Servers running x86 processors from AMD and Intel used to rule the market, both unit and money-wise, less than a decade ago, but fast forward to today, Arm-based machines command well over 45% of the server market, according to data released by IDC. While technically x86 machines still control 52% of the market in terms of revenue, the real winner is a different category altogether: GPU- and ASIC/FPGA-accelerated systems, which generated over 70% of the global server revenue in the first quarter of 2026.

Server market reaches $122.6 billion in a single quarter, Dell leads the game

IDC estimates that the global server market generated a record $122.6 billion in revenue in the first quarter of 2026, up 30.4% year-over-year, as spending on AI infrastructure remained particularly strong.

Sales of ODM Direct servers — custom machines ordered by hyperscalers that run merchant or custom silicon — accounted for 50.2% of the revenue (down from 64.1% in Q1 2025) and reached $61.53 billion, up modest 2.1% year-over-year*. By contrast, sales of standard servers from well-known brands grew at a much higher pace, which suggests that branded vendors such as Dell, HPE, Supermicro, and others won a larger portion of AI infrastructure deployments than they did a year earlier. That was probably made possible by accelerating enterprise AI deployment and sovereign AI projects, which tend to buy machines from branded vendors, as well as hyperscalers increasingly turning to well-known suppliers for AI hardware.

When it comes to vendor rankings, Dell remained the largest server supplier by revenue with a 16.5% share of the market after its revenue surged 244.1% year-over-year to $20.3 billion, which was driven by exceptionally strong AI server demand. Supermicro remained in second place with $9.3 billion in revenue and a growth of 128.9%.

Lenovo ranked third with $5.6 billion and 36.5% growth, while IEIT Systems (which is a part of the sanctioned Inspur Group) dropped to fourth after revenue declined 7.0% to $4.0 billion. HPE was No.5 with $3.7 billion in revenue, up 17.2%. Other vendors — from Asus to Atos and from ASRock Rack to Gigabyte — commanded 14.8% of the market with $18.11 billion in revenue, up from 13% and $12.21 billion in the same quarter a year ago.

Arm-based machines rapidly gain revenue share

As AI servers dominated the market in Q1 2026, systems with various types of accelerators accounted for over 70% of the revenue. However, the rise of Arm-powered machines is the elephant in the room that is hard to miss, as it represents a tectonic shift in the whole market, both to the Arm instruction set architecture (ISA) in general and custom-built Arm CPUs designed by hyperscalers.

Non-x86 platforms generated $58.7 billion in revenue, a 107.6% increase year-over-year, which lifted their share of the market to 47.9%. Most of the non-x86 systems are Arm-based AI machines (think Nvidia's NVL72) as well as systems running custom CPUs, AWS, Google, and Microsoft, just to name a few. Still, also keep in mind IBM Z mainframes and IBM Power Systems (including storage) that use CPUs featuring proprietary non-x86 and non-Arm ISAs and which still generate $1 billion or more in revenue. IDC claims that Arm-based machines accounted for more than 95% of non-x86 revenue, so it is safe to say that Arm-based machines commanded over 45% of server revenues in Q1 2026.

One of the reasons why Arm-based machines now command a huge chunk of the server market is because they are used inside such systems as Nvidia's NVL72 'Blackwell' that sell for up to $6.5 million per unit. Each NVL72 rack-scale solution carries 36 compute trays with two Blackwell GPUs and one Grace CPU per unit, so while unit-wise each we are only talking about 36 processors, dollar-wise one NVL72 machine is as expensive as 928 entry-level 1P server (for $7,000) for cloud or edge applications or 433 higher-end 2P servers (for $15,000) for cloud or virtualization applications.

Given the fact that Nvidia will continue bundling its own Arm-based Vera CPUs with NVL72 'Vera Rubin' machines that will be more expensive than their Blackwell ancestors, we will not be surprised that Arm-based machines will account for well over 50% of the server market revenue in the second half of this year or in 2027. Also, keep in mind that Nvidia plans to sell server racks featuring only Vera CPUs for agentic AI applications, which will further drive sales of Arm-based machines.

Accelerated servers: The real winner

Since AI servers dominate server sales, it is not surprising that sales of accelerated servers are increasing. Systems equipped with GPUs produced $68.9 billion in revenue during the quarter (up 24.8% compared to the same period a year earlier) and accounted for 56.2% of all server sales. Servers based on other accelerator types, including custom ASICs and FPGAs, expanded to $17.7 billion, up 122.1% YoY. As a result, accelerated servers earned $86.6 billion in Q1 2026, which is around 70.6% of all server revenue.

X86 servers remain unit volume champions, but suffer from shortages

In contrast, x86 server revenue declined 2.9% to $63.9 billion, though IDC attributes this weakness to supply limitations rather than deteriorating demand. The market research firm claims that the industry's primary constraint is no longer customer appetite for general-purpose servers, but rather the availability of key components, including CPUs, DRAM, NAND memory, and hard drives.

Without any doubt, x86 servers remain working horses for the industry. In fact, many of them use accelerators, including ASICs, FPGAs, and GPUs, as they are used for a wide range of workloads, including AI, supercomputing, simulations, encryption, video transcoding, and many more.

AMD and Intel shipped nearly 20 million EPYC and Xeon SP processors for data center systems in 2025, according to Dean McCarron, the head and principal analyst at Mercury Research. He believes Nvidia is on track to ship four million Grace and Vera CPUs this year, which is considerably lower compared to shipments of AMD and Intel. It is hard to estimate how many custom Arm-based CPUs are deployed by AWS, Alibaba, Google, and Microsoft, but it is safe to say that we are talking millions of CPUs here; otherwise, the companies would not be able to justify development and production of custom silicon.

From a volume perspective, x86 servers remain the most popular machines, and it will probably take some time before ARM can challenge x86 in mainstream general-purpose servers. Nonetheless, it is safe to say that Arm-based data center CPUs are catching up with x86 parts in terms of volumes.

Summary

The global server market hit a record $122.6 billion in the first quarter of 2026 as AI infrastructure spending continued. Accelerated systems powered by GPUs, custom ASICs, and FPGAs generated more than 70% of server revenue, while Arm-based platforms — including Nvidia's Grace Blackwell as well as custom CPUs from Arm, Google, and Microsoft — captured nearly half of the market.

Although x86 servers based on AMD EPYC and Intel Xeon processors remain dominant in shipment volumes, supply shortages of CPUs, memory, and storage components constrained revenue growth, which further enabled Arm-powered AI-optimized systems to gain share. But while at 20 million data center processors per year, x86 volumes are untouchable for Arm today, things may change in the coming years. Nvidia is on track to ship 4 million CPUs in 2026, and other developers of custom Arm-based CPUs are certainly not standing still.

*There is one significant difference with IDC's 'ODM Direct' classification. IDC classifies revenue according to which company invoices the customer, not necessarily who manufactures the hardware. As a result, while many AI servers are built by ODMs like Compal, Foxconn, or Quanta, they are sold under brands like Dell or HPE. As a result, while the latter get more business from enterprises or sovereign AI deployments, this does not mean that big ODMs are losing business; they are actually gaining it, as the appetites of hyperscalers like AWS, Google, Meta, or Microsoft are not going anywhere, just demand from new entrants emerges.

Ever since buying out VMware in 2023, Broadcom has raised prices for existing customers. Although the play did result in massively increased revenue for VMware, it also triggered a slow but steady exodus away from the platform, as enterprises everywhere balk at the price tags and/or sue Broadcom.

UK supermarket chain Tesco is the latest entity to up and replace VMWare across its fleet of 40,000 servers, a move it expects to complete in 2027, Ars Technica reports.

This marks the largest publicly known migration off the platform in recent years, and arrives against the backdrop of the $134-million (£100m) lawsuit that Tesco filed last year against Broadcom. In the filing, Tesco claims that Broadcom committed a breach of contract by not honoring the perpetual licenses it bought in 2021, and that Broadcom's actions are at odds with multiple competition laws. Tesco also named VMware itself and Computacenter, the reseller, in the suit.

Tesco's story is a familiar one since the Broadcom acquisition. Much like other customers, the supermarket chain had reportedly bought perpetual licenses for vSphere Foundation and Cloud Foundation, plus support services until 2026 with a four-year extension option. In 2023, Broadcom bought VMware and stopped perpetual licensing in lieu of subscriptions.

This resulted in highly elevated pricing figures for Tesco, apparently to the tune of a 175% price hike for VMware and a 350% upcharge for mainframe software. After the 2025 lawsuit, the situation predictably deteriorated further, and Broadcom stopped supporting Tesco's VMware suite, forcing the chain to find an unnamed third party for that job. CEO Hock Tan's shop goes as far as to refuse security updates to customers without subscriptions, too.

There's no word on which software Tesco's systems administrators have selected for the job, though the Ars Technica report says that the choice appears to be incompatible with both Veeam and Zerto, suggesting it's a lesser-known offering. Additionally, TechRadar notes that HP Enterprise just announced it's offering its customers a year's free licensing to Morpheus VM Essentials, along with a $1 license for Zerto migration software. The timing is interesting, as HP is looking to snag former VMWare customers.

Ever since the Broadcom acquisition, VMware has forced subscription bundling for support contracts, effectively doubled per-core pricing, specified a 72-core minimum for purchase (thereby nuking small business setups), added a three-year contract minimum, and enforced penalties for late renewals. Moving an entire server infrastructure off a hypervisor platform is significant. Even still, a few corporations have done so, and others are reportedly considering switching away.

Data center water use is one of the hot topics right now in the U.S. and is one of the primary reasons why such projects are being blocked across the nation. However, Amazon claimed that it’s improving its water efficiency and that it uses the least amount of water on a per kWh basis among the AI tech giants, including Microsoft, Google, and Meta. Aside from that, it also said that despite using 2.5 billion gallons of water across the world in its data centers, it said that that amount is still a fraction of what other industries in the U.S. are using annually.

Amazon pointed at EPA data from 2017 that said Americans use 9 billion gallons of water daily just for landscape irrigation. This amounts to nearly 3.3 trillion gallons of water every year just for watering plants and gardens and doesn’t even include the irrigation needed for food production. This makes its 2.5-billion-gallon water consumption a literal drop in the bucket, amounting to just 0.075% of water used for keeping gardens green and thriving.

These gardens, however, have a direct benefit in that they could potentially keep ambient temperatures lower (versus an all-cement environment) and add to the aesthetics of the immediate area — the only question is how many people these developments benefit. Moreover, many new data centers in the U.S. are reportedly being built in areas suffering from water issues. While the overall data center water consumption might seem small if you look at the larger picture, we still cannot turn a blind eye to its potential effects on the local community.

For example, a Meta data center has allegedly caused a Georgia town’s deep-well water source to muddy — a sign that the level of the local water table is hitting low levels, allowing mud to be siphoned by the residents’ water pumps. Another site reportedly sucked 29 million gallons of water in 15 months, which caused low water pressures for residents. Reports like these make it harder for other data center projects to secure permits from local governments to start construction, especially as communities fear that these issues could happen in their area, too.

Amazon boasts water efficiency improvements

As the pushback against data centers’ egregious resource consumption has become front-and-center in many permitting fights, Amazon is showing that it’s taking steps to reduce its water consumption and claims that it’s on track to being “water positive” by 2030. It achieved this by implementing innovations in its data centers to reduce water use without reducing performance.

The company says that it mostly uses air cooling for its data centers, which uses up a lot less electricity compared to water cooling. But when ambient temperatures rise, it switches to Direct Evaporative Cooling, with the company spraying water on an absorbent medium, which the company describes as “a sophisticated, giant sponge,” and then runs the hot air through it to reduce temperatures by five to 10 degrees. It says that this is more power efficient than using chillers, reducing power use by about 20% to 25% during the hottest time of the day when power usage by other users is at its highest, too.

It also said that it raised the temperature thresholds in its data centers. While this meant that its servers ran hotter compared to previous years, it didn’t impact on the longevity of their hardware and they were still able to deliver the same amount of computing power as before. Through trial and error, the company discovered that it could keep running its servers using air cooling only, with its water-cooling systems only kicking in once the ambient temperature reaches 85 degrees F (or more than 29 degrees C). This resulted in 50% reduction in water use, says Amazon water specialist Beau Schilz.

Efforts like this are essential for data center operations to become sustainable. After all, despite the pushback against new data center projects, we cannot forever put a moratorium on them as AI’s demand for compute is increasing. But until AI hyperscalers can earn back the trust of the people and prove that they can build infrastructure without increasing costs for everyone else and hoarding the resources that communities need for their survival, then they can only expect opposition to continue.

Researchers from the University of California San Diego (UCSD) collaborated with Google to recycle “old” Pixel smartphones and give them a second life as a low-cost data center. According to Google Research, retired smartphones are part of the “embodied carbon” that is associated with manufacturing and its carbon footprint. In fact, humanity’s penchant for mobile devices and replacing them every few years is one of the biggest contributors to e-waste, so the group from UCSD planned to give these discarded devices a second life as a “general-purpose computing platform.”

The study revealed that smartphones from just three years ago still deliver a higher single-core performance compared to servers like the Asus RS720A-E11, which can be equipped with Nvidia H200 or Nvidia RTX Pro 6000 GPUs and two AMD EPYC server processors, that you frequently find in the most powerful data centers. While the latter delivers performance that a mobile device can’t even dream of, the fact that the former still scored higher in the SPEC benchmarking suite on a per-core basis meant that researchers could still use them for compute tasks with a little creativity.

The first thing they did was to strip these gadgets of non-essential components — displays, batteries, cameras, speakers, chassis, etc. Only the motherboard remains, as it plays host to the SoC needed for running compute. The Android operating system is then replaced with a general-purpose Linux distro used in data center applications, which removes unnecessary bloat found in the original consumer device and allows for the deployment of orchestration software like Kubernetes. Benchmarking results revealed that 25 to 50 old phones wereequal to the computing power of a single dual-socket server-class CPU.

UCSD determined that a 20-phone cluster can support one application that a 75+ student class requires. So, instead of hosting it on the cloud, which would entail additional costs and resource use on the data center side, it could instead run these apps on a local deployment of these used smartphones. The research team plans to use 2,000 phones to build a local data center that can support “a hundred such classes at once.” Aside from getting the advantage of running apps locally and owning the hardware needed for them, the group also says that it’s only a “fraction of the usual cost,” likely referring to building a local server made from new components. This is especially true today, with the increased pricing for memory and storage chips.

The research team says that it expects to launch the full system later this year and is looking to see how consumer parts can withstand continuous use in a data center application. But even if the experiment is successful, we don’t foresee AI hyperscalers switching to servers made from used phone parts as they would often want to work with fewer parts and the reliability delivered by specialized hardware. Still, this is a great option for universities and educational institutions, as well as smaller entities that do not have the resources to secure brand-new parts and compete against tech giants with billions of dollars to burn.

This isn’t the first time scientists have looked at giving old phones a second life — another group of researchers looked at converting old phones into “tiny data centers” last year, even using one set of four old devices for underwater monitoring. After all, even though the SoCs found in these devices are considered “outdated” by modern standards, they should still be more than capable enough for many mundane tasks. NASA even repurposed the Qualcomm 801 SoC, a mid-range chip from 2014 and found in the Ingenuity Mars helicopter, to help the Perseverance rover find its way around the Red Planet like some sort of processor for a makeshift GPS. And for smartphones that no longer work, people are finding ways to extract the gold and other resources found on their boards for recycling.

What's a common characteristic of clouded leopards, Komodo dragons, and Sumatran tigers? Their habitats are all close to data centers — at least, if the proposal by DC Blox's to build a facility right next to the Nashville Zoo comes to pass. The 1.6-acre facility (69,000 sqft.) buildout is in the permit stage, but already raising more than a few issues both with the zoo itself and concerned Nashville residents, as evidenced by an online petition that counts over 322,000 signatories as of this writing.

The Nashville Zoo published a post in its blog detailing its major concerns. Besides the obvious concern about the power draw and its impact on the local grid, the zoo claims the facility could add to noise and light pollution, as well as lower the water quality in the surrounding areas. Altogether, those factors might have a negative impact on a number of hosted species and breeding programs, including the zoo's famous clouded leopards, animals "notoriously sensitive to any mechanical noise."

Nashville Zoo CEO Rick Schwartz reportedly said the entity was looking to use part of the 23.5-acre lot in question for an education and conservation center and that it was in talk for years with current owners MarketStreet Enterprises to buy part of the property for that purpose. MarketStreet's website lists the Nashville Zoo as one of its current tenants.

Technical details appear to be relatively sparse, but the datacenter would draw an estimated 50 MW of power from the local grid and reportedly include its own power substation. The 1.6-acre figure would hold the build's first phase, a relatively small parcel out of a 23.5-acre lot. DC Blox claims the project would use "closed-loop or waterless" cooling and comply with all local ordnances.

While 50 MW is definitely a lot of power, it's not anywhere in the same ballpark as that required by the vast majority of AI datacenters. This could mean the Nashville datacenter might be a "standard," relatively inoffensive co-location facility, keeping in theme with DC Blox's other operations. The proposed area is also quite compact in comparison to thousand-acre constructions demanding tens of gigawatts. Nashville already hosts another couple dozen datacenters, though none with a plan to pull up to 50 MW from the grid.

As an example, DC Blox's Birmingham facility does in fact use air cooling, and could draw up to 60 MW of power when fully built out. Having said that, the firm's website claims its operations are "AI-ready," meaning that the power and cooling requirements could rise in time, though theoretically such expansions would require additional permits. For its part, DC Blox says the datacenter would "not be an AI factory placing a burden on local resources," and further noted that the area had an existing permit for a datacenter, though there appear to be no records that such a facility was ever actually built.

Notably, while there appear to be no city-specific laws regarding datacenters, Tennessee law at least requires that datacenters pay for any grid infrastructure upgrades. Nashville's leadership is nevertheless considering adding restrictions to new datacenter investments, including but not limited to water, power, and noise regulations. Those would even ban buildouts larger than 500,000 sqft altogether.

The project is only in the permit stage, and in an odd spot at that. Local law generally requires that only land owners can apply or support a building permit, and yet DC Blox filed one while not yet owning the area, which is currently being sold by MarketStreet Enterprises. There's no confirmation on whether MarketStreet has co-signed or authorized the filing. Nashville Mayor Freddie O'Connell reportedly called the application "unusual" and is asking the Metro's legal department to look further into the situation.

Astera Labs demonstrated its recently introduced Scorpio X-Series 320 Lane Smart Fabric Switch, which appears to be the industry’s largest open memory-semantic fabric switch, at Computex 2026 in Taipei. The PCIe 6.0 switch with 320 lanes can be used to build large multi-GPU scale-up clusters, large shared KV-cache memory pools, and disaggregate data center infrastructure using custom topologies.

The switch provides 320 PCIe 6.0 lanes and 20 Tbps of switching bandwidth, up from 144 lanes and 9 Tbps for previous-generation devices. Astera Labs says the increased number of lanes enables larger scale-up domains, enabling the connection of up to 80 accelerators using a single switch. By contrast, older 144-lane switches support up to 32 accelerators per switch. For clusters with more than 64 accelerators, the company says the new device reduces switch hops from as many as three to one and cuts switch count by a factor of four to six while still providing all-to-all connectivity akin to that provided by Nvidia’s NVL72 systems (albeit with lower bandwidth and higher latencies). The switch can support both standard and custom accelerators as long as they use standard PCIe connectivity.

On the trade show floor, Astera Labs is showing off its switching capabilities with Intel’s Arc B70 Pro graphics cards. However, real-world deployments based on the Scorpio X-Series 320-lane PCIe switches will likely use more advanced Intel hardware. In general, the switch can be used to build clusters from all types of accelerators that do not support their own NVLink or UALink-like interconnections, including AMD’s Instinct MI350P and Nvidia’s RTX 6000 Blackwell. Astera has yet to showcase a full working cluster featuring 80 accelerators, as the company only got the Scorpio X-Series 320-lane PCIe switch from the fab eight weeks ago. Also, finding 80 similar accelerators is not easy. Nonetheless, based on the company’s demonstration, the switch appears to be working.

A key feature of the Scorpio X-Series is Hypercast, which is a hardware-based data replication engine intended to accelerate communication-intensive operations common in AI models. According to Astera Labs, MoE networks tend to route tokens across hundreds of experts and create large amounts of multicast traffic between accelerators. In such cases, traditional switching architectures either require repeated data transmissions or slow multicast-group reconfiguration, whereas Hypercast is designed to handle these communication patterns directly in hardware, reduce GPU networking overhead, and improve accelerator efficiency, Astera Labs claims.

The company also added In-Network Compute engines that offload collective operations such as AllReduce, ReduceScatter, AllGather, AllScatter, and all-to-all exchanges. These features can reduce communication latency by more than 50% in certain workloads, according to Astera.

Another important feature of the Scorpio X-Series 320-lane PCIe switch is its memory-semantic connectivity, which enables connected processors to access fabric-attached resources using native load and store operations rather than software-controlled transactions. This greatly simplifies usage of the device and improves real-world performance by reducing overhead and improving fabric efficiency at scale.

Astera says that the production ramp of the Scorpio X-Series 320-lane PCIe switch is set for the second half of 2026. Currently, the company is sampling the switch with leading hyperscalers.

Supermicro is showing off its upcoming machines based on AMD’s 6^th Generation EPYC ‘Venice’ processors, Instinct MI450 accelerators, and, of course, Nvidia’s Vera Rubin-based solutions. Arguably, the most important product that the company demonstrated at its CEO’s keynote is the VR200 NVL72 rack that uses Vera CPUs and Rubin GPUs. The machine, just like other upcoming liquid-cooled systems from Supermicro will use the company’s all-new coolant that has a 1000x higher electrical impedance compared to today’s mixtures, which may be quite important for next-generation AI machines.

“Our new coolant [uses a new] formula [and] reaches up to 1,000 times higher electrical impedance than a standard cooler,” said Charles Liang, chief executive of Supermicro, during his keynote speech at Computex. “In case there are small leaks, when you have a high electrical impedance, the system will not [shut down] [and will] keep running.”

Conventional water-based coolants used in direct liquid cooling systems have electrical conductivity (albeit lower than water), so if coolant leaks onto a motherboard, GPU, power delivery circuitry, or connectors, it can create leakage currents or even short circuits. A coolant with 1000 times higher electrical impedance is far more resistant to current flow, which reduces the likelihood that a minor leak will immediately shut down a system or damage electronic components. This is important as modern rack-scale AI solutions like Nvidia’s VR200 NVL72 are rumored to cost around $8 million, so their protection is crucial. Also, reducing downtime in AI data centers is important as these machines must make money for their owners.

Unfortunately, Supermicro’s claim is difficult to evaluate because the company does not disclose specifications of the coolant, such as conductivity (µS/cm), resistivity (MΩ·cm), or dielectric strength (kV/mm). It also remains tight-lipped about the baseline coolant used for comparison. Since modern water-glycol coolants are already fairly resistant, a 1000-fold improvement sounds significant, but without the actual details, it is impossible to determine the practical magnitude of the advancement. What Supermicro claims is that a minor coolant leak would be less likely to force an immediate server shutdown, which means lower downtime risk in large AI deployments. However, we have no idea whether the new coolant behaves more like a dielectric fluid and less like water-glycol coolants.

The new coolant will be used in all new liquid-cooled systems from Supermicro, including the upcoming AMD Helios and Nvidia VR200 NVL72 machines. Given that Nvidia nowadays does not leave many ways for its partners to differentiate, the coolant with 1000 times higher electrical impedance than conventional coolants will likely be among the key selling points of Supermicro’s Nvidia-based offerings.

Speaking of Nvidia’s AI systems, Supermicro is on track to release its Vera Rubin-based systems in the second half of this year, just like other makers. Supermicro intends to offer both NVL72-rack-scale machines based on Nvidia’s Vera CPUs and Rubin GPUs, as well as Rubin DGX systems with different processors.

As for AMD’s side of matters, Supermicro is also on track to release MI455X-based Helios rack-scale solutions in the second half of the year, though the company does not disclose whether these use UALink or UALink over Ethernet interconnections. In addition, Supermicro is also on track with 1-way and 2-way servers based on AMD’s 6^th Generation EPYC ‘Venice’ processors that are made using TSMC’s N2 process technology. These machines are also expected to hit the market in 2026.

The very first set of Nvidia Vera CPU benchmarks have just been released by Phoronix, with results from a set of common Linux benchmarks. While the tests were curated by Nvidia at its Santa Clara headquarters, the early data from those tests indicates that Vera is highly competitive compared to AMD's EPYC and Intel Xeon offerings, at least in the workloads Nvidia is targeting with the chip.

Phoronix was invited to NVIDIA's Santa Clara headquarters to test the upcoming 88-core CPUs. Vera is notable for all kinds of reasons, but most especially because it doesn't license an Arm processor core. Instead, like Apple's chips, it uses the ARM instruction set on a fully custom CPU core known as "Olympus." This isn't the first time Nvidia has produced a custom CPU core; that would be "Denver" in the Tegra K1 from 12 years ago. However, where Denver was a desktop-class CPU constrained by a mobile power budget, Vera is a server-class monster fed with a server-class power budget, and the proof is in the benchmark results.

Phoronix's Michael Larabel was able to test Vera across a range of benchmarks, including code compilation tests, synthetic memory benchmarks, AV1 video encoding, Python, Java OpenJDK, file compression, Lua JIT, and some database benchmarks. In most of the tests, the Vera processors are highly competitive against the competition, which includes AMD EPYC "Turin" processors in single- and dual-CPU configurations as well as an Intel "Granite Rapids" Xeon chip. As Larabel says, Vera offers "competitiveness to Intel/AMD x86_64 CPUs that [he has] never seen out of any other ARM or non-x86_64 processors." It doesn't win outright in most tests, but it's running very close to the EPYC configurations, which is shockingly good for a first-generation custom server core from Nvidia.

What's even more impressive is if you switch things over to a per-thread view of the data, which Phoronix does for some tests. Historically, Arm server vendors have achieved similar performance versus AMD and Intel in multi-threaded workloads by simply cramming a huge pile of cores onto a chip, but the single-threaded performance of x86 was beyond reach. Well, in a timed Gem5 compilation, only the AMD EPYC 9575F was able to beat Nvidia's Vera on a "performance per core" metric. In a Linux kernel build, Vera tops the stack of server chips by a decent little margin. It's real eyebrow-raising stuff, and speaks to the strength of the Vera core.

Looking at the geometric mean of test results, Vera comes out on top by a fair margin, in part thanks to very strong performance in LuaJIT FFTs, ClickHouse database server, and the Renaissance JVM benchmark, where it absolutely dusted the competition. This is genuinely exceptional, but there are caveats to this data. Larabel notes that Nvidia limited the scope of these initial benchmarks to its "intended markets and target use-cases," but it's also true that these tests are benchmarks that Larabel is familiar with and already runs on server CPUs for performance analysis. He says that today's results are a "small subset" of his typical preferred approach, which is to say that this data is valid, but may not represent all workloads or use-cases.

One key detail that the site wasn't able to test comes down to power efficiency. This is a big deal for consumer systems, but it's a much bigger deal for AI data center operators who are currently struggling with build-outs because the power infrastructure worldwide simply isn't robust enough to support megawatt AI training clusters popping up all over the place. Nvidia says that Vera has a 450W TDP, while the fast SOCAMM2 memory it uses consumes another 50 watts; the Xeon and EPYC chips used for comparisons are rated for 500W before accounting for platform memory power. Real power consumption can be very removed from TDP figures, though, so we'll have to wait and see how that plays out.

Another point in favor of Vera is software support. According to Larabel, Vera has "great upstream open-source support," which is encouraging for the future of these chips. Over the decades, we've seen many examples of promising hardware that falls by the wayside due to insufficient software support. All of the testing was performed on a mainline Linux kernel with no need for nasty Device Trees or hacky bespoke driver solutions.

While Nvidia's Vera looks legitimately formidable next to current-generation chips, both AMD and Intel have new chips on the way that look downright gobsmacking. AMD's planning to pack some 256 Zen 6 CPU cores into a single socket with the dense version of EPYC "Venice", while Intel's "Clearwater Forest" will combine up to 288 Darkmont cores with its bleeding-edge 18A process technology in a direct attempt to reclaim the performance-per-watt crown. The end of the year and next year's server landscape are going to be very, very different from today.

On Wednesday, the Taiwan Keelung District Prosecutors' Office executed search warrants across 12 locations and is now seeking to detain three individuals in the island’s first formal crackdown on illicit AI semiconductor exports to China. According to a fresh report from Bloomberg, the trio is accused of forging shipping documents to sneak AI servers manufactured by US-based Super Micro Computer Inc. (Supermicro) into China, Hong Kong, and Macau, in direct violation of Washington’s trade restrictions. While the scale of this specific bust is relatively small, purportedly involving around 50 servers, the political and economic implications are far more significant.

If the name Supermicro sounds familiar in the context of chip smuggling, that’s because the company is currently sitting at the center of the largest tech-evasion case in U.S. history. Just months ago, the U.S. Department of Justice indicted Supermicro co-founder and Senior VP Wally Liaw for orchestrating a breathtaking $2.5 billion smuggling ring that reportedly used front companies in Thailand to route restricted NVIDIA hardware to Chinese tech giants like Alibaba.

Though Taiwanese prosecutors claim this new 50-server case was initiated independently of the U.S. investigation, it targets the exact same vulnerability. Smugglers have long used Taiwan as a transit hub, banking on the hope that local compliance teams look the other way. The fact that Taipei is now using local forgery and fraud laws to lock people up signals a major policy shift under President Lai Ching-te, who is under immense pressure from Washington to secure the global AI supply chain.

For those tracking this story closely, Taiwan's sudden aggression isn't really a surprise. Reports of underground hardware in China include thriving repair shops working on illicit Nvidia GPUs and firms stripping restricted NVIDIA silicon off dead boards to build custom "Franken-cards," show chip smuggling is a significant issue.

Yesterday's raids in Taiwan suggest the government is finally taking a harsher stance on the issue. The timing of the Taiwan crackdown is pretty ironic, though; just this week, NVIDIA dropped its Q1 Fiscal 2027 financial results, and the numbers show that the company essentially has a license to print money, regardless of what happens on the black market.

Nvidia pulled in a staggering $81.6 billion in total revenue, an 85% increase year-over-year, with a mind-melting $75.2 billion of that coming strictly from its Data Center division. Driven by insatiable demand for its Blackwell architecture GPUs, Nvidia's growth has gone entirely parabolic. The most critical piece of data for the smuggling narrative was buried in Nvidia's forward guidance, though; management explicitly stated that the company is assuming zero data center compute revenue from China moving forward.

The analysis here is simple: Nvidia has successfully decoupled its legal financial future from China. Nvidia doesn't need the market, and some would argue it doesn't need the regulatory headache. For the black market, the walls are closing in fast. When chip smuggling was just a U.S. DOJ priority, enforcement was limited by geography. Now that key manufacturing and transit hubs like Taiwan and Singapore are actively hunting down middlemen using local criminal fraud laws, the supply chain is fracturing. Getting banned Hopper or Blackwell chips into mainland data centers just became exponentially more fraught.

Amazon’s data centers in Bahrain and the UAE have been hit multiple times by drone and missile strikes from Iran since the U.S. started bombing the country in February 2026. This left the company’s ME-CENTRAL-1 and ME-SOUTH-1 disrupted, with the AWS Health Dashboard indicating that it will take months before they can go back online.

“The Middle East (UAE) Region (ME-CENTRAL-1) has suffered damage as a result of the conflict in the Middle East and is currently unable to reliably support customer applications,” Amazon said in its latest update. “While some workloads continue to function normally, we strongly recommend customers migrate all accessible resources to other Regions and restore inaccessible resources from remote backups as soon as possible. Relevant billing operations are currently suspended while we restore normal operations in this AWS Region. This process is expected to take several months.”

These sites were first hit in early March, a few days after the U.S. started its bombing campaign on Iran. A few weeks after that, AWS Bahrain was hit once more, this time by drone strikes, further disrupting operations. The last reported strike happened in early April, putting several zones into a “hard down” status. Although both sites remain partially operational, it seems that Amazon has refrained from billing customers who used these servers and continues to encourage them to move to other regions and avoid further disruption due to the conflict.

Even if the missile and drone strikes haven’t completely leveled the two data centers, Amazon is still on the hook for months of repairs. Aside from these direct hits, the shockwave and post-explosion fires they caused have compounded the damage to the sensitive equipment used in these sites. The company also reported that fire suppression systems have caused flooding and water damage to various equipment, while cooling systems have suffered from mechanical failures.

There’s currently an uneasy truce in the area, as the two parties are at the negotiating table trying to end the war. However, hostilities could resume at any time, especially as the U.S. and Iran cannot agree on several points, making it risky for Amazon to start repairs right now. But even then, the ongoing chip shortages could make it harder for the company to get the components it needs to get these services up and running again, unless they have backup servers in their inventory ready to be deployed if peace in the region has been achieved.

The residents of a small town in Pennsylvania are currently amid a fight for their community, as data center developers have applied for permits to build six campuses within the area. The proposed data centers will be built in the town of Archbald, which has a population of just 7,000 people, and will include 51 data warehouses. The Washington Post reports that each of these data warehouses will be about the size of a Walmart Supercenter, all of which would encompass about 14% of the town’s 17-square-mile area. The Archbald town council approved zoning changes that allowed data centers back in 2023, but some of the people in the town of 7,000 have started pushing back on these developments when they started applying for permits in 2025.

Archbald used to be at the epicenter of coal mining and transportation in the area. However, it has since turned into a quiet town with the decline of the coal industry in the early 1900s, which is why many were surprised by the developers’ interest in putting up data centers in the area.

But because a main power line cuts directly through the town, data centers would have easier access to the power they need without needing to invest significantly in grid infrastructure. The 500-kV Susquehanna-Roseland power line connects the 2.5-gigawatt Susquehanna nuclear power plant to New Jersey and goes directly over Archbald. When paired with the abundant land and fresh water available in the area, it just made sense for AI hyperscalers to build their infrastructure where the resources they need are readily available.

The community is passionately pushing back against these developments, though. Many are concerned about the effects of these sites on local utilities, as well as their potential for noise and light pollution. Communities neighboring data centers have started complaining of higher electricity rates and reduced power quality, and it has gotten to the point that President Donald Trump personally asked hyperscalers to promise that they’ll take steps to address the issue via the “ratepayer protection pledge.” Nevertheless, it seems that these commitments aren’t enough, as many local governments have started pushing back on these projects, resulting in delays costing billions of dollars.

The community’s fight against data centers in Archbald has gotten so heated to the point that four out of the seven town council members have resigned from their positions, alongside several planning board members. Some of the former councilors said that they resigned because the attacks have started going personal, especially on social media, and also cited the shooting of an Indiana politician’s home over another data center fight.

While the developers and the town government are grappling over permits, one project took advantage of a loophole and started cutting trees on their property. While this was a legal move, newly elected Archbald town council member Larry West stated that it revealed the town’s dirty past, which, according to The Washington Post, took decades for the coal dust to clear. “Now, it’s happening again, but this time, it’s data centers,” West said.

Maine Governor Janet Mills just vetoed a bill that would ban all new data center projects in the state. This moratorium applies to all data centers that exceed 20MW of consumption and would have halted all approvals until October 2027, while a state council studies their impact on the local power and water supply, as well as their effect on surrounding air quality. According to Bangor Daily News, the governor said she supported the temporary moratorium on this power-hungry infrastructure if the legislature exempted a site that’s already under construction in Jay, a town in Franklin County that sits around 60 miles north of Portland.

“A moratorium is appropriate given the impacts of massive data centers in other states on the environment and on electricity rates,” the Democrat governor said in a statement. “But the final version of this bill fails to allow for a specific project in [Jay] that enjoys strong local support from its host community and region.”

The site where the Jay data center is being built was once a paper mill that suffered a catastrophic accident, leading to its closure in 2023. It’s reported that the $550 million project will use existing infrastructure, limiting its impact on the electric grid and energy bills, while also contributing to Jay’s property tax revenue. Furthermore, it’s expected to create 800 construction jobs and open 100 additional high-paying permanent positions once the data center begins operations.

It’s reported that the Democratic governor has never had a veto overturned by the legislature, which is led by the same party. However, the Maine House of Representatives voted against the exemption, 115-29, which is more than enough to meet the two-thirds majority required to override Mills. Aside from disagreeing with the majority of the Maine Democratic Party, this also exposes her to a political challenge as she attempts to secure a nomination as the Democratic contender for a U.S. Senate seat, facing political newcomer Graham Platner.

Despite her move to block the bill, Mills said she will issue an executive order to create a state council to study the impact of data center projects in the state. She also signed a bill that put into law a measure preventing data center projects from taking advantage of the state’s business development tax incentive programs.

Residents and local governments have started pushing back against data center projects after many communities with nearby infrastructure have reported increased electricity costs and poor power quality. It has gotten so bad that President Donald Trump convened some of the biggest hyperscalers at the White House and made them promise to “pay their own way” for their electricity needs.

IBM and Arm on Thursday announced a strategic collaboration to co-develop dual-architecture enterprise platforms that would enable software designed for the Arm ecosystem to work on IBM Z mainframes and LinuxONE systems in emulation mode. The collab is designed to enable enterprises to run AI and cloud-native workloads originally developed for Arm on mission-critical IBM Z enterprise hardware with ultimate reliability, availability, and security.

Nowadays, a lot of AI frameworks as well as data-intensive cloud-native applications are developed for the Arm ecosystem, whereas IBM Z platforms (based on the Z390x or z/Architecture ISA) excel in reliability, availability, and serviceability but have a narrower native software stack. This is why enterprises increasingly operate a mix of legacy transaction processing alongside AI inference and microservices, which are typically deployed on separate Arm or x86 servers, according to IBM.

Running Arm workloads on IBM Z is designed to enable running a broad software ecosystem on IBM's Z mainframe systems, particularly those that are based on the Telum II processor and Spyre AI accelerator, through virtualization or emulation without porting them to IBM Z, which is costly, time consuming, and not common for the modern industry that relies more on x86 and Arm and less on IBM Z. Therefore, by bringing these newer workloads onto the same system, IBM reduces architectural complexity, lowers integration overhead, and simplifies operations.

Furthermore, this approach keeps workloads close to where critical data already resides: financial systems, government databases, and high-value transactional engines, which reduces latencies, minimizes security and compliance risks, and eliminates the need to replicate datasets across external platforms.

"IBM's defining role in shaping enterprise infrastructure spans decades, showcasing the breadth and commitment required to support our clients' most intensive and sensitive workloads," said Christian Jacobi, Chief Technology Officer and IBM Fellow, IBM Systems Development. "This moment marks the latest step in our innovation journey for future generations of our IBM Z and LinuxONE systems, reinforcing our end-to-end system design as a powerful advantage."

The model is not intended for performance-hungry applications. In addition, emulation and virtualization introduce a host of additional performance penalties, so do not expect IBM Z systems running Arm workloads on Telum II CPUs and Spyre accelerators to demonstrate leading performance. That being said, enterprise decision-making does not prioritize performance per se, but rather total cost of ownership, operational stability, reliability, risk mitigation, and scalability.

As a result, the trade-off may well be justified, particularly for those companies that already use IBM Z for mission-critical workloads and yet have to run additional workloads on different types of hardware. At the end of the day, IBM customers do not want to replace all of their hardware and mission-critical applications, but rather want their already deployed hardware and software to evolve, which includes running modern applications alongside legacy software. Whether or not this could lead to eventual inclusion of Arm-based CPUs or accelerators into IBM servers is something that remains to be seen, but IBM does not talk about it at this point.

San Francisco-based startup Aikido Technologies, which is focused on building offshore wind turbines, is experimenting with adding data centers to its power platforms. According to IEEE Spectrum, the company plans to launch a 100-kilowatt unit that combines a wind turbine with an AI server off the coast of Norway in the North Sea by the end of 2026. This move will address the power and space challenges many AI hyperscalers are facing right now, especially as many projects get mired in “not in my backyard” fights.

Aikido is using a semi-submersible design for its offshore wind turbines, similar to what many oil and gas companies use when drilling in high seas. This design comes with three ballast-filled legs, filled with fresh water to help maintain buoyancy and stay upright. From there, it’s secured to the seabed via chains and anchors, ensuring that it will remain in the general area, even as the wind and ocean batter it.

The firm says that it can add up to a 3- to 4-MW data hall in the upper part of each leg, meaning each wind turbine can potentially become a 9- to 12-MW data center. The fresh-water ballast is still stored in the lower part of each leg, which is then pumped towards the AI chips for cooling. The warm water is then pumped back into the ballast, with the chilly waters of the North Sea cooling it down. It also added an air-conditioner to manage the temperature of other components that aren’t part of the water-cooling loop.

“We have this power from the wind. We have free cooling. We think we can be quite cost competitive compared to conventional data-center solutions,” Aikido CEO Sam Kanner told IEEE Spectrum. “This crunch in the next five years is an opportunity for us to prove this out and supply AI compute where it’s needed.”

However, building a wind-powered offshore data center is not without its challenges. First off, wind power isn’t exactly consistent throughout the year, so each 'data center' will have batteries for storing excess energy and delivering it in times of low production. If the lean season extends far longer than anticipated, it’s also connected to the grid, allowing it to use power from other sources. Aside from this, the sea can be quite unforgiving, and salt water is particularly corrosive, possibly leading to higher maintenance costs.

Nevertheless, experiments like this can potentially solve the power and space problems that most land-based data centers face at the moment. In fact, China has thought of a similar approach, with a wind-powered underwater data center prototype launched in Shanghai in October last year. Although this might seem like an ambitious project, it’s still far more feasible than Elon Musk’s plan of launching a million data center satellites orbiting around the earth.

Microsoft is currently looking at high-temperature superconductors (HTS) for transmitting the massive amounts of electricity that it needs for its data centers. According to the company blog, since superconductors have zero resistance, adoption of that exotic tech would mean that the HTS cables would not suffer voltage drops or generate heat as electricity travels through them.

The advantages of HTS cables means that they can be lighter and take up less space compared to traditional copper and aluminum wires. For example, overhead lines typically need 70 meters of space to prevent the electrical fields of the individual cables from interfering with each other, among other reasons. HTS cables, on the other hand, only requires a 2-meter-wide trench.

HTS has been studied for several decades now, but it seems that recent advancements have made it more viable to deploy at scale. The biggest challenge that this technology faces is the cryogenic technology required to keep the conductors at their optimal temperature. Classic elemental superconductors, like mercury, need to operate below 10 Kelvin — that’s around -263 degrees C or less than -440 degrees F. And even though HTS do not need to stay as cool as traditional ones, conductors made with those materials still require temperatures around -200 degrees C or less than -320 degrees F.

As Mark Zuckerberg predicted a couple of years ago, power constraints are one of the biggest constraints hampering AI growth. Even Microsoft CEO Satya Nadella has said that the company has idle AI GPUs in its inventory because it did not have enough electricity to install them all.

AI data centers’ massive demand for power has started affecting ordinary users, with politicians taking notice of the toll it has put on ordinary Americans. Because of this, President Donald Trump has called on AI tech companies to “pay their own way” when it comes to their power consumption.

Microsoft was the first AI hyperscaler to respond to this promising the public that it will follow its “Community-First AI Infrastructure” framework when building its data centers. The company’s first and biggest promise in this framework is “We’ll pay our way to ensure our datacenters don’t increase your electricity prices.” This means that the company will have to spend more to carry the burden of investing in the power plants and other infrastructure required for its planned data centers. But aside from just building more and more power lines and facilities, Microsoft has apparently decided that it wants to reduce the waste brought about by inefficiencies in the system

It seems that data centers’ massive electrical demand is making HTS technology economically viable to deploy, especially if it will reduce the massive amounts of space that substations and other conventional power infrastructure require. More importantly, it would allow Microsoft to build more data centers without needing more electricity from the grid as it’s still working on the research and development of small-modular reactors (which isn’t even guaranteed to work). If high-temperature superconductors become viable, it would be yet another example of the mind-boggling economics of AI expanding the frontiers of every part of the technology stack.

The competition to calculate digits of Pi was initially an informal pursuit but grew more serious over time. Our server-oriented colleagues at StorageReview have proven that storage performance can make or break a Pi run, setting their latest record at a whopping 314 million digits with a single server that ran for four months.

Calculating Pi quickly became a way to benchmark the floating-point performance of CPUs. As the calculations grew ever larger, however, the task became more complicated, as RAM, I/O architectures, and storage systems came into play. That's a point StorageReview clearly illustrated by achieving a record with a single 2U server over a four-month calculation run.

The said machine is a Dell PowerEdge R7725 unit fitted with two AMD Epyc 192-core chips, for a total of 384 cores, along with 1.5 TB of DDR5 memory, an amount that could probably buy a small country at today's prices. The storage array is where this setup shines, though, with 2.5 petabytes of storage thanks to a 40-drive array of Micron 6550 Ion SSDs at 61.44 TB each.

It's long been the case that calculating Pi to such long extents requires a significant amount of bytes to store intermediate computations. After all, you're dealing with factors that are trillions of digits long. Past approaches, such as Google's 100-trillion record in 2022, used cloud server instances, and Linus Media Group and Kioxia's 300-trillion-digit run earlier this year used a Weka cluster with shared storage. But StorageReview opted to prove a point by using plain ol' simple fast local SSDs.

It's worth noting that one key factor enabling the 314-trillion-digit run is that, unlike the previous generation, the 17th-generation Dell servers used don't have a PCIe switch in their storage backplane; instead, they use a direct connection to the CPUs' PCIe lanes. With 40 bays, that works out to 2 to 4 lanes per SSD, but that still worked out to a meaty 280 GB/s of read/write performance, much higher figures than StorageReview's own past experiments.

There were additional relevant optimizations as well. The team tweaked the machine's scratch array for the patterns generated by the y-cruncher software at large digit counts. It also changed the server's standard air-cooling configuration to a CoolIT AHx10 setup, resulting in higher steady-state load clocks for the Epyc chips.

The power consumption was only 1,600W, a high number on its own but actually a pretty impressive figure as far as efficiency is concerned. Last but not least, the OS was changed from Windows Server to Ubuntu 24.04.2, a simple switch that resulted in better I/O performance on its own. We can't help but wonder if using the latest Windows Server release with native NVMe support would have been comparable. If you're curious about more details, go ahead and read the entirety of StorageReview's write-up.

Cyberattacks on infrastructure seem to be becoming a part of daily life. The latest one hit the Romainian water management authorities, taking around 1,000 computers down, affecting 10 of the country's 11 regional offices with hostile data encryption (ransomware).

No criminal group has yet claimed the action, but the attackers left a message demanding the institution contact them in seven days. The affected management systems include email, web services, databases, and Geographic Information Systems. Predictably, Windows workstations and domain name servers were also hit.

Thankfully, Romania's taps still flow freely, as no actual water control systems were reportedly affected, and that "[control] activity [is] carried out within normal parameters, through dispatches and voice communications." That's better luck than Denmark had in 2024, when a similar infrastructure attack resulted in actual burst pipes.

Romania's National Directorate for Cyber Security (DNSC) says that the attack vector remains unidentified, but that the data encryption was performed using Windows' own BitLocker rather than some more esoteric tool. The DNSC and the Romanian Intelligence Service are investigating the matter and attempting to restore the infrastructure's systems, though no further details have yet emerged.

Although this latest attack hasn't been claimed by any particular group, at first sight it appears to follow the general pattern of similar attacks on western nations' infrastructure, in what some European countries have described as a "hybrid war."

Just last week, Denmark formally accused Russia of two cyber-attacks on its infrastructure. In 2024, the pro-Russian Z-Pentest hit Danish water control systems, managing to change water pressure and burst three pipes in Køge, south of Denmark, leaving 500 homes without water for a few hours. And in 2025, another pro-Russian group NoName057(16) executed a distributed denial-of-service attack on Danish websites ahead of elections. Germany took a similar stance earlier this month, attributing an attack on Germany's air traffic control systems in 2024 to the Russian Fancy Bear group.

Regardless of who's behind the attack, most any systems administrator will tell you that unfortunately, cyber-security is by default one of the lowest-priority items on any installation, and generally seen as as an obstacle to work. Lessons are almost always only learned when an attack lands, and even then, they're often quickly forgotten. Perhaps the silver lining of these recent cyber-attacks is increased public awareness.

There's pretty big news for Windows Server administrators. After some delays, starting with Windows Server 2025 and its latest October Cumulative update, the operating system will finally support native NVMe I/O, marking the end of an era where requests were translated to SCSI bus commands, even with the highest-powered drives.

The feature has now reached General Availability and is built right into the OS, though it's not enabled by default. Sysadmins willing to take the plunge only have to tweak a registry key, or add a group policy MSI, and they can enjoy up to 80% higher IOPS and up to 45% lower CPU utilization under a high I/O load. This should be a shot in the arm for scenarios involving high-performance file serving, virtualization, AI and ML workloads, and databases.

Those figures came from a single test setup that though powerful, didn't even seem particularly exotic for this arena: a two-socket Intel system with 208 logical cores, 128 GB of RAM, and a Solidigm D7-PS1010 3.5TB PCIe 5.0 solid-state drive. Even with a single I/O thread, the system saw gains of 45% IOPS, increasing to 78% at eight threads, and 71% at 16 threads. Meanwhile, CPU load under 4K random reads saw a 41% reduction with eight threads and 47% with 16.

Microsoft's engineering team claims that "the whole I/O processing workflow is redesigned for extreme performance." Besides the raw performance gains, the improvements to the I/O locking mechanism should result in lower latency and round-trip times overall.

A thread on Reddit contains some interesting comments from anecdotal tests; some claim they saw no difference in their systems, while others hypothesized that only PCIe 5.0 NVMe devices could really make a significant use of the improved I/O stack. As for consumer-grade drives, one tester actually saw reduced performance with a Western Digital SSD, suggesting that some drives may be tuned to the old way of doing things.

There's no word on when this feature will arrive in plain Windows 11. Given the varying quality of consumer drives' firmware, there will probably be a lot of testing involved to bring the new I/O stack to home PCs, and it's equally likely that it'll be enabled or disabled depending on the drive in question.

Although common home computing scenarios don't have much to gain from added I/O, there might be tangible gains in a few specific areas. First, the redesigned I/O locking mechanism may result in a smoother overall experience, with fewer instances where a single application slamming the drive still brings system responsiveness to a halt — or when multiple tasks are occurring simultaneously, such as at the end of the Windows boot process, when always-on applications are starting. Game times can also see a reduction given the lower CPU load, and DirectStorage might see renewed interest for the same reason. We'll be sure to test all of this when the time comes.

Would you like to run 235B parameter LLMs at home, but your lowly $10,000 budget restricts you to “consumer GPUs that can barely handle 70B parameter models”? This was the situation developer David Noel Ng found himself in, until he stumbled across an “obviously fake” Nvidia Grace-Hopper platform being sold on Reddit, of all places. Ng took the gamble and, according to his blog post, it paid off royally. He’s managed, with a bit of tinkering and fixing up, to get an enterprise system that would usually cost ~ $80,000 for a tenth of that sum. The included 960GB of LPDDR5X memory, alone, is now worth more than he paid for the full system. Hilariously, he even lowballed the seller, noting the original listing on Reddit was for 10,000 EUR before offering just 7,000 EUR.

Why Ng made the offer on an ‘obviously fake’ listing

As we mentioned in the intro, the deal Ng found seemed a little too good to be true. However, he researched the seller, who seemed to be a legitimate server equipment reseller within two hours' driving distance, so he quickly made an offer to get first in line.

There were some underlying, but not insurmountable, issues with the Grace Hopper system as sold, which meant it wouldn’t be widely popular on a consumer marketplace. Specifically, it was “a Frankensystem converted from liquid-cooled to air-cooled” operation. It also looked a bit of a mess, wasn’t rackable, and ran using a 48V power supply.

On the other hand, even if this were just a collection of components, the offer seemed irresistible. The specs of the system, as sold, were as follows:

• 2x Nvidia Grace-Hopper Superchip
• 2x 72-core Nvidia Grace CPU
• 2x Nvidia Hopper H100 Tensor Core GPU
• 2x 480GB of LPDDR5X memory with error-correction code (ECC)
• 2x 96GB of HBM3 memory
• 1,152GB of total fast-access memory
• NVLink-C2C: 900 GB/s of bandwidth
• Programmable from 1000W to 2000W TDP (CPU + GPU + memory)
• 1x High-efficiency 3000W PSU 230V to 48V
• 2x PCIe Gen4 M.2 22110/2280 slots on board
• 4x FHFL PCIe Gen5 x16

Naturally, the $80,000 value of the rig is probably a fairly modest estimate. Ng notes that the two H100 chips alone are "about 30-40,000 euro each."

Getting the Frankensystem working

A significant section of Ng’s blog post is devoted to receiving, cleaning, and building a new working cooling system for the Frankensystem. It makes for a fascinating read. Suffice to say, the Nvidia Hopper system, with its awesome potential, was acquired as a dusty, extremely noisy, very hot-running machine. And it was demonstrated as such before Ng took it home.

With care and patience, five liters of Isopropyl alcohol, four repurposed but cheap Arctic AiO liquid coolers, a pair of custom CNC-milled copper parts, a kilo (~2 pounds) of 3D printed parts, microscope-assisted soldering, an LED lighting strip, and some know-how, Ng eventually triumphed. You can see the finished, reassembled Grace Hopper system, pictured at the top.

Memory gold mine

Ng seems extremely happy with the finished system and its AI performance. He says he can now “run 235B parameter models at home for less than the cost of a single H100.” The cherry on the cake, though, is that since buying the system, memory prices “have become insane,” meaning that the 960GB of DDR5X in this system would cost more than Ng paid for the whole caboodle.

The Chicago Mercantile Exchange, or CME, had to pause trading activities across the globe because of a cooling issue at one of the data centers that serve it. According to CNBC, CME is the largest derivatives exchange operator in the world, so this disruption has been felt in cities as far away as London and Kuala Lumpur. This exchange lets you trade almost everything on the market, including agriculture, energy, equity indices like S&P 500 and Nikkei 225, foreign exchange, interest rates, metals, crypto, and more.

“Due to a cooling issue at CyrusOne data centers, our markets are currently halted. Support is working to resolve the issue in the near term and will advise clients of Pre-Open details as soon as they are available,” CME said on its X account. It also told CNBC in a statement, “On November 27, our CHI1 facility experienced a chiller plant failure affecting multiple cooling units. Our engineering teams, along with specialized mechanical contractors, are on-site working to restore full cooling capacity. We have successfully restarted several chillers at limited capacity and have deployed temporary cooling equipment to supplement our permanent systems.”

At the time of writing, BrokerTec US Actives, which primarily electronically trades liquid U.S. Treasury securities, and BrokerTec EU, dealing in European government bonds and repo agreements, are now open. However, all the rest of the trading systems remain down.

The timing of the outage — happening early in the morning on the last day of the week, right after a holiday — meant that it would likely have a limited impact on U.S. trading. On the other hand, Asian and European markets are expected to feel the brunt of the interruption. A Kuala Lumpur-based trader told CNBC that prices haven’t moved much since the start of the issue, and that they expect that to continue until it has been resolved.

This isn’t the first time that the exchange has had to stop trading because of an issue with its electronic systems, but this has certainly been the biggest, so far. This also revealed the risks of relying on just a single system, especially for such a massive operation that moves trillions of dollars and whose actions can affect entire economies.

Apple CEO Tim Cook took to X yesterday, October 23, to announce that Apple’s “American-made advanced servers” have begun shipping from a new Houston facility to Apple’s own data centers.

Apple has spent years outsourcing virtually all its server manufacturing abroad, but what’s being made in Houston is no generic rack; it’s the backbone of Private Cloud Compute, the company’s answer to cloud AI and a critical piece of infrastructure for scaling Apple Intelligence when local NPUs aren’t enough.

Originally announced in 2024, Apple PCC is part of a $600 billion US investment pledge by Apple that includes significant capital outlay for homegrown server capacity.

Apple has been unusually specific in describing how PCC works. When your iPhone or Mac needs to send a request off-device, that data is handed off to a clean-room OS build, verified by a chain of trust starting in a Secure Enclave. The image itself is write-protected and stripped down, with no persistent storage and no telemetry. When the job’s done, the server forgets the session ever happened.

Apple says it will publish the software images of every production PCC node and has released a Virtual Research Environment so that security researchers can inspect and attempt verification independently. Apple says even it can’t access your data once it’s inside a PCC instance — a statement that will undoubtedly be tested in the coming months as the hardware sees wider deployment.

What Apple hasn’t detailed is what’s powering these servers. The company has confirmed that it’s using “custom Apple silicon,” but hasn’t named the chip or node. Based on the capabilities and the security model, it’s likely derived from the M-series.

This lack of detail doesn’t stop Apple PCC from being a big deal. While Microsoft and Google continue to lean on traditional GPU-heavy cloud instances for AI inference, Apple is trying something different with a hybrid of on-device model execution and cloud-side fallback, built on a software stack that won’t store or log user data. The idea is to extend the reach of local AI without violating the company’s privacy commitments and without resorting to third-party hardware for back-end acceleration.

Microsoft has cancelled its initial plan of building a data center in Caledonia, Wisconsin, after residents and elected officials objected to the plan. According to the Milwaukee Journal Sentinel, Redmond is cancelling its application for the current site, although it’s still looking at building one in a different location in Caledonia or in another part of Racine County.

“Based on the community feedback we heard, we have chosen not to move forward with this site,” Microsoft said to the publication. “We remain committed to investing in Southeast Wisconsin and look forward to working with the Village of Caledonia and Racine County leaders to identify a site that aligns with community priorities and our long-term development goals.”

Microsoft’s plan for the original Caledonia site, called Project Nova, was to build a 244-acre facility containing three data center structures and a 15-acre electrical substation. Data Center Dynamics reports that the company planned to construct it a little over a mile west of the We Energies Oak Creek Power Plant. This makes it a convenient location, especially as data centers consume a ton of electricity.

The residents' complaints were not listed in the news report, but it’s not hard to imagine their concerns. Data centers are notorious for causing spikes in power bills. Even if the affected community can afford to pay more in electricity bills, they’re also known to reduce the power quality in the area, affecting the lifespan of appliances and other electrical equipment in surrounding homes.

“I commend Microsoft for listening to the community,” Village Trustee Fran Martin told Data Center Dynamics. “And for also addressing the concerns that the community had and that I had about that particular site.” While Redmond is no longer pushing forward with the current location, it seems to still be looking to build Project Nova in Southeast Wisconsin, particularly within the general area of Racine County.

Microsoft is already moving forward with another data center project just south of Caledonia in Mount Pleasant. The company dubbed it the “world’s most powerful” AI data center, set to go online next year. Microsoft CEO Satya Nadella said that it will have more performance than the current fastest supercomputer, which is likely a dig at Musk’s Colossus. Given that the xAI site in Memphis, Tennessee, already uses 300 MW of power, with plans to hit 1 GW in the near future, it’s safe to say that Microsoft’s Wisconsin data centers could use even more electricity.

Chinese company Highlander is building a server pod near Shanghai, where it will be submerged underwater to reduce the power consumption used by the facility for cooling. According to the South China Morning Post, Highlander is expected to sink the underwater pods in October. The servers will operate commercially, with state-owned institutions like China Telecom being among the first customers.

This isn’t the first underwater data center project; Microsoft concluded its own experiment off the coast of Scotland in 2024. Although Microsoft said that it has learned a lot from the project, it did not proceed with a commercial deployment of the system.

On the other hand, China started its own underwater project in Hainan in 2023, and it is still active. The Shanghai project will be the second underwater center in the country, but it will be the first one to have clients.

The biggest benefit to using ocean water to cool data centers is the massive reduction in cooling costs. “Underwater operations have inherent advantages,” Highlander vice president Yang Ye told the Morning Post. Ye also said that it will reduce cooling energy consumption by around 90%. Highlander also said that 95% of its energy requirements will come from renewable sources, further reducing its carbon footprint.

There are also significant challenges that Highlander had to overcome. “The actual completion of the underwater data center involved greater construction challenges than initially expected,” said Engineer Zhou Jun, who worked at the Shanghai project.

Salt water and electronics do not mix well, so the company had to take special precautions to protect its systems. One technique Highlander uses involves a coating with glass flakes to protect the steel capsule from corrosion. It will also have an above-sea section that serves as an access point for maintenance crews.

Although placing data centers underwater will help reduce their power consumption, some experts are concerned about the effects of ocean warming due to continuous heat output. There hasn’t been enough research on this, University of Hull Marine Ecologies Andrew Want told SCMP.

Nevertheless, the Highlander claims that an independent assessment of its test project showed that temperature changes were still within acceptable limits. But as data centers expand and hit megawatt- and gigawatt-scales, thermal pollution will become an increasingly serious matter.

Solidigm has announced what it calls the industry's first liquid-cooled enterprise SSD, the D7-PS1010, in the blade-like E.1 form factor with a PCIe 5.0 interface. The SSD board is ensconced in a wraparound cold plate, with hose connectors at the ends, making for a hot-swappable design that directly cools the controller and NAND chips.

The company is pointing this SSD at AI servers with Direct-Attached Storage (DAS, as opposed to using a separate storage). To that effect, Solidigm talks up a partnership with Supermicro, which mentions the E.1S variant of these SSDs is a perfect fit for its Nvidia HGX B300-based servers. The hot-swap design in a compact form factor also helps with maintenance, as conventional liquid-cooled designs can be a bit fiddly whenever they need servicing.

The cooling apparatus of D7-PS1010 can cool both sides of the PCB at once and should make for much-improved thermal efficiency compared to traditional fan-based designs. Solidigm claims this design ought to let designers come up with server layouts that entirely eschew fans in their storage bays, potentially making the rack units smaller than ever.

Intel pulled the wraps off its first all-E-core Xeon at Hot Chips on Monday, August 25, in Cupertino, California. Dubbed Clearwater Forest, it's the first Xeon built on the company’s next-gen 18A process.

Scheduled to ship in 2026, the processor packs a punch with 288 efficiency cores and a two-socket ceiling of 576 cores, as well as over 1,152 MB combined last-level cache. On paper, it’s Intel’s next attempt to close the performance-per-watt gap in the data center, and it’s a shot across the bow at AMD’s EPYC line.

Darkmont cores bring modest IPC gains, massive scale

Architecturally, Clearwater is a spiritual sequel to Intel’s 144-core all-E-core Xeon, Sierra Forest, but with a few notable upgrades. The new Darkmont E-cores offer a wider 3x3 decode engine, deeper out-of-order windows, and enhanced execution ports, resulting in approximately 17% better IPC than the Crestmont generation.

Four cores share 4MB of L2 cache, and the chip doubles L2 bandwidth compared to the Sierra. Multiply that across 288 cores, and you’ve got a throughput monster designed to eat multithreaded web services and AI inference jobs alive.

The choice of process matters just as much as core count. Clearwater is one of the first real-world testbeds for Intel’s 18A node, which pairs RibbonFET transistors with backside power delivery.

This is effectively Intel’s answer to the density and power efficiency questions that have enabled TSMC and AMD to dominate in recent years. By separating power and signal routing, Intel claims higher cell utilization and lower IR drop, features that hyperscalers care about when power bills stretch into megawatts.

A new definition of Xeon?

After a rapid expansion left China with excess capacity of compute power in underused data centers, the country is reevaluating its approach to data center development. Authorities are now working on a national plan to regulate growth, optimize existing resources, and link facilities into a unified computing network that can sell unused compute power to those who need it, reports Reuters. 

From boom to bust: Many Chinese datacenters sit idle, or underused

Driven by the 'Eastern Data, Western Computing' strategy, various startups built hundreds of large-scale data centers for AI and cloud computing in the People's Republic. This initiative encouraged local governments to build data centers in the country's western regions, where electricity is cheaper, to serve demand from the eastern economic zones. Analysts argued that while building in remote provinces lowers energy costs, it often compromises latencies, which in turn lowers demand for their services, as many applications are latency sensitive. 

Many data centers were built on the assumption that state-owned companies and government agencies would purchase computing power. However, demand has fallen short of expectations, leaving many data centers idle or operating at a 20% - 30% load, way below their capacity. 

And despite the weak outlook, investments continued and state spending on data centers totaled $3.4 billion in 2024, which was even more than in 2023. This year, around $1.73 billion has been spent on data center infrastructure, according to government procurement data reviewed by Reuters. 

Now, concerns are growing among local governments about the long-term viability of their investments. Over the past 18 months, more than 100 data center projects have been scrapped, a sharp increase from just 11 cancellations in all of 2023. 

As a result, the National Development and Reform Commission (NDRC) is conducting a comprehensive review of the sector, imposing new restrictions to prevent overbuilding. For example, reviews of projects are now more thorough, and there is a ban on small-scale computing infrastructure funded by local authorities. The goal is to filter out projects lacking economic justification and ensure that new facilities meet basic efficiency criteria, such as minimum utilization thresholds and purchase contracts.

Cloud platform

The global server market experienced an unprecedented surge to nearly $100 billion in the first quarter as companies heavily invested in AI-related infrastructure, according to IDC, and 'accelerated' servers running Arm-based processors comprise one of the most rapidly growing categories, with Arm-powered server shipments rising 70% this year.

It appears that the vast majority of these Arm-powered machines are Nvidia's GB200 NVL72 rack-scale solution, based on the Grace Blackwell platform, which features an Nvidia Grace CPU and eight B200 AI GPUs per server.

Overall server purchases totaled $95.2 billion in Q1 2025, reflecting a 134.1% increase from the same period in 2024. This figure represents the fastest quarterly growth ever observed in this market, according to IDC. The widespread deployment of GPU-accelerated AI servers is fueling momentum, including those used by hyperscalers, with Nvidia's Arm-based Grace CPUs contributing to a 70% year-over-year increase in Arm server shipments this year.

Based on this surge, the annual projection for the 2025 server market was revised upward to $366 billion, representing a 44.6% increase compared to the previous year.

Spending on servers based on the x86 instruction set architecture (ISA) is expected to grow 39.9% for the year, reaching $283.9 billion. Meanwhile, systems using Arm and other non-x86 CPUs will gain even more, at a rate of 63.7% year-over-year, with projected sales of $82 billion in 2025.

Sales of GPU-based AI and HPC servers are projected to grow by 46.7% in 2025, accounting for nearly half of all spending in this segment. This trend has been amplified by the need for massive computing power to support new AI workloads and training pipelines.

Arm-based platforms are also gaining momentum, with shipment volumes projected to rise 70.0% compared to 2024. By the end of 2025, Arm systems are expected to represent approximately 21.1% of total server units shipped worldwide. This is considerably lower than Arm's long-term expectations of 50% market penetration. However, 21.1% is still a huge slice of the pie, considering that the lion's share of these CPUs are Nvidia's Grace processors.

The market is projected to keep growing

Spending on servers is expected to continue rising sharply in the coming years, starting at around $249 billion in 2024 and reaching $588 billion by 2029. The largest category in this period is Accelerated x86 (AI servers based on GPUs or AI accelerators with AMD or Intel processors), which is projected to grow from $112 billion in 2024 to $324 billion by 2029.

However, Accelerated Arm machines are also set to expand rapidly, increasing more than threefold from $32 billion to $103 billion by 2029, reflecting a rather rapid adoption of Arm-based systems for AI and cloud workloads.

Keep in mind that Accelerated Arm machines released between 2027 and 2028 can use not only Nvidia's processors, but also CPU designs from the 'NVLink Fusion camp,' such as those from Fujitsu, Marvell, MediaTek, and Qualcomm. Of course, it remains to be seen whether they will be able to capture significant market share.

Accelerated Other Non-x86 (including FPGA and ASIC servers) is also set to grow, albeit modestly, reaching $31 billion in 2029. Demand for Accelerated AI servers will be driven by more advanced LLMs and LRMs, in addition to the speculation that artificial general intelligence (AGI) is possible. AGI would require even more compute performance than today's AI technologies, according to IDC.

"The Stargate project re-announcement promised to invest up to $500 billion in AI infrastructure to help create artificial general intelligence (AGI)," said Kuba Stolarski, research vice president, Worldwide Infrastructure Research.

"Shortly thereafter, the release of DeepSeek's R1 reasoning model caused concerns about the necessity of investing in so much infrastructure. [DeepSeek] R1 needed more infrastructure than was reported, and the evolution from simple chatbots to reasoning models to agentic AI will require several orders of magnitude more processing capacity, especially for inferencing. Improvements in the efficiency of model creation were expected and, in fact, a goal in the industry. Efficient models will use fewer resources, and therefore may scale better in multi-user environments, enabling high-level reasoning and possibly eventually leading to AGI."

Meanwhile, spending on traditional servers (Non-Accelerated x86) will continue to rise steadily from $91 billion to $130 billion, but it'll become a smaller share of the overall market value. Unfortunately, IDC doesn't make forecasts about the adoption of general-purpose servers based on Arm CPUs, such as those from Arm itself, custom CPUs from hyperscale CSPs, or companies like Ampere Computing.

The U.S. leads the way

Researchers from the University of Tartu in Estonia have developed a way to repurpose old smartphones, chaining them together to build "tiny data centers" used for processing data on location and in real-time. According to the University, the team said that the most sustainable environmental solution is for users to hold on to their phones for as long as possible, but the fast pace of technological change and the fickleness of fashion means that this is easier said than done. So, they decided to think outside the box and decided to find a different sustainable solution instead.

“Innovation often begins not with something new, but with a new way of thinking about the old, re-imagining its role in shaping the future,” said Pervasive Computing Associate Professor Huber Flores. What the researchers did was get four old smartphones and remove their batteries to reduce the risk of leaks and contamination. They then 3D-printed some casings and holders that will host the devices and custom circuitry to regulate the power. This setup will make it easier to reuse the base and replace the phones if and when they break.

We’re unsure what other changes the team made to the phones, but it seems that they also created custom software for them. They also likely attached external sensors to gather data, especially as the first test was done underwater. The University said that the first trial was a success, with the setup able to monitor marine life without needing human intervention. Previously, this was only possible with a scuba diver recording video and then bringing the footage to the surface for study.

Diverse edge processing possibilities

When it comes to data centers, you probably don't think of their mobility potential. However, an impressive data center cube, the DUG Nomad, is made to take on the go. While it's not quite the same as taking your laptop to a cafe, this 10-foot-tall box is designed to take a full-on data center into some seriously rugged environments, all while keeping the hardware inside comfortable with immersion cooling.

Storage Review recently got to take a close-up look at one of the 10-foot models and shared plenty of juicy details about the setup. According to DUG, the Nomad cube is designed to optimize performance for AI tasks but with the ability to operate smoothly in extreme environments. The hardware is positioned in a way that makes it easy to maintain, while the components used to build the rig are relatively standard. For example, the cube itself is constructed from your average shipping container.

You can read more about the specs on the DUG website, which has published a data sheet with the Nomad's specifications. Immersion cooling is a huge aspect of its design, as this is its primary cooling method. It relies on "specialized dielectric fluid," which aims to provide consistent temperature control and longevity for the components inside.

DUG confirms that the automotive brand Castrol provides the oil used in the data center. This helps maintain the hardware's temperature and protects it from environmental factors like oxidation or corrosion, which is crucial when operating in dusty or humid environments.

Although the hands-on look was of the 10-foot model, there are other sizes available, including 20-foot and 40-foot containers. The 10-foot unit features a 26U rack inside and uses Hypertec's Ciara Trident servers. These are designed specifically for immersion cooling and allow easy access to crucial components like power connections and storage units.

If you want to get a closer look at the DUG Nomad, check out the full review shared by Storage Review. There you'll find a comprehensive look at the components inside and get an idea of what the unit is capable of. You can also find a video showing off the DUG Nomad over at YouTube.

Computer security researchers have been left scratching their heads after applying Microsoft’s latest raft of Patch Tuesday updates. As highlighted by Will Dormann on Mastodon, April’s updates to Windows 10 and 11 have left some unexpected detritus on the C:\ drive; An empty ‘inetpub’ folder has been left behind by the update process.

The errant inetpub folder concerned Dormann, who reacted with a “LOLWUT,” as this folder is associated with systems with Microsoft’s Internet Information Services (IIS) installed. IIS is a web server platform with a long history of security vulnerabilities.

IIS was built by Microsoft to host websites, web applications, and services on your PC. You can run and test projects locally before going global with your site or app. As a service often used on public-facing sites, and as it is associated with the world's most popular desktop OS, IIS has been continuously targeted by hackers.

It is understandable that seeing an empty inetpub folder appear on updated Windows installs - where these PCs never had IIS installed before - causes alarm bells to ring in security circles. One of the key questions about this folder appearing is whether Microsoft used it for some update purpose – a purpose not mentioned in the KB5055523 release notes. Or, perhaps the folder appeared due to a bug.

Digging around Microsoft's help pages, it appears that this isn't the first time inetpub has appeared on Windows machines which have never touched ISS software. We saw similar issues discussed in 2016.

Whatever the case, and whether the sudden appearance of this empty folder is something to be suspicious of, file managing neat-freaks can still feel righteous in their annoyance about this. Tidiness is next to godliness, as Linus Torvalds might say in slightly different language involving ‘turds.’.

Nvidia showed off a mockup of its future Rubin Ultra GPUs with the NVL576 Kyber racks and infrastructure at GTC 2025. These are intended to ship in the second half of 2027, more than two years away, and yet, as an AI infrastructure company, Nvidia is already well on its way to planning how we get from where we are today to where it wants us to be in a few years. That future includes GPU servers that are so powerful that they consume up to 600kW per rack.

The current Blackwell B200 server racks already use copious amounts of power, up to 120kW per rack (give or take). The first Vera Rubin solutions, slated for the second half of 2026, will use the same infrastructure as Grace Blackwell, but the next Rubin Ultra solutions intend to quadruple the number of GPUs per rack. Along with that, we could be looking at single rack solutions that consume up to 600kW, as Jensen Huang verified during a question-and-answer session, with full SuperPODS requiring multi-megawatts of power.

Kyber is the name of the rack infrastructure that will be used for these platforms.

There are no hard specifications yet for Rubin Ultra, but there are performance targets. As discussed during the keynote and in regards to Nvidia's data center GPU roadmap going beyond Blackwell Ultra B300, Rubin NVL144 racks will offer up to 3.6 EFLOPS of FP4 inference in the second half of next year, with Rubin Ultra NVL576 racks in 2027 delivering up to 15 EFLOPS of FP4. It's a huge jump in compute density, along with power density.

Each Rubin Ultra rack will consist of four 'pods,' each of which will deliver more computational power than an entire Rubin NVL144 rack. Each pod will house 18 blades, and each blade will support up to eight Rubin Ultra GPUs — along with two Vera CPUs, presumably, though that wasn't explicitly stated. That's 176 GPUs per pod, and 576 per rack.

The NVLink units are getting upgrades as well and will each have three next-generation NVLink connections, whereas the current NVLink 1U rack-mount units only have two NVLink connections. Either prototypes or mockups of both the NVLink and Rubin Ultra blades were on display with the Kyber rack.

No one has provided clear power numbers, but Jensen talked about data centers in the coming years potentially needing megawatts of power per server rack. That's not Kyber, but whatever comes after could very well push beyond 1MW per rack, with Kyber targeting around 600kW if it keeps with the current 1000~1400 watts per GPU of the Blackwell series.

Qualcomm is assembling a team of developers to design server processors for the data center. Evidence of this comes via a vacancy posted on the company's own website (noticed by Longhorn). The company is looking for a server system-on-chip (SoC) security architect, and the details of the listing outline that the company is "developing reference platforms based on Qualcomm's Snapdragon SoC, delivering a comprehensive solution that includes hardware, software, reference designs, user guides, SDKs, and more."

The job listing confirms the existence of a 'Qualcomm Data Center' team that is designing a 'high-performance, energy-efficient server solution for data center applications.' While we are, of course, speculating, Qualcomm may plan to use the high-performance energy-efficient cores designed by its Nuvia team for data center applications. This same team developed the company's Snapdragon X processors for laptops, but it originally designed server processors before it was brought into the Qualcomm stable. 

In fact, Qualcomm used to develop and sell Arm-based data center CPUs back in the day without much success, which is presumably why it stopped. However, Amazon’s Graviton processors have proven that Arm-based solutions can thrive in the data center, and the market is wide open for more innovative entrants. Apparently, Amazon's success made Qualcomm rethink and return to the drawing board with data center CPUs. 

"We are seeking experienced SoC Security Architects to join our team," the job listing reads. "If you possess a deep understanding of hardware security architecture and are passionate about architecting and designing complex SoCs at advanced process nodes, we would be pleased to hear from you. This critical role involves architecting the next-generation security system and hardware infrastructure by collaborating with other platform architects to optimize overall Power, Performance, Area (PPA) efficiency and security assurance while ensuring compliance with industry standards."

Interestingly, when acquiring Nuvia, Qualcomm specifically noted that it planned to use Nuvia's Phoenix cores in its processors for client PC and not data centers, despite the fact that Nuvia originally designed its cores with the data center in mind. Nowadays, the cores dubbed 'Oryon' power Qualcomm's Snapdragon X processors for client PCs.

But now that Qualcomm has apparently changed its mind about data center products and is assembling a team of SoC developers (the core team is already there), it is reasonable to expect the company to develop a data center solution within the next couple of years. While we do not know for sure, the first devices to use Qualcomm's CPUs would likely be the company's platforms for 5G and then eventually 6G base stations, a space where the company controls the hardware and software stack. As for Qualcomm's development of chips for data centers currently dominated by x86, only time will tell if that is the company's ultimate intention. 

HPE has secured a deal exceeding $1 billion to build AI servers for the X social platform previously known as Twitter, Bloomberg reports. The company beat out prominent AI server suppliers Dell and Supermicro for the contract. Since the deal is confidential, HPE did not comment to Bloomberg.

The deal was reportedly finalized in late 2024, following a competitive bidding process involving Dell and Supermicro. HPE, historically behind Dell and Supermicro in AI server sales, has ample reason to consider this agreement a significant endorsement of its technology. While $1 billion is a lot of money even by AI industry standards, this is a big win for HPE and a significant loss for Dell and the troubled Supermicro.

Bloomberg does not provide details about the servers HPE will supply to the microblogging platform, but most big companies rely on Nvidia AI GPUs. In general, GPUs are believed to account for half of an AI server's cost, so assuming we are dealing with GPU-based servers, a $1 billion deal would imply $500 million worth of GPUs.

Assuming that HPE and X will get a Blackwell GPU at $50,000 a unit (the actual cost of these GPUs varies), X will get a cluster of AI servers with 10,000 Blackwell GPUs. Such a cluster would offer rather formidable performance: up to 90 FP4 ExaFLOPS and 45 INT8/FP8 ExaTOPS. It remains to be seen how X plans to use this computational superpower. Of course, we are speculating as we do not know what is inside those machines.

Twitter used to buy custom servers from companies like MiTAC (Tyan), Supermicro, and Wiwynn. After Elon Musk acquired Twitter in 2022, the company ceased to buy servers from MiTAC (Tyan) and Wiwynn but retained its business deals with Supermicro. Interestingly, Musk's AI venture xAI contracted Dell to build its Colossus supercomputer for artificial intelligence with 100,000 Nvidia H100 GPUs.

AMD's latest commits to the LLVM GitHub repository include mentions of a new "GFX950" GPU—likely the firm's recently announced Instinct MI355X or the now-launched MI325X accelerators, per Phoronix. These patches push early enablement for the accelerators in LLVM - to better optimize the compiler back-end and improve software compatibility.

LLVM is a collection of modular tools and libraries that optimize code for specific hardware and architectures. It acts as a language-agnostic intermediary between high-level languages and machine code. AMD's latest commits to the LLVM repository feature a new GFX950 GPU. Going by Team Red's nomenclature, Phoronix suggests that this is likely the internal codename for the MI325X or MI350 (MI355X) accelerators - though the latter is more likely since the MI325X has been available since October.

Looking into the commits, we find that AMD has added support for the "v_prng_b32 instruction," - offering hardware acceleration for random number generation and MFMA (Matrix Fused Multiply-Add) instructions for matrix-related operations in machine learning. Additionally, there are mentions of "V_CVT_F32_BF16" instructions to convert FP32 numbers to the BF16 format, and the LDS (Local Data Share) memory has been increased to 160kB.

It is great to see that AMD is prepping its accelerators for launch and we should hear more news in the coming months - possibly at CES 2025 - as the official launch window draws near. The MI355X from the MI350 family boasts 288GB of HBM3E memory - fabricated on TSMC's N3 node with support for FP4 and FP6 data types. AMD touts an 80% uplift as compared to the MI325X in FP16 and FP8 computations.

These chips will go neck and neck against Nvidia's Blackwell B300 chips by Q2 or Q3 next year. As it stands, the MI355X is expected to deliver 9.2 PetaFLOPS of FP4 compute performance - on par with Nvidia's B200 offerings. AMD is ahead in terms of memory capacity - featuring 288GB of HBM3E - presumably across eight 12-Hi stacks which is 50% more than the B200 but rumored to be on par with the B300. However, Blackwell's debut has been marred by purported overheating issues and a design flaw - which could push volume B200 supply to Q1/Q2 2025.

Elon Musk's new expensive project, the xAI Colossus AI supercomputer, has been detailed for the first time. YouTuber ServeTheHome was granted access to the Supermicro servers within the 100,000 GPU beast, showing off several facets of the supercomputer. Musk's xAI Colossus supercluster has been online for almost two months, after a 122-day assembly.

What's Inside a 100,000 GPU Cluster

Patrick from ServeTheHome takes a camera around several parts of the server, providing a birds-eye view of its operations. The finer details of the supercomputer, like its power draw and pump sizes, could not be revealed under a non-disclosure agreement, and xAI blurred and censored parts of the video before its release. The most important things, like the Supermicro GPU servers, were left mostly intact in the footage above.

The GPU servers are Nvidia HGX H100s, a server solution containing eight H100 GPUs each. The HGX H100 platform is packaged inside Supermicro's 4U Universal GPU Liquid Cooled system, providing easy hot-swappable liquid cooling to each GPU. These servers are loaded inside racks which hold eight servers each, making 64 GPUs per rack. 1U manifolds are sandwiched between each HGX H100, providing the liquid cooling the servers need. At the bottom of each rack is another Supermicro 4U unit, this time with a redundant pump system and rack monitoring system.

These racks are paired in groups of eight, making 512 GPUs per array. Each server has four redundant power supplies, with the rear of the GPU racks revealing 3-phase power supplies, Ethernet switches, and a rack-sized manifold providing all of the liquid cooling. There are over 1,500 GPU racks within the Colossus cluster, or close to 200 arrays of racks. According to Nvidia CEO Jensen Huang, the GPUs for these 200 arrays were fully installed in only three weeks.

Because of the high-bandwidth requirements of an AI supercluster constantly training models, xAI went beyond overkill for its networking interconnectivity. Each graphics card has a dedicated NIC (network interface controller) at 400GbE, with an extra 400Gb NIC per server. This means that each HGX H100 server has 3.6 Terabit per second ethernet. And yes, the entire cluster runs on Ethernet, rather than InfiniBand or other exotic connections which are standard in the supercomputing space.

Of course, a supercomputer based on training AI models like the Grok 3 chatbot needs more than just GPUs to function. Details on the storage and CPU computer servers in Colossus are more restricted. From what we can see in Patrick's video and blog post, these servers are also mostly in Supermicro chassis. Waves of NVMe-forward 1U servers with some kind of x86 platform CPU inside hold either storage and CPU compute, also with rear-entry liquid cooling.

Outside, some heavily bundled banks of Tesla Megapack batteries are seen. The start-and-stop nature of the array with its milliseconds of latency between banks was too much for the power grid or Musk's diesel generators to handle, so some amount of Tesla Megapacks (holding up to 3.9 MWh each) are used as an energy buffer between the power grid and the supercomputer.

Colossus's Use, and Musk's Supercomputer Stable

The xAI Colossus supercomputer is currently, according to Nvidia, the largest AI supercomputer in the world. While many of the world's leading supercomputers are research bays usable by many contractors or academics for studying weather patterns, disease, or other difficult compute tasks, Colossus is solely responsible for training X's (formerly Twitter) various AI models. Primarily Grok 3, Elon's "anti-woke" chatbot only available to X Premium subscribers. ServeTheHome was also told that Colossus is training AI models "of the future"; models whose uses and abilities are supposedly beyond the powers of today's flagship AI.

Colossus's first phase of construction is complete and the cluster is fully online, but it's not all done. The Memphis supercomputer will soon be upgraded to double its GPU capacity, with 50,000 more H100 GPUs and 50,000 next-gen H200 GPUs. This will also more than double its power consumption, which is already too much for Musk's 14 diesel generators added to the site in July to handle. It also falls below Musk's promise of 300,000 H200s inside Colossus, though that may become phase 3 of upgrades.

The 50,000 GPU Cortex supercomputer in the "Giga Texas" Tesla plant is also under a Musk company. Cortex is devoted to training Tesla's self-driving AI tech through camera feed and image detection alone, as well as Tesla's autonomous robots and other AI projects. Tesla will also soon see the construction of the Dojo supercomputer in Buffalo, New York, a $500 million project coming soon. With industry speculators like Baidu CEO Robin Le predicting that 99% of AI companies will crumble when the bubble pops, it remains to be seen if Musk's record-breaking AI spending will backfire or pay off.

Fujitsu is collaborating with Supermicro to build liquid-cooled servers by 2027, according to a report by The Register. These liquid cooler servers will be based on Fujitsu's upcoming ARM-based Monaka processor, which is slated to be released in the same timeframe.

The combination of liquid cooling and Fujitsu's energy-efficient Monaka chips is aimed at combating sky-high demand for data center capacity, which has become greater than what can be supplied thanks to various factors, including AI. One of the biggest obstacles preventing accelerated data center capacity development the ability to meet the growing power consumption of modern datacenter chips. By combining the efficiency of the ARM architecture with liquid cooling, Fujitsu and Supermicro hope to offer a market-leading server portfolio for their customers. 

Monaka is the name of Fujitsu's next generation ARM-based datacenter processor. The new chip is aimed at AI, HPC, and datacenter deployments featuring 150 Armv9-A cores with SVE2. Monaka is designed to take full advantage of the power efficiency of the ARM architecture, and Fujitsu has set an ambitious goal of having Monaka be twice as power-efficient as its competitors' chips — not its competitors' current chips, but those that will be made in 2026 and 2027. Monaka will be built on TSMC's 2nm fabrication process.

Fujitsu apparently originally designed these CPUs with air cooling in mind. However, the manufacturer is now shifting gears in this partnership with Supermicro. The main goal is to reduce the size of Monaka-based servers; liquid cooling paired with highly power-efficient processors allows designers to build highly compact cooling solutions. 

It's also likely that liquid cooling Monaka could result in greater power efficiency gains as compared to air cooling. Testing by SMC has revealed that Nvidia's GPU servers are 50% more power efficient when using submersion liquid cooling compared to air cooling. We don't know how exactly these servers will be setup, but Fujitsu and Supermicro have an opportunity to make some of the densest and most power-efficient servers in the world by 2027, if Fujitsu can deliver on its goals for Monaka. 

Those looking for PC management solutions like KVMs, particularly in the server space, may be interested to hear of Sipeed's new KVM expansion card, which just opened up for preorders ahead of a planned October-November shipping window [h/t LinuxGizmos]. The first 500 preordered units will ship in October, and all subsequent preorder customers will receive shipments in November. Sipeed's new expansion card, the Sipeed NanoKVM-PCIe, is powered by RISC-V architecture and offers several connectivity solutions with full 1080p streaming support if you use it over the Internet. 

A few different models are also available during this preorder window, and all enjoy a 10% discount. The base model is the NanoKVM-PCIe Eth, available for $40. Then there's NanoKVM-PCIe Eth + Wi-Fi ($45), NanoKVM-PCIe Eth + PoE ($50), and the all-inclusive NanoKVM-PCIe Eth + Wi-Fi + PoE ($55). "Commission fees" of $2-$3 also apply per model, though truthfully, this sounds like something that should have been included in pricing from the get-go.

In any case, the NanoKVM-PCIe does look pretty nice. While the connectivity and power methods will differ depending on your specific model, the flexibility is quite nice. The device can be powered off the PCIe slot, USB, 9-Pin, or PoE power. It should function even when the PC is powered off, allowing for Wake over LAN functionality.

Regarding connectivity, only the Wi-Fi models include Wi-Fi, but if you opt for those, you'll enjoy reasonably high-speed Wi-Fi 6 support once you attach an antenna. Alternately, you can stick with Ethernet, which is most common in these server- and enterprise-centric use cases, where you're expected to manage multiple PCs as terminals. Wired and Wireless will stream at 1080p and 60 FPS, with a range of 100-160 milliseconds of video delay. At 60 FPS, 160 ms of delay is equivalent to roughly 9.6 frames, so gamers shouldn't even consider something like this— not that this is targeted at them, anyway.

One particularly nifty feature of the device is a minuscule OLED display alongside the rest of the I/O that gives status readings, including current streaming resolution and FPS. It's more than a little gimmicky, but it's nice if such a thing suits your configuration needs.

The U.S. Department of Justice (DoJ) is reportedly investigating Supermicro after Hindenburg Research accused the company of manipulating its financial reports and violations of U.S. export regulations to Russia and China, reports the Wall Street Journal. Supermicro has denied any wrongdoing, but the investigation is in its early stages, and stock prices have dropped in response. The DoJ and Supermicro have not formally confirmed the ongoing probe.

DoJ reportedly probes Supermicro

Hindenburg Research published a report titled Supermicro: Fresh Evidence Of Accounting Manipulation, Sibling Self-Dealing And Sanctions Evasion At This AI High Flyer back in August. The report accused Supermicro of improper accounting practices, related-party transactions, and failure to comply with U.S. export regulations. Supermicro is accused of selling high-tech products to China and Russia. This is after Russia invaded Ukraine in early 2022 and follows strict U.S.-imposed export controls on high-performance processors bound for China. The Hindenburg investigation appears comprehensive — the short seller investor says it lasted three months and included interviews with former Supermicro employees.  

Following Hindenburg's report, Supermicro postponed filing its annual report, stating that it needed to review its internal financial controls, which raised suspicions. Now, the Department of Justice has started its own inquiry into the company, focusing on allegations of financial misconduct. According to the report, a prosecutor contacted individuals with potential knowledge of these practices, including information about a former employee who accused Supermicro of violating accounting rules.

Dealing with Russia

Following the Russian invasion of Ukraine in 2022, the U.S. implemented strict export restrictions on high-performance computers and related technology to Russia. As a result, approximately 46 companies involved in handling these products are now under U.S. sanctions, with two-thirds of these exports consisting of components deemed critical by the U.S. government, potentially intended for military use. 

Despite Supermicro's claims of halting sales and recording no revenue from Russia since the war began, the company's exports to Russia have surged threefold, according to Hindenburg's findings. 

One significant recipient of Supermicro products is Niagara Computers, a supplier linked to a major Russian supercomputer used at a previously secret and now-sanctioned Kurchatov Institute nuclear technology research center. Niagara Computers received $46.3 million in products since the start of the war. Allegedly, these sales were initially facilitated through a California-based distributor but were later funneled through three new Turkish shell companies, one of which was sanctioned for smuggling.  

Additionally, around $30 million in components were allegedly shipped through a Hong Kong-based shell entity to VneshEcoStyle, one of Russia's largest importers of dual-use technology, which is also now under sanctions. This company makes no secret that its business is to 'find the equipment abroad and to deliver it in Russia,' which essentially means evading sanctions.

Trading with China

As the U.S. government has been cracking down on technology sales to China and Supermicro has come under scrutiny, the company's business practices to supply technology to Tianxia purportedly become more sophisticated.  

Hindenburg Research's report alleges that Supermicro exported AI, HPC, and surveillance servers equipped with Nvidia processors to China, even to entities with ties to the Chinese military through resellers. 

The report suggests that Supermicro employed intermediaries (for example, using Taiwan-based Leadtek, which derives about 70% - 80% of its revenue from China) and shell companies to obscure the final destinations of its products, thus evading U.S. export control regulations. It alleges that this practice enabled Supermicro to bypass restrictions and continue business in China despite tightening export laws. 

The report suggests that these actions demonstrate systemic governance failures within Supermicro and a willingness to circumvent trade restrictions to boost revenue and profits. This, of course, raises geopolitical concerns about the role of U.S. companies in advancing the technological capacity of strategic rivals. Hindenburg suggests that Supermicro potentially risked violating U.S. laws designed to prevent sensitive technologies from aiding foreign militaries. However, the report provides no direct evidence of Supermicro selling restricted components (Nvidia's A100 or H100 processors) to China. 

Stock drops 

Supermicro has benefited greatly from the rise of artificial intelligence, with its market value soaring from $4.4 billion in recent years to $67 billion by March 2024. The company is a major supplier of AI servers, which has driven much of its recent success.  

Despite Supermicro's growth, the report has cast doubt on its business practices. Apparently, just selling AI servers to interested parties was not enough. Based on Hindenburg's research, the company sold its servers to entities possibly linked to China and Russia's military amid the war against Ukraine, in which China aids Russia with technology. 

Supermicro's stock dropped by 12% after news of the DOJ investigation became public due to the WSJ report. The AI stock boom, which had significantly boosted the company's valuation, has cooled recently as investors adjust their expectations regarding returns on AI investments. However, when it comes to Supermicro, the company's financial results were allegedly boosted by illegal business practices and supplying U.S. adversaries. 

In a status update on Monday, Alibaba Cloud services said it was still gradually restoring access to servers and data affected by an explosion and fire last week. Specifically, servers were being "carefully dried" after sprinklers, and firefighters drenched them to quench a fire. We hope that works well, as unintentional liquid cooling can harm electronics.

Alibaba Cloud monitoring first flagged "network access anomalies" at its Zone C Singapore facilities last Tuesday, September 10. It quickly became aware of a fire in progress and updated customers to reassure them that firefighters were on the scene to quench the reported lithium battery explosion and fire. It seems to have taken just 30 minutes for Alibaba to switch its cloud network and security products to use data centers elsewhere. However, it asked customers who had control over such things to migrate their production workloads ASAP.

Several hours later, on the day of the explosion and subsequent fire, Alibaba updated customers to announce some hardware was being affected by some "abnormalities in high-temperature environment." Over 12 hours after the first reports on the network abnormalities, Alibaba implemented an emergency power shutdown in the affected building zones. It said that water spray from firefighting was "posing a risk of electrical short circuits."

The Register revealed further details about the data center disaster via local Singaporean media. It recounted reports that firefighting robots were deployed - thus keeping personnel safe from any potential further explosions and toxic fumes. It also pointed out that service providers like Lazada and Bytedance (TikTok) have experienced significant disruptions while cloud resources were shuffled around.

The weekend saw the first engineering access to fire-affected machines since the day of the lithium-ion-fired catastrophe. An on-site team began preparations for drying the equipment, wiring, powering on, verification, debugging, etc.

According to Alibaba Cloud's status page for the affected Zone C data center, all services are 'normal' at the time of writing. At worst, the status page indicates that 15 of the data center's services were abnormal after the accident.

Elon Musk’s supercomputing exploits continue to press forward this week, as the technocrat shared a video of his newly renamed “Cortex” AI supercluster on X. The recent expansion to Tesla’s “Giga Texas” plant will contain 70,000 AI servers and will require 130 megawatts (MW) of cooling and power at launch, upscaling to 500 MW by 2026.

Musk’s video of the Cortex supercluster shows off the in-progress assembly of a staggering number of server racks. From the fuzzy video, the racks seem to be laid out in an array of 16 compute racks per row, with four or so non-GPU racks splitting the rows. Each computer rack holds 8 servers. Somewhere between 16-20 rows of server racks are visible in the 20-second clip, so rough napkin math estimates 2,000 GPU servers can be seen, less than 3% of the estimated full-scale deployment. 

Musk shared in Tesla’s July earnings call that the Cortex supercluster will be Tesla’s largest training cluster to date, containing “50,000 [Nvidia] H100s, plus 20,000 of our hardware.” This is a smaller number than Musk previously shared, with tweets from June estimating Cortex would house 50,000 units of Tesla’s Dojo AI hardware. Previous remarks from the Tesla CEO also suggest that Tesla’s own hardware will come online at a later date, with Cortex expected to be solely Nvidia-powered at launch. 

The Cortex training cluster is being built to “solve real-world AI,” per Elon’s Twitter. In Tesla’s Q2 2024 earnings call, this means training Tesla’s Full Self Driving (FSD) autopilot system for Tesla—which will power consumer Teslas and the promised “Cybertaxi” product—and training AI for the Optimus robot, an autonomous humanoid robot expected to begin limited production in 2025 to be used in Tesla’s manufacturing process. 

Cortex first turned heads in the press thanks to the massive fans under construction to chill the entire supercluster, shown off by Musk in June. The fan stack cools the Supermicro-provided liquid cooling solution, built to handle an eventual 500 MW of cooling and power at full power. For context, an average coal power plant may output around 600 MW of power.  

Cortex joins Elon Musk's stable of supercomputers in development. So far, the first of Musk's data centers to become operational is the Memphis Supercluster, owned by xAI and powered by 100,000 Nvidia H100s. All of Memphis' 100,000 servers are connected with a single RDMA (remote direct memory access) fabric, and are likewise cooled with help from Supermicro. Musk has also announced plans for a $500 million Dojo supercomputer in Buffalo, New York, another Tesla operation. 

The Memphis Supercluster is also expected to upgrade its H100 base to 300,000 B200 GPUs, but delays on Blackwell's production due to design flaws have pushed this massive order back by several months. As one of the largest single customers of Nvidia AI GPUs, Musk seems to be following Jensen Huang's CEO math: "The more you buy, the more you save." Time will tell whether this rings true for Musk and his supercomputer collection.

X (formerly Twitter), an Elon Musk company, is embroiled in a legal dispute with Wiwynn, a server maker that belongs to Wistron, over $61 million in unpaid hardware bills, reports Business Insider. Wiwynn is not the only company that makes such accusations against X; MiTAC, another Taiwanese server supplier, was similarly impacted in the fourth quarter of 2022, reports United Daily News.

The conflict between Twitter and Wiwynn arose after Elon Musk took control of Twitter in 2022 and implemented significant cost-saving measures, including renegotiating supplier contracts, which led to a halt in payments for server parts worth around $120 million, reports The Register. Wiwynn claims that X failed to honor a Master Purchase Agreement in 2014 under which it would submit its infrastructure requirements to Wiwynn, approve its proposed list of custom servers, and pay for it. After Musk's acquisition of the social media platform, the company allegedly stopped paying for these products and cut off communication with Wiwynn. This abrupt cessation of payments has now led to the lawsuit.

The Taiwanese firm had accumulated $120 million of server components to fulfill X's orders, expecting Twitter to take responsibility for the costs. However, after the payment stoppage, Wiwynn could only cancel about $40 million of these parts that were not delivered and then resell the remaining $19 million worth of equipment. The lawsuit seeks damages for the remaining $61 million, the cost of components custom-designed for X that could not be sold elsewhere.

Wiwynn was not alone, as MiTAC also encountered payment issues after Musk's takeover in the fourth quarter of 2022, reports UDN. The company recorded a substantial NT$1.4 billion ($43.71 million) write-down due to delayed payments from Twitter, which contributed to significant financial losses. Yet, MiTAC did not directly accuse X of not paying the bills. Eventually, the company mitigated losses by reselling or reclaiming the reserved components initially intended for Twitter, which allowed it to post strong results in Q2 2023.

X and Wiwynn have so far declined to comment on the legal proceedings.

As shipments of high-performance processors based on x86 and Arm architectures to China face challenges, the country is gradually adopting locally designed data center platforms. As a result, Chinese companies also must adopt locally developed operating systems. Tencent Cloud recently unveiled TencentOS Server V3, which supports Huawei's Kunpeng, Sugon's Hygon, and Phytium's FeiTeng CPUs, reports DigiTimes.

The TencentOS Server V3 is designed primarily for large-scale server clusters powered by China's three main server CPU lines: Arm-based Huawei's Kunpeng, x86-based Sugon's Hygon, and Arm-based Phytium's FeiTeng CPUs. The operating system optimizes CPU usage, power consumption, and memory usage. To better optimize its operating system and data centers for domestic processors, Tencent partnered with Huawei and Sugon to develop high-performance platforms for domestic databases.

Also, the TencentOS Server V3 can run GPU clusters, which will be helpful for Tencent's AI efforts. The latest version of the OS fully supports Nvidia GPU virtualization to increase the utilization of processors when they are used for simple yet resource-consuming services, such as optical character recognition (OCR). According to DigiTimes, this innovation has reduced Nvidia cluster card procurement costs by nearly 60%.

DigiTimes reports that with nearly 10 million machines in operation, TencentOS Server is one of China's most widely deployed Linux OS. However, this is not the only server-oriented Linux distribution developed in Tianxia. For example, Huawei has developed its operating system, OpenEuler. Last year, OpenEuler held a 36.8% share of the Chinese server OS market, leading over CentOS/Red Hat (20.7%), Windows (19.3%), and Ubuntu/Debian (10.1%).

Huawei began its development of EulerOS in response to U.S. sanctions, and the first open-source version was released in December 2019, followed by a commercial release in September 2021. The report says that openEuler-powered servers will gain substantial market share in various sectors, including over 50% in finance, more than 70% in public services, and over 40% in energy and power industries by 2023.

The adoption of domestic server operating systems signifies China's dedication to technological independence and improved cybersecurity. Tencent's efforts in server OS development, alongside Huawei's advancements, indicate a strong trend towards self-reliance in technology within China. This move addresses security and supply concerns and ensures greater control over critical digital infrastructure.

When it comes to data center CPUs, Nvidia is a new kid on the block that has to prove that it can develop competitive processors for servers. AMD, an established player, apparently considers Nvidia's 72-core Arm-powered Grace Hopper Superchip a competitor, which is why it decided to publish a blog that compared the performance of its EPYC processor to Nvidia's Grace platform in a series of benchmarks, claiming up to a 2x performance advantage. AMD also points out x86's compatibility advantages over chips based on the Arm architecture. While, of course, such a comparison has to be taken with a grain of salt, it gives us some idea about the performance of Nvidia's Grace CPU.

Based on tests conducted by AMD, its EPYC 9754 (128-cores) and 9654 (96 cores) processors deliver over twice the performance of Nvidia's Grace CPU Superchip (72 cores) in various workloads, including general-purpose computing, server-side Java, transactional databases, decision support systems, web servers, in-memory analytics, video encoding, and high-performance computing (HPC). 

For instance, in the SPECpower_ssj2008 benchmark, single- and dual-socket AMD EPYC systems outperform Nvidia Grace systems by approximately 2.50x and 2.75x, respectively. A dual-socket AMD EPYC 9654 system outperforms a similar Nvidia system by about 2.27x on the same tests, highlighting their superior energy efficiency. 

AMD says its EPYC processors hold over 300 world records for performance and efficiency across various benchmarks, including business applications, technical computing, data management, data analytics, digital services, media and entertainment, and infrastructure solutions. Still, considering the fact that the main purpose of Nvidia's Grace Hopper project is to run AI training and inference workloads, not exactly general-purpose server workloads, the results aren't shocking. However, Nvidia's Hopper is renowned for its prowess in AI applications.

AMD's EPYC Zen 4 processors are built on the x86-64 architecture, ensuring compatibility with a wide array of software written for x86 architecture. By contrast, Nvidia's Grace is based on the Arm architecture, and while Arm is catching up with x86 in terms of data center software support, x86 is still ahead in terms of compatibility.

AMD also notes that due to their high performance, energy efficiency, and compatibility with existing software and infrastructure, AMD's EPYC processors offer a lower total cost of ownership (TCO) than Nvidia's Grace. 

AMD's own tests show that its EPYC processors outperform Nvidia's Grace CPU Superchip across various key workloads, provide extensive compatibility due to their x86-64 architecture, and offer a proven, future-proof solution for diverse data center needs, which is why the company claims they are a better choice for general data center workloads. However, when it comes to AI training and inference, Nvidia's Grace Hopper platform can offer significant advantages over AMD's Instinct. After all, AI is what Nvidia's Grace Hopper platform was developed to address from the start.

Micron Technology has introduced its new multiplexed rank dual inline memory modules (MRDIMMs) designed to combine high performance, low latency, high capacity, and predictable power consumption for next-generation Intel Xeon 6 server platforms. As the number of cores per socket rise, the new modules will be particularly useful to ensure decent per-core memory bandwidth, which will be valuable for artificial intelligence (AI) and high-performance computing (HPC) applications. 

MRDIMMs are JEDEC-standardized next-generation memory modules that essentially feature two DDR ranks that run in a multiplexed mode to double the speed. To enable such operation, the modules are equipped not only with more memory devices, but also with an MRCD chip that allows simultaneous access to both ranks and MDB chips that enable muxing and demuxing. The host CPU interacts with an MRDIMM module at an 8800 MT/s data transfer rate (for the first generation, the next generation expected to reach 12800 MT/s), but all the components on the module work at half the rate, enabling to tighter latencies and keeping power consumption of the module in check.  

Lower latencies greatly increases the real-world performance of memory subsystems. Micron and Intel claim that a 128 GB DDR5-8800 MRDIMM can offer an up to 40% lower loaded latency than a 128 GB DDR5-6400 RDIMM. As for power consumption, keeping it in chip is important for modern server platforms as combined consumption of memory modules can be comparable or even higher than power consumption of some server CPUs. 

"Micron's latest innovative main memory solution, MRDIMM, delivers the much-needed bandwidth and capacity at lower latency to scale AI inference and HPC applications on next-generation server platforms," said Praveen Vaidyanathan, vice president and general manager of Micron’s Compute Products Group. "MRDIMMs significantly lower the amount of energy used per task while offering the same reliability, availability and serviceability capabilities and interface as RDIMMs, thus providing customers a flexible solution that scales performance. Micron’s close industry collaborations ensure seamless integration into existing server infrastructures and smooth transitions to future compute platforms." 

Micron plans to offer a comprehensive line-up of DDR5-8000 and DDR5-8800 MRDIMMs in both standard and tall form factors (TFF) based on 16 Gb, 24 Gb, and 32 Gb DDR5 DRAM devices made using its proven 1β technology. Standard-height MRDIMMs will be available in 32 GB, 64 GB, 96 GB, 128 GB, and 256 GB density without using any 3D-stacked DRAM devices, which means higher performance, lower power, and lower costs. TFF MRDIMMs with a 56.9 mm height and aimed at over 1U servers will be available in 128 GB and 128 GB capacities. 

Tall form-factor MRDIMMs have a larger surface area compared to standard-sized modules, which, for the same airflow, results in a 24% lower temperature. This design reduces the likelihood of thermal throttling and enhances energy efficiency, making them particularly suitable for AI and HPC machines. 

Upcoming MRDIMMs from Micron and other suppliers (including both first-party makers like Samsung and SK Hynix and third-party suppliers like Adata) are set to be compatible with Intel's Xeon 6 platforms. On the software side of matters, DDR5 MRDIMMs do not require any software development or turning, though expect large server OEMs like Dell, HPE, and Lenovo to run comprehensive tests to ensure compatibility and stable performance across hundreds of workloads. 

Micron's MRDIMMs are currently sampling and are expected to ship in volume in the latter half of 2024. Subsequent generations promise even greater performance, with up to 45% better memory bandwidth per channel – at 12800 MT/s — compared to current RDIMMs at 6400 MT/s. 

China is set to increase its national computing power by 30% by 2025, growing from 230 ExaFLOPS to 300 ExaFLOPS, reports The Register. This ambitious plan involves significant advancements in technology, though for now it leaves more questions than answers as gaining an additional 70 ExaFLOPS of compute power is not an easy thing to do.

Currently, China has over 8.1 million data center racks providing 230 ExaFLOPS of processing power, according to data revealed at the Global Digital Economy Conference 2024 by Wang Xiaoli from the China Academy of Information and Communications Technology. 

China's plan is to reach 300 ExaFLOPS by 2025, which is a considerable increase when we look back over previous growth rates. China's compute power grew from 180 ExaFLOPS in 2022 to 197 ExaFLOPS in August 2023, which means that the new goal will require a significant acceleration. Unfortunately, Wang Xiaoli did not say how China is going to add 70 ExaFLOPS of compute horsepower in less than a year from now.  

How China plans to add 70 ExaFLOPS of compute capacity amid sanctions by the U.S. that require companies like AMD, Intel, and Nvidia as well as their partners to get an export license to sell their high-performance processors or servers to Chinese entities, both public and private, is a big question. The country can barely produce its own high-performance processors, as contract chipmakers like Semiconductor Manufacturing International Co. (SMIC) cannot procure advanced wafer fab tools as well. 

China's plan to enhance its computing capacity aims to support broader AI deployment and economic transformation, particularly in rural areas, which means that the plan is important for the whole country. This increase in compute power is expected to drive significant technological and economic benefits across China, but, again, if China manages to get appropriate hardware. 

But while it is problematic for Chinese companies to get high-performance computing processors and servers, they can get energy storage technologies. Yovole Network, a Shanghai-based cloud computing data center service provider, has partnered with Tesla to implement its Megapack energy storage technology. Yovole Network is also incorporating advanced technologies such as hydrogen energy, photovoltaic storage, indirect evaporative cooling, and liquid cooling in its data centers, which indicates that it has some massive compute horsepower inside. These technologies are essential for supporting the increased processing power while managing energy efficiency. 

This week, Chinese CPU developer Loongson published 105 programs from 53 developers that natively support its 5000- and 6000-series processors based on the proprietary LoongArch architecture. As the list revealed, Lenovo has quietly deployed Loongson's processors in its datacenters and is running cloud services on them, reports The Register. The scale of the deployment is unclear, but the revelation highlights Lenovo's commitment to using Chinese CPUs.

For now, Lenovo offers three software packages that support Loongson's LoongArch-based platforms: Wentian WxSphere Server Virtualization System Software V8.0 (16-core 3C5000L/3C5000), Wentian WxCloud Cloud Computing Management Platform V3.0 (16-core 3C5000L/3C5000), and Wentian WxStack Hyper-converged System Software V8.0 (quad-core 3A6000). For Lenovo, this is enough to deploy Loongson's 5000-series CPUs commercially for its cloud services and prepare to deploy the next-generation Loongson's 6000-series processors. 

Loongson has quietly gained traction in China with mini PCs aimed at the channel, NAS, and the education sector. These moves align with China's increasing urgency to replace Western technology with homegrown solutions, driven by policy objectives and necessity due to U.S.-led sanctions. 

Deploying 16-core 3C5000 processors for cloud services is something new, but it shows that Lenovo is confident in these CPUs and their successors, which will feature up to 128 cores. Lenovo's support for Loongson's architecture is crucial in making Chinese hardware a viable alternative to existing enterprise technologies. This support is expected to challenge companies like AMD and Intel, especially given China's vast market, which includes major telecommunications companies with extensive customer bases. 

It is unclear whether it makes much financial sense to use 16-core CPUs for cloud services nowadays, as there are more powerful equivalents from traditional x86 CPU vendors specifically architected for such workloads. However, Lenovo needs to learn how Loongson's CPUs behave with its instances today and try out the next-generation DragonChain microarchitecture-based processors that will be rolling out over the next couple of years.  

Notably, Lenovo's software stack is not the only cloud platform in China to support Loongson's processors; there are ten more platforms from various vendors, so there are more Loongson-based cloud deployments in the country. 

Optical interconnects are crucial for next-generation AI and HPC data centers, but adding them to CPUs and GPUs is complicated, and much of the infrastructure for mass adoption is still lacking. To that end, Intel introduced its first fully integrated optical input/output (I/O) chiplet at the Optical Fiber Communication Conference (OFC) 2024. The optical compute interconnect (OCI) chiplet can be attached to CPUs and GPUs to enable high bandwidth, low power consumption, and extended-reach I/O connectivity.

Intel's OCI chiplet is one of the industry's first fully integrated optical I/O solutions for co-packaging with compute processors. This chiplet supports 64 PCIe 5.0 channels, each transmitting at 32 GT/s in both directions, totaling 4 Tbps, over distances of up to 100 meters using fiber optics. The chiplet uses dense wavelength division multiplexing (DWDM) wavelengths and consumes only five pico-Joules per bit, significantly more energy-efficient than pluggable optical transceiver modules, which consume about 15 pico-Joules per bit, according to Intel.

This device is crucial for next-generation data centers and AI/HPC applications. It will enable high-performance connections for CPU and GPU clusters, coherent memory expansion, and resource disaggregation. These features will be handy for operating supercomputers for large-scale AI models and machine learning tasks that require tremendous data bandwidth.

Traditional electrical I/O systems, which rely on copper traces for connectivity, offer high bandwidth density and low power but are limited to short distances of about one meter. In contrast, Intel's optical I/O chiplet can transmit data over much longer distances with higher efficiency and lower power consumption.

The current optical I/O chiplet is largely a prototype, and Intel is collaborating with select customers to further develop and integrate this device with next-generation systems-on-chips (SoCs) and system-in-packages (SiPs).

"The ever-increasing movement of data from server to server is straining the capabilities of today’s data center infrastructure, and current solutions are rapidly approaching the practical limits of electrical I/O performance," said Thomas Liljeberg, senior director of product Management and Strategy, Integrated Photonics Solutions (IPS) Group. "However, Intel’s groundbreaking achievement empowers customers to seamlessly integrate co-packaged silicon photonics interconnect solutions into next-generation compute systems. Our OCI chiplet boosts bandwidth, reduces power consumption and increases reach, enabling ML workload acceleration that promises to revolutionize high-performance AI infrastructure."

Intel's silicon photonics initiative is backed by over 25 years of research and extensive deployment in data centers. The company's hybrid laser-on-wafer technology and direct integration approach offer high reliability and cost efficiency, which Intel says sets it apart from competitors.

So far, Intel has shipped over 8 million photonic integrated circuits (PICs) with more than 32 million integrated on-chip lasers. These PICs were integrated into pluggable transceiver modules and used in large data centers at major cloud providers for 100, 200, and 400 Gbps applications. Next-generation PICs, supporting 200G per lane, are being developed for 800 Gbps and 1.6 Tbps applications.