-debuts-autopilot-platform-for-developing-chips-autonomously-using-ai-new-agentengineer-platform-is-poised-for-general-availability-by-the-end-of-2026" target="_blank">moving towards automating ever greater portions of the chip design process</a>. Amazon ensuring that process is optimized for Amazon hardware places it in a much more favorable position in a world where chip design is easier and faster. Especially with many of the major AI companies <a href="https://www.tomshardware.com/tech-industry/anthropic-to-build-its-own-co-designed-custom-ai-accelerator-for-inferencing-workloads-samsung-reported-to-be-partnering-with-the-claude-ai-maker-for-manufacturing" target="_blank">looking to develop their own inferencing hardware</a> to ease costs from pricey Nvidia GPUs.</p><h2 id="building-the-future-together">Building the future, together</h2><p>Many of the major AI developments in 2026 have centered around the use of AI agents,  <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/metas-multi-billion-dollar-graviton-deal-exposes-new-bottleneck-in-ai-infrastructure" target="_blank">leading to hardware shortages</a> and a race to fill that gap with optimized hardware. The Synopsys/Amazon deal could put both companies in a position to take advantage of that, designing and developing new AI hardware, while better integration could benefit customers using both companies' services and components.</p><p>The core of the deal, however, is in chip design collaboration. Amazon will expand its use of Synopsys' designs to include blueprints of application-optimized IP: silicon designs that it can incorporate into its own chips. It will also take advantage of Synopsys' AI-powered engineering software to accelerate its development of custom AI chips and AWS infrastructure hardware. </p><p>Considering Amazon already markets its <a href="https://www.tomshardware.com/pc-components/cpus/amazon-unveils-192-core-graviton5-cpu-with-massive-180-mb-l3-cache-in-tow-ambitious-server-silicon-challenges-high-end-amd-epyc-and-intel-xeon-in-the-cloud" target="_blank">Graviton 5</a> for CPU-intensive agentic AI workloads, accelerating the development of next-generation designs could help further cement Amazon's position as it <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/amazon-launches-trainium3-ai-accelerator-competing-directly-against-blackwell-ultra-in-fp8-performance-new-trn3-gen2-ultraserver-takes-vertical-scaling-notes-from-nvidias-playbook" target="_blank">looks to compete with Nvidia on AI data center deployment</a>.</p><p>In the announcement, Amazon also cites its Trainium chips for AI training, and Nitro for cloud security, networking, and storage, suggesting the collaboration between the two companies could augment multiple chip lines. </p><p>Taking a step back from the chip design process, this deal will also see Amazon and Synopsys collaborate on better applying AI within their own workflows, optimizing the silicon-to-system process to make chip design faster and more efficient. Amazon will deploy Synopsys' AI-powered EDA, physics-based simulations, and agentic AI solutions to improve the capabilities of its engineering teams. The companies claim this will help them design, analyze, optimize, and validate new chip designs more efficiently.</p><p>"As our chip designs grow more ambitious and AI reshapes the engineering process itself, Synopsys helps us move faster across the design cycle, helping us deliver more capable, efficient computing for customers worldwide," said Amazon SVP Peter DeSantis in a joint press release.</p><p>With Synopsys and its <a href="https://www.tomshardware.com/tech-industry/semiconductors/the-state-of-agentic-ai-in-chip-design-tools-in-2026-cadence-synopsys-and-siemens-all-pitch-autonomous-engineers" target="_blank">competitors Cadence and Siemens all pushing for faster, more autonomous chip design</a>, we may see a shortening of the typical design cycle for new enterprise hardware. If that proves true, keeping up with that new pace will be paramount for companies like Amazon and Synopsys.</p><h2 id="it-goes-both-ways">It goes both ways</h2><p>Alongside Amazon's expanding use of Synopsys technologies, Synopsys itself will adopt AWS compute and storage services to accelerate its own IP and software development efforts. It will also use Amazon Bedrock to build and deploy AI applications and agents to develop its own product offerings. </p><p>This embeds Amazon cloud services and AI tools further within Synopsys' engineering workflows, while Synopsys' intellectual property becomes more deeply integrated in Amazon's custom silicon. With their joint plan to accelerate Synopsys multiphysics solutions on Amazon's Trainium and Graviton, Amazon's hardware could be more attractive to customers running those engineering workloads. </p><p>Growing that relationship holds further financial incentives for Synopsys, too. The IP agreement introduces a license-plus-royalty model, so as production volume increases, Synopsys royalty revenue could scale with it. With Amazon as its lead customer for the application-optimized IP, this deal could act as a strong endorsement for its chip design blueprints, making it easier to pitch its AI-powered tools, integrated with its IP, to other chip developers.</p><p>For Amazon, this partnership should go beyond accelerating its own custom chip designs. It could strengthen the case for AWS services, with optimization of Synopsys tools and services a useful benefit, as well as both companies benefiting from jointly optimizing the chip design process with agentic AI augmentation.</p><p>Faster chip design doesn't necessarily mean better chips or a shorter time to market, but if the collaboration with Synopsys helps Amazon improve the performance or efficiency of its design, even modest gains could make its AWS infrastructure more attractive and competitive.</p><p>But with no announcements or suggested timeline for new chip development as of yet, both firms will need to demonstrate the effectiveness of this partnership before anyone can measure how accurate that billion-dollar estimation truly is.</p> ]]></dc:content>
                                                                                                                                            <link>https://www.tomshardware.com/tech-industry/amazon-and-synopsys-ink-multi-year-billion-dollar-deal-in-multi-year-ip-agreement-to-accelerate-ai-chip-design-efforts-synopsys-to-adopt-amazon-bedrock-to-deploy-ai-agents-harnessing-aws-compute-and-storage-capabilities</link>
                                                                            <description>
                            <![CDATA[ Amazon and chip design tool maker Synopsys have inked a multi-year partnership worth over a billion dollars. As part of the arrangement, Amazon will license Synopsys' chip designs and its design tools to create and optimize new AI chips. Synopsys will adopt Amazon Bedrock to build and deploy AI agents, and adopt AWS computing and storage services, while optimizing its own tools for Amazon hardware. ]]>
                                                                                                            </description>
                                                                                                                                <guid isPermaLink="false">Psiidz2yYA4qK46cdBPyKX</guid>
                                                                                                <enclosure url="https://cdn.mos.cms.futurecdn.net/yobatAEjDQzCaExZNuCuof-1920-80.jpg" type="image/jpeg" length="0"></enclosure>
                                                                        <pubDate>Fri, 02 Oct 2026 13:50:00 +0000</pubDate>                                                                                                                                <updated>Mon, 05 Oct 2026 12:09:47 +0000</updated>
                                                                                                                                            <category><![CDATA[Tech Industry]]></category>
                                                                                                                    <dc:creator><![CDATA[ Jon Martindale ]]></dc:creator>                                                                                    <dc:source><![CDATA[ https://cdn.mos.cms.futurecdn.net/YeutDv8zJmhi7xH35MSt8Z-320-70.jpg ]]></dc:source>
                                                                <dc:description><![CDATA[ &lt;p&gt;After building his first computers in his teens, Jon Martindale has spent the past two decades covering the latest advances in technology. From displays to PC components, blockchain to AI, and tablets to standing desk accessories, Jon has covered just about every facet of the tech space in his varied career. He has bylines at Forbes, USNews, Lifewire, DigitalTrends, PCWorld, and a range of other sites. He brings that same level of expertise and professional insight to Toms Hardware.Away from writing, Jon is an avid reader, board gamer, and fitness enthusiast. He lives in rural Gloucestershire with his wife, two children, and French Bulldog cross.&lt;/p&gt; ]]></dc:description>
                                                                                                                                <cf:isSponsored>false</cf:isSponsored>
                <cf:hasAffiliateLinks>false</cf:hasAffiliateLinks>
                <cf:isPaid>true</cf:isPaid>
                                                                                                                                                                                    <media:content type="image/jpeg" url="https://cdn.mos.cms.futurecdn.net/yobatAEjDQzCaExZNuCuof-1920-80.jpg">
                                                            <media:credit><![CDATA[Synopsys / Amazon]]></media:credit>
                                                                                                                                                                                                                                    <media:description><![CDATA[Amazon and Synopsys logos together. ]]></media:description>                                                            <media:text><![CDATA[Amazon and Synopsys logos together. ]]></media:text>
                                <media:title type="plain"><![CDATA[Amazon and Synopsys logos together. ]]></media:title>
                                                    </media:content>
                                                    <media:thumbnail url="https://cdn.mos.cms.futurecdn.net/yobatAEjDQzCaExZNuCuof-1920-80.jpg" />
                                                                                                                                                                    <content:encoded >
                            <![CDATA[
                            <article>
                                <p>Amazon and chip design tool maker Synopsys are entering a multi-year deal, said to be worth over a billion dollars, in which the two companies will deepen their cooperation on AI chip design, while better optimizing their tools for one another's services. As part of the deal, Amazon will license Synopsys' IP and expand its use of Synopsys' electronic design automation (EDA) software tools for AI chip design and agentic AI technologies.</p><p>The deal goes both ways. From its side of the equation, Synopsys will work with Amazon to optimize its multiphysics solutions for <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/amazon-launches-trainium3-ai-accelerator-competing-directly-against-blackwell-ultra-in-fp8-performance-new-trn3-gen2-ultraserver-takes-vertical-scaling-notes-from-nvidias-playbook">Amazon Trainium</a> and <a href="https://www.tomshardware.com/pc-components/cpus/amazon-unveils-192-core-graviton5-cpu-with-massive-180-mb-l3-cache-in-tow-ambitious-server-silicon-challenges-high-end-amd-epyc-and-intel-xeon-in-the-cloud">Graviton </a>chips, will adopt AWS cloud computing and storage services, and will begin using Amazon Bedrock to build and deploy AI applications and agents for its own development work.</p><p>Although there's clearly some element of cooperative back-scratching with this deal, Synopsys and its competitors are <a href="https://www.tomshardware.com/tech-industry/semiconductors/synopsys-debuts-autopilot-platform-for-developing-chips-autonomously-using-ai-new-agentengineer-platform-is-poised-for-general-availability-by-the-end-of-2026" target="_blank">moving towards automating ever greater portions of the chip design process</a>. Amazon ensuring that process is optimized for Amazon hardware places it in a much more favorable position in a world where chip design is easier and faster. Especially with many of the major AI companies <a href="https://www.tomshardware.com/tech-industry/anthropic-to-build-its-own-co-designed-custom-ai-accelerator-for-inferencing-workloads-samsung-reported-to-be-partnering-with-the-claude-ai-maker-for-manufacturing" target="_blank">looking to develop their own inferencing hardware</a> to ease costs from pricey Nvidia GPUs.</p><h2 id="building-the-future-together">Building the future, together</h2><p>Many of the major AI developments in 2026 have centered around the use of AI agents,  <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/metas-multi-billion-dollar-graviton-deal-exposes-new-bottleneck-in-ai-infrastructure" target="_blank">leading to hardware shortages</a> and a race to fill that gap with optimized hardware. The Synopsys/Amazon deal could put both companies in a position to take advantage of that, designing and developing new AI hardware, while better integration could benefit customers using both companies' services and components.</p><p>The core of the deal, however, is in chip design collaboration. Amazon will expand its use of Synopsys' designs to include blueprints of application-optimized IP: silicon designs that it can incorporate into its own chips. It will also take advantage of Synopsys' AI-powered engineering software to accelerate its development of custom AI chips and AWS infrastructure hardware. </p><p>Considering Amazon already markets its <a href="https://www.tomshardware.com/pc-components/cpus/amazon-unveils-192-core-graviton5-cpu-with-massive-180-mb-l3-cache-in-tow-ambitious-server-silicon-challenges-high-end-amd-epyc-and-intel-xeon-in-the-cloud" target="_blank">Graviton 5</a> for CPU-intensive agentic AI workloads, accelerating the development of next-generation designs could help further cement Amazon's position as it <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/amazon-launches-trainium3-ai-accelerator-competing-directly-against-blackwell-ultra-in-fp8-performance-new-trn3-gen2-ultraserver-takes-vertical-scaling-notes-from-nvidias-playbook" target="_blank">looks to compete with Nvidia on AI data center deployment</a>.</p><p>In the announcement, Amazon also cites its Trainium chips for AI training, and Nitro for cloud security, networking, and storage, suggesting the collaboration between the two companies could augment multiple chip lines. </p><p>Taking a step back from the chip design process, this deal will also see Amazon and Synopsys collaborate on better applying AI within their own workflows, optimizing the silicon-to-system process to make chip design faster and more efficient. Amazon will deploy Synopsys' AI-powered EDA, physics-based simulations, and agentic AI solutions to improve the capabilities of its engineering teams. The companies claim this will help them design, analyze, optimize, and validate new chip designs more efficiently.</p><p>"As our chip designs grow more ambitious and AI reshapes the engineering process itself, Synopsys helps us move faster across the design cycle, helping us deliver more capable, efficient computing for customers worldwide," said Amazon SVP Peter DeSantis in a joint press release.</p><p>With Synopsys and its <a href="https://www.tomshardware.com/tech-industry/semiconductors/the-state-of-agentic-ai-in-chip-design-tools-in-2026-cadence-synopsys-and-siemens-all-pitch-autonomous-engineers" target="_blank">competitors Cadence and Siemens all pushing for faster, more autonomous chip design</a>, we may see a shortening of the typical design cycle for new enterprise hardware. If that proves true, keeping up with that new pace will be paramount for companies like Amazon and Synopsys.</p><h2 id="it-goes-both-ways">It goes both ways</h2><p>Alongside Amazon's expanding use of Synopsys technologies, Synopsys itself will adopt AWS compute and storage services to accelerate its own IP and software development efforts. It will also use Amazon Bedrock to build and deploy AI applications and agents to develop its own product offerings. </p><p>This embeds Amazon cloud services and AI tools further within Synopsys' engineering workflows, while Synopsys' intellectual property becomes more deeply integrated in Amazon's custom silicon. With their joint plan to accelerate Synopsys multiphysics solutions on Amazon's Trainium and Graviton, Amazon's hardware could be more attractive to customers running those engineering workloads. </p><p>Growing that relationship holds further financial incentives for Synopsys, too. The IP agreement introduces a license-plus-royalty model, so as production volume increases, Synopsys royalty revenue could scale with it. With Amazon as its lead customer for the application-optimized IP, this deal could act as a strong endorsement for its chip design blueprints, making it easier to pitch its AI-powered tools, integrated with its IP, to other chip developers.</p><p>For Amazon, this partnership should go beyond accelerating its own custom chip designs. It could strengthen the case for AWS services, with optimization of Synopsys tools and services a useful benefit, as well as both companies benefiting from jointly optimizing the chip design process with agentic AI augmentation.</p><p>Faster chip design doesn't necessarily mean better chips or a shorter time to market, but if the collaboration with Synopsys helps Amazon improve the performance or efficiency of its design, even modest gains could make its AWS infrastructure more attractive and competitive.</p><p>But with no announcements or suggested timeline for new chip development as of yet, both firms will need to demonstrate the effectiveness of this partnership before anyone can measure how accurate that billion-dollar estimation truly is.</p>
                                                            </article>
                            ]]>
                        </content:encoded>
                                                </item>
            </channel>
</rss>