At its GTC 2025 keynote in DC on Tuesday, Nvidia unveiled its next-generation Vera Rubin Superchip, comprising two Rubin GPUs for AI and HPC as well as its custom 88-core Vera CPU. All three components will be in production this time next year, Nvidia says.

"This is the next generation Rubin," said Jensen Huang, chief executive of Nvidia, at GTC. "While we are shipping GB300, we are preparing Rubin to be in production this time next year, maybe slightly earlier. […] This is just an incredibly beautiful computer. So, this is amazing, this is 100 PetaFLOPS [of FP4 performance for AI]."

Indeed, Nvidia's Superchips tend to look more like a motherboard (on an extremely thick PCB) rather than a 'chip' as they carry a general-purpose custom CPU and two high-performance compute GPUs for AI and HPC workloads. The Vera Rubin Superchip is not an exception, and the board carries Nvidia's next-generation 88-core Vera CPU surrounded by SOCAMM2 memory modules carrying LPDDR memory and two Rubin GPUs covered with two large rectangular aluminum heat spreaders.

Markings on the Rubin GPU say that they were packaged in Taiwan on the 38^th week of 2025, which is late September, something that proves that the company has been playing with the new processor for some time now. The size of the heatspreader is about the same size as the heatspreader of Blackwell processors, so we cannot figure out the exact size of GPU packaging or die sizes of compute chiplets. Meanwhile, the Vera CPU does not seem to be monolithic as it has visible internal seams, implying that we are dealing with a multi-chiplet design.

A picture of the board that Nvidia demonstrated once again reveals that each Rubin GPU is comprised of two compute chiplets, eight HBM4 memory stacks, and one or two I/O chiplets. Interestingly, but this time around, Nvidia demonstrated the Vera CPU with a very distinct I/O chiplet located next to it. Also, the image shows green features coming from the I/O pads of the CPU die, the purpose of which is unknown. Perhaps, some of Vera's I/O capabilities are enabled by external chiplets that are located beneath the CPU itself. Of course, we are speculating, but there is definitely an intrigue with the Vera processor.

Interestingly, the Vera Rubin Superchip board no longer has industry-standard slots for cabled connectors. Instead, there are two NVLink backplane connectors on top to connect GPUs to the NVLink switch, enabling scale-up scalability within a rack and three connectors on the bottom edge for power, PCIe, CXL, and so on.

In general, Nvidia's Vera Rubin Superchip board looks quite baked, so expect the unit to ship sometime in late 2026 and get deployed by early 2027.

Coming on the heels of the Vera Rubin-based supercomputers for Los Alamos National Laboratory, Nvidia announced on Tuesday at GTC 2025 that, together with partners, it would build seven ExaFLOPS-class AI supercomputers for Argonne National Laboratory. Two out of five systems will be built by Oracle and will use over 100,000 Blackwell GPUs, delivering a combined performance of up to 2,200 ExaFLOPS.

The first of five AI supercomputers for Argonne National Laboratory is Equinox, which will pack 10,000 Blackwell GPUs and serve as the first phase of the project, coming online in 2026. The second phase of the project — called Solstice — will be a 200 MW system packing over 100,000 Blackwell GPUs. The two systems will be connected to deliver an aggregate performance of 2,200 FP4 ExaFLOPS for AI computations.

"We are proud to announce that Nvidia, the U.S. Department of Energy and Oracle are partnering to build two AI factories at Argonne National Laboratories featuring Blackwell," said Dion Harris, the head of data center product marketing at Nvidia. "This collaboration aims to significantly boost America's scientific research and development productivity and establish U.S. leadership in AI. Phase one features the Equinox system, which is 10,000 Blackwell GPUs; phase two, providing 200 MW of AI infrastructure, totaling 2,200 ExaFLOPS of AI performance."

The systems will be used to build three-trillion-parameter AI simulation models as well as for classic scientific computing.

One interesting thing to note about the Equinox and Solstice supercomputers is that they will be built by Oracle, a company that nowadays is not widely known as a vendor that designs and builds completely bespoke supercomputers for customers, as traditional HPC vendors like Atos, Dell, or HPE do. Oracle's primary business emphasis is on cloud infrastructure enabling AI/HPC workloads rather than custom HPC system integration from the ground up. While Oracle has its Cloud@Customer option, these machines also run Oracle's software and are managed by the company. Whether Equinox and Solstice will be managed by Oracle remains to be seen.

In addition, the Argonne National Laboratory will expand its Argonne Leadership Computing Facility — which will be available to researchers and scientists through competitive national programs — with Nvidia-based supercomputers, including Tara, Minerva, and Janus. For now, it is unclear which platform these systems will use or whether they will be built by HPE or Oracle, but we can be sure they will deliver formidable performance.

"Argonne's collaboration with Nvidia and Oracle represents a pivotal step in advancing the nation's AI and computing infrastructure," said Paul K. Kearns, director of Argonne National Laboratory. "Through this partnership, we are building platforms that redefine performance, scalability and scientific potential. Together, we are shaping the foundation for the next generation of computing that will power discovery for decades to come."

Nvidia used to be a humble supplier of graphics processors, providing reference designs for its partners' graphics cards. As the company grew, it moved from data center-grade GPUs for AI and HPC to reference designs for servers, then to rack-scale solutions. At its GTC 2025 summit, the company announced Omniverse DSX Blueprint: a reference design for gigawatt-class AI data centers, or what Nvidia calls AI factories.

The Omniverse DSX is Nvidia's blueprint for gigawatt-scale data centers, purpose-built for training and running large AI models, which was designed and simulated using the company's Omniverse framework. The blueprint combines digital-twin simulation in Omniverse with real-world engineering data to create a unified environment where partners can plan, build, and optimize every aspect of an AI data center —from power and cooling systems to compute and networking layouts.

In practice, DSX acts as a reference design and software-driven control layer for data centers ranging from 100 MW to multi-GW capacity. It was validated at Nvidia's AI Factory Research Center in Manassas, Virginia, and underpins real deployments, including the 2 GW Switch site in Georgia and the 1.2 GW Stargate facility in Abilene, Texas.

To make its DSX reference design more adaptable for different data centers, it has two configurational frameworks, DSX Boost and DSX Flex:

• DSX Boost is an internal configuration that adjusts power management and workload distribution across the data center to achieve either about 30% lower power consumption or 30% greater GPU density per megawatt, therefore increasing token-generation throughput without physical expansion.
• DSX Flex is an external configuration linking the data center to regional power grids and renewable-energy sources to tap roughly 100 GW of under-utilized grid capacity by balancing supply and demand dynamically.

The Omniverse DSX reference design for a gigawatt-class data center is meant to enable new players to build AI factories from scratch using hardware from Nvidia and its partners, without much experience in building such facilities, which have unique requirements. For such companies, DSX Blueprint will help maintain compatibility across processors, networking, and cooling systems, and align energy requirements with minimal redesign effort.

To that end, expect the DSX blueprint to accommodate not only the current Blackwell generation of hardware, but also future product generations, including Vera Rubin, though not in the first version. While Nvidia offers some flexibility with DSX blueprints (as it does with graphics cards), customers are meant to adhere to the reference design as closely as possible.

 "Blueprints are reference designs," said Kari Briski, vice president of generative AI software for enterprise at Nvidia. "So, they are meant to be a reference of what we think is the best performance, software put together to achieve a use case. So, of course you can choose and replace as you feel necessary. But we have what we do with the blueprints as we test them end to end, we run performance and QA on them to ensure that all the components and APIs work together."

Unsuspecting YouTube viewers looking for Nvidia's GTC keynote on Tuesday might well have found themselves accidentally watching an AI-generated Jensen Huang deepfake promoting a cryptocurrency scam, after YouTube promoted the video over the official stream.

As spotted by eagle-eyed Dylan Martin on X, the stream (now disabled) was actually hosted by a channel called Offxbeatz.

"Heads up: There's a fake Nvidia GTC DC keynote stream happening now on YouTube hosted by a channel called NVIDIA Live," Martin warned. "It appears to be a deepfake of Jensen Huang promoting a "crypto mass adoption event." Obviously don't do anything to connected to that QR code."

Around 20 minutes after the real keynote went live, there were some 90,000 people watching the fake live stream, with Martin noting that it was even the top result on YouTube if you searched for Nvidia GTC DC, a perfectly plausible query for anyone trying to watch the video. At one point, there were some 95,000 people watching the fake crypto scam stream and only 12,000 people watching the real stream as Jensen Huang took to the stage to share the latest from Nvidia. (Expand the below tweet to see the thread.)

A marvel of the modern AI world we live in, Martin was even able to generate an Otter AI transcription of the scam. Fake Jensen welcomed everyone to the event before stating, "But before we get into the keynote, I've got a surprise that's too exciting to wait. We're postponing the main talk for just a moment to announce something truly special, a crypto mass adoption event that ties directly into Nvidia's mission to accelerate human progress."

Fake Jensen went on to emphasize that the move was not just a random stunt, hailing Nvidia GPUs for powering Ethereum smart contracts, high-speed Solana transactions, and efficient cross-border payments with XRP. This is, of course, all nonsense. The scam, replete with a QR code for duped viewers to engage with, offered a crypto distribution scheme, calling on viewers to send in supported cryptocurrencies.

Nvidia's GTC conference has focused heavily — almost exclusively — on AI this year. Everything shown seems to have an AI connection. And one thing we know about AI is that it needs a lot of memory, and a lot of storage to hold the increasingly large models. That's where Kioxia's new LC9 data center drive comes into play.

The LC9 uses Kioxia's BiCS 8 QLC NAND with 2 Tb dies. It's unclear how many dies are in a package, probably 16, yielding 4TB capacity packages. Even with that much density, you would still need 32 such packages to reach the 122.88 TB of capacity offered by the top LC9 model (leaving a decent amount of spare flash to help with performance and endurance).

What's new with the LC9 is that it's also a PCIe 5.0 compliant drive, and it's dual-port as well. That means it can function as a single x4 device, or it can alternatively run with two x2 links. Kioxia showed the drive running a sustained read test and pushing close to the interface's maximum 15 GB/s (give or take).

The drive isn't rated for massive amounts of data writes, with only a 0.3 DWPD (drive writes per day) endurance. That's still plenty, as it means with a 5 year warranty the drive can handle around 67,000 TBW, which is more than enough for read-intensive applications. In contrast, some data center drives are designed to accommodate multiple DWPD, especially those intended for write-heavy workloads, so the write endurance rating tells us the target usage for this drive.

The LC9 instead focuses on providing high read speeds for a lot of data. That's useful for AI models that continue to grow in size. With its voracious appetite for both memory and storage — and a lot of companies were talking about ways to offload portions of the AI stack to fast SSD storage — it feels like it's only a matter of time before someone creates a single LLM that will require the entire capacity of Kioxia's LC9 122.88 TB drive.

Jensen Huang, chief executive of Nvidia, confirmed at GTC 2025 that the company plans to spend hundreds of billions of dollars on chips made in the U.S. over the next four years, reports the Financial Times. This decision comes as the company works to lessen its reliance on Asian manufacturing due to potential tariffs under the Trump administration and geopolitical instability surrounding Taiwan. The company is also producing Blackwell systems in the US.

Nvidia to spend hundreds of billions on American chips

"We are in it," Huang said at a GTC press conference, answering a question about production at TSMC Arizona, reports Reuters. "We are now running production silicon in Arizona." Huang also confirmed to the Financial Times that Blackwell systems are being produced in the US.

Huang did not elaborate on which chips are being manufactured at TSMC’s Fab 21 in Arizona, nor did he disclose volumes that Nvidia is producing there. The phrase ‘running production silicon’ means that actual chips (not test or prototype chips) are being manufactured. It does not necessarily imply high-volume production.

But while the volumes of Nvidia’s chips made in the U.S. remain unclear, Nvidia is set to increase manufacturing in America over the course of the next four years.

“Overall, we will procure, over the course of the next four years, probably half a trillion dollars’ worth of electronics in total,” Jensen Huang told the Financial Times. “And I think we can easily see ourselves manufacturing several hundred billion of it here in the U.S.”

It should be noted that while unit sales of discrete GPUs for client PCs are generally decreasing, the die sizes of flagship graphics processors are increasing, so the silicon real estate that Nvidia produces is also expanding. Additionally, sales of Nvidia’s gigantic data center GPUs are on the rise, and the company expects this to continue for the next several years. Therefore, it is not surprising that the company plans to spend roughly $500 billion on chips in the next four years.

Nvidia is mostly known for its GPUs for client PCs and data centers. In addition to GPUs, Nvidia also designs its own CPUs, DPUs, NVLink switches, networking chips, and system-on-chips (SoCs) for vehicles as well as various embedded applications. The vast majority of Nvidia’s silicon is produced by TSMC, though some chips are made by other foundries.

However, Nvidia’s products and Nvidia-based products also use numerous other chips not designed by Nvidia or produced by TSMC.

For example, the company uses CPUs developed by AMD and Intel, as well as GDDR and HBM memory made by Micron, Samsung, and SK hynix. It also uses a variety of components from other suppliers, including retimers, system management controllers, clock generators, power management ICs, analog devices, and various controllers/microcontrollers, to name a few.

Companies producing memory (Micron, SK hynix) and other components (Analog Devices, GlobalFoundries, Texas Instruments) are all building new production capacity in the U.S. (Micron’s fab is coming online in 2027, SK hynix is expected to follow in 2028, and TI’s SM1 fab is expected to be operational in 2025).

As a result, Nvidia may increasingly source these components from American facilities. Furthermore, Nvidia’s next-generation x86 servers will use AMD or Intel CPUs that could be produced by either TSMC in Arizona or Intel in Arizona. 

That said, with the expanded production of semiconductors in the U.S., it would not be surprising if Nvidia spent several hundred billion dollars on silicon made in America in the coming years. However, a key question is whether by ‘hundreds of billions’ Jensen Huang meant closer to $125 billion or between $250 billion and $300 billion, as many new fabs in the U.S. are coming online in the latter half of the decade.

Thought it was hard to get an Nvidia RTX 50-series Blackwell GPU? Things are potentially going to get even more difficult, as Nvidia has now revealed its Blackwell RTX Pro series of graphics cards. These will target both laptops and desktops, as well as standalone PCs and data center products.

We've heard from some people at GDC/GTC that Nvidia is working to improve the supply of all of its Blackwell GPUs, and some even suggested we "might" see supply finally start to catch up to demand by May/June — meaning we could see MSRP-priced models on sale and readily available. We'll believe that when we see it.

We expected the Blackwell professional announcement, and certainly Nvidia would know it was incoming and would plan for the increased production required. But we would also say the same of the RTX 50-series launches. Nvidia should have known demand would be high, and yet the supply has been woefully insufficient.

And given the choice between shipping GB202/GB203/GB205 GPUs as consumer parts with an ostensible $550~$2,000 MSRP, or shipping professional and data center parts that can cost five times as much (or more), we suspect the latter category will be served first whenever possible.

Details on the specifications and configurations (some of them, anyway) were shared after the keynote, and the higher solutions will use the same GB202 chip as the GeForce RTX 5090, but with a dramatically altered memory configuration. The top RTX Pro 6000 will also be equipped with 24Gb (3GB) GDDR7 chips, the same chips that are currently going into the RTX 5090 Laptop GPU.

24Gb chips potentially bumps the memory configuration from 32GB on a 512-bit interface to 48GB, or from 16GB on a 256-bit interface (for lower tier parts) to 24GB, and for a 192-bit interface there will be 18GB options. But that's only part of the potential upgrade.

As we've traditionally seen with professional and data center solutions, Nvidia will ship some products with memory chips in "clamshell" mode — with GDDR7 chips on both sides of the PCB. That doubles the maximum capacity for every interface width, yielding up to 96GB for GB202 and its 512-bit interface, up to 48GB for GB203 and its 256-bit interface, and up to 36GB for GB205's 192-bit interface.

And that's not some hypothetical number. Nvidia has stated that it will have a Blackwell RTX Pro GPU with 96GB of GDDR7 memory, with ECC enabled. Laptops on the other hand look like they'll stick to similar capabilities as the RTX 50-series mobile solutions, with an RTX Pro solution sporting up to 24GB — the same as the RTX 5090 Laptop GPU, which uses the GB203 silicon.

At present, the four workstation models only use two of the Blackwell chips: GB202 and GB203. The RTX Pro 6000 uses a near-complete GB202, with 188 of the potential 192 SMs. It also has the full 128MB L2 cache, and the aforementioned 96GB of GDDR7 ECC memory, with 16 chips on each side of the PCB. It also has the full complement of four NVDEC/NVENC units for video encoding/decoding. Power use for the workstation and server variants is set to 600W max (the Server Edition has a configurable TDP), while the Max-Q variant clamps the power limit to 300W.

Nvidia didn't disclose clocks or theoretical TFLOPS on the other three RTX Pro GPUs, so we've just estimated 2.5GHz for now. The step-down RTX Pro 5000 still uses GB202, just trimmed down quite a lot. It has 110 SMs enabled, which is only 57% of the computational elements. It also has four of the 32-bit memory interfaces disabled, giving it a 384-bit interface. Nvidia sticks with the more readily available 16Gb (2GB) GDDR7 chips as well, with 12 on each side of the PCB yielding 48GB of total VRAM — with no mention of the L2 cache size, though we suspect it will have 96MB. Along with the other reductions, the RTX Pro 5000 has just two NVENC/NVDEC units enabled. It has a 300W TDP.

Nvidia's RTX Pro 4500 switches to the GB203 chip, the same chip found in the RTX 5080 and 5070 Ti. It's also a nearly complete solution, with 82 of the potential 84 SMs enabled alongside all eight memory channels. Like the RTX Pro 5000, it uses 2GB GDDR7 modules in clamshell mode, for 32GB of total VRAM. It also features two NVENC/NVDEC units, and a power limit of just 200W — surprisingly low, considering it's otherwise similar to the RTX 5080 that has a 360W TGP.

Last  up, the RTX Pro 4000 also uses the GB203 chip, with some severe trimmings in some areas. It has 70 SMs enabled, the same as the RTX 5070 Ti, but the memory interface is only 192 bits wide, the same as the RTX 5070. Nvidia sticks with the 2GB GDDR7 modules as well, which in clamshell mode gives the GPU 24GB total VRAM. Dual NVENC/NVDEC are again present, and the power limit gets slashed to just 140W.

All of the models feature a full PCIe 5.0 x16 slot, with 16-pin 12V-2x6 connectors. Nvidia had the RTX Pro 6000 variants on display, which appear to be launching first. The other RTX Pro models will presumably launch in the coming months, and it's not clear when exactly the laptop solutions will become available.

Nvidia also announced a change in branding, if you didn't notice. Where the previous professional and data center solutions were sold under various RTX names (RTX A6000/A5500/A5000/etc. for Ampere, then RTX 6000/5000/4500/etc. for Ada), the new Blackwell generation of professional and data center GPUs will use RTX Pro nomenclature.

Nvidia already listed the RTX Pro 6000/5000/4500/4000 series GPU names for desktops. For laptops, there will be RTX Pro 5000/4000/3000/2000/1000/500 models, and for data center so far there's only an RTX Pro 6000. That last will, naturally, be the full-fat model with 96GB of GDDR7 ECC memory, taking over from the Nvidia L40.

We're glad to see the change in naming, as things were becoming a bit obfuscated after Nvidia killed off its Quadro branding several generations back. Now, RTX Pro will very clearly indicate that something is different from the standard GeForce RTX lineup.

There's still the matter of those numbers, however. RTX Pro 6000 quite obviously implies a Blackwell GPU for now, but what will happen with the next generation Rubin (or whatever the codename ends up being for the non-DC parts) GPUs? This is where the Ampere RTX A6000 branding made sense, but Ada parts dropped the "A" and things became a little fuzzy. Hopefully, that's something Nvidia addresses when those future products finally arrive.

Solidigm, a US-based subsidiary of SK hynix specializing in storage solutions for data centers has introduced what it claims to be the world's first liquid-cooled eSSD (Enterprise SSD) solution for AI servers. Solidigm presented its D7-PS1010 E1.S eSSDs at GTC 2025, paving the way for fully liquid-cooled servers in the future.

SSDs are an extremely vital component of servers designed for AI due to their fast data access requirements. While SSDs are prone to endurance limits, AI workloads are generally more read-intensive. This coupled with the lack of mechanical parts, higher efficiency, and predictable response times makes SSDs the go-to choice for AI-centric data centers. In an attempt to foster fanless server designs, Solidigm has pioneered the world's first liquid-cooled eSSD solution.

Traditional DLC (Direct Liquid Cooling) solutions for eSSDs fail to adequately cool both sides of drives. Additionally, the design is not hot-swappable, incurring downtime during replacements. Solidigm's D7-PS1010 E1.S eSSD overcomes these limitations as its cold plates have been engineered to keep temperatures at both sides under check. Furthermore, the design is highly serviceable, allowing technicians to hot-swap SSDs from the rear using a spring-loaded mechanism.

Solidigm says these eSSDs can eliminate fan-based 1U rack designs, decreasing HVAC and air-cooling costs for data centers. Industry-wide adoption of D2C (Direct to Chip) cooling solutions can set a standard for future data centers to be built with liquid-cooling infrastructure in mind. Just a few days back, CoolIT revealed a single-phase DLC cold plate, advertised to dissipate 4kW of heat for next-gen AI accelerations, like Nvidia's Blackwell Ultra B300 chips.

The Solidigm D7-PS1010 E1.S will also launch in a 15mm form factor for already-existing air-cooled servers and other storage solutions. Servers still employ certain components that are not liquid-cooled, wherein heat from these components must be dissipated through HVACs and other air-cooling solutions. We're still faraway from immersion cooling, however. While they have several advantages of traditional DLC solutions, existing infrastructure cannot be easily molded to suit their requirements.

Attendees at Nvidia's GTC event can see this SSD at Booth #1602. Solidigm says the D7-PS1010 E1.S is set to launch in the second half of this year for AI servers, presumably alongside Nvidia's B300 accelerators.

Jensen Huang, chief executive of Nvidia, stated that his company had not been approached to participate in a group effort to acquire a stake in a company that would operate Intel's foundry unit. He dismissed claims that Nvidia was working with industry peers and TSMC on such a deal at a press conference at GTC, reports Reuters. 

Separately, Paul Liu, a TSMC board member and the head of Taiwan's National Development Council, denied claims that the company is considering purchasing Intel's struggling foundry unit, reports DigiTimes.

"Nobody has invited us to a consortium," Huang said, according to Reuters. "Nobody invited me. Maybe other people are involved, but I do not know. There might be a party. I was not invited." 

While speaking to Taiwan's Legislative Yuan Economic Committee on March 19, Paul Liu stated that the acquisition of Intel Foundry has never been discussed at the board level and compared it to mixing two incompatible substances, something that one in the semiconductor industry can consider both figuratively as Intel and TSMC have vastly different corporate cultures and literally as the two companies use different chemical substances for manufacturing. As a result, industry experts believe such an acquisition would be more harmful than beneficial to TSMC.

Liu explained that stabilizing the company's most advanced production technologies already takes between 18 and 24 months. Exporting such a node to the the U.S. would add another year, making American operations three years behind those in Taiwan, according to Liu. He also addressed concerns about a possible second Trump presidency leading to a shift of Taiwan's semiconductor industry to the U.S., urging lawmakers to see the U.S. as a partner (which accounts for the lion's share of TSMC's revenue) rather than a competitor.

Earlier this year there were rumors that the U.S. government pushed TSMC to take over Intel Foundry and operate it. The plan involves Intel spinning off its Intel Foundry unit, which makes chips for itself and external clients. TSMC would acquire less than half of the new entity, while the remaining shares would go to industry partners. The list of industry partners included AMD, Broadcom, Nvidia, and Qualcomm, according to Reuters. However, speaking at the conference, Huang denied any involvement in discussions about a potential consortium to take control of Intel's fabs. 

For Nvidia, which produces billions of dollars worth wafers every year, a dual sourcing supply strategy could make sense. However, designing large AI/HPC GPUs for different process technologies used by Intel and TSMC would significantly affect Nvidia's costs. Investing in a fab joint venture would also be odd for the company, which started as — and remains — a fabless chip designer.

The talks reportedly started before TSMC's March 3 announcement of its $100 billion U.S. investment. This plan includes five additional Fab 21 modules, two advanced packaging facilities, and a research center. However, according to Reuters' own report, the conversations about the fab joint venture continued after the announcement as TSMC was seeking agreements with major fabless chip design companies. As a result, Taiwanese lawmakers raised concerns following such media reports and pressured Liu for clarifications.

The Nvidia Blackwell RTX Pro 6000 GPU was announced during the GTC 2025 keynote. These will use the same GB202 die that goes into Nvidia's RTX 5090 graphics card, but with some significant changes in some of the other aspects. There will be three variants of the RTX Pro 6000: the Blackwell Workstation Edition, Max-Q Workstation Edition, and Blackwell Server Edition.

The core specifications for the RTX Pro 6000 are the same across all three models. You get 188 SMs enabled, out of a potential 192 maximum from GB202. That's 10.6% more SMs, shader cores, tensor cores, RT cores, etc., relative to the RTX 5090. Clock speeds weren't given, but Nvidia does list up to 125 TFLOPS of FP32 compute via the shaders, and 4000 AI TOPS from the tensor cores. That works out to a boost clock of around 2.6 GHz, but that won't be the same for all three variants.

The RTX Pro 6000 features the full 128MB L2 cache of GB202, along with four NVENC and four NVDEC video blocks. RTX 5090 only has 96MB of L2 cache and three each for NVENC/NVDEC. It's very close to a fully enabled chip, with only 2% of the SMs disabled.

The memory configuration is the same for all three variants. As discussed in the initial RTX Pro 6000 announcement, Nvidia uses 24Gb (3GB) GDDR7 chips rather than the 2GB chips used on the consumer GeForce RTX 50-series cards. That increases the memory capacity to 48GB per PCB side, and with chips on both sides of the PCB in 'clamshell' mode, there's 96GB total. The memory has the same 28 Gbps clocks as most of the 50-series parts, with 1792 GB/s of total bandwidth.

The Blackwell Workstation Edition looks basically the same as the RTX 5090, except with a glossy black finish in places rather than a matte black. TDP (TGP) for the card is 600W, 25W higher than the 5090, but otherwise, the two cards look about the same. You also get four DisplayPort 2.1b outputs, whereas the 5090 typically offers at least one HDMI 2.1b output.

For the Max-Q Workstation Edition, the TGP gets capped at 300W. Half the power will naturally mean lower typical boost clocks for a lot of workloads, though there will undoubtedly be cases where it will still run nearly as fast as the 600W card. It also has a standard FHFL (full-height, full-length) dual-slot form factor with dual-blower fans at the back of the card. It also has four DP2.1b outputs.

Finally, the Blackwell Server Edition has a similar form factor to the Max-Q card but ditches the fans, instead relying on the server fans to provide airflow and cooling. That's usually in ample supply for servers, and noise levels are usually less of a concern — you get high RPM fans moving lots of air in a regulated environment to make everything run sufficiently cool. The power on the Server Edition is configurable up to 600W, so some installations might opt for lower power to optimize the efficiency if they're power-limited.

All three models use the same 16-pin connector found on desktop RTX cards. Servers and workstations tend to be built to much tighter specifications, and so far there haven't been any widespread reports of servers or workstations with melting connectors. That suggests perhaps that the biggest issues with 16-pin connectors are component quality and proper installation — companies are less likely to cheap out on the cables in a server or workstation, so there aren't impurities causing hot spots and melting.

Pricing hasn't been discussed, but we typically see professional and server solutions like the RTX Pro 6000 selling for 4X~5X more than the equivalent consumer GPUs. It wouldn't be surprising if the various RTX Pro 6000 cards cost $10,000 or more. We'll find out exactly where they fall in the coming days.

An Asus exec at GTC 2025 has shown off what might be both the ultimate PC collector's item and the most valuable consumer graphics card on Earth. Asus Director of Marketing, Ernest Cheng, shared a picture of one of the extravagant ROG Astral GeForce RTX 5090 Dhahab OC Edition graphics cards. Lifting this to a higher level of geek Nirvana, this particular sample bears Nvidia CEO Jensen Huang's autograph. We think this lavish gold graphics card sample is probably worth over $16,000 after looking at previous auctions.

As Cheng points out above, this 'Golden ROG RTX 5090 Astral' is now one of its kind, bearing Jensen's autograph and his "RTX ON!" inscription. However, it was already one of the rarest graphics cards among the rare-as-hen's-teeth RTX 5090 hardware that has reached retail.

As a reminder of the calibre of product we have in the photo, you can check back through our extensive RTX 5090 review, and recap the ROG Astral GeForce RTX 5090 Dhahab OC Edition story we ran in February (linked top).

The ROG Astral line is a new quad-fan flagship family from Asus, and an RTX 5090 OC model carries an official MSRP of $3,099, which is over 50% more expensive than a reference design. Now, add the golden shroud and design featuring detailed engravings of skyscrapers, camels, and Arabic calligraphy – and you have true opulence. This Middle Eastern exclusive will obviously carry a further price premium to the $3,099. Add in Jensen's autograph, and the sample graphics card in our picture will surely be worth a king's ransom.

One thing collectors will be aware of, though, is an overabundance of signatures can devalue the addition of any autograph. Jensen seems to be pretty open to signing things from fans willy-nilly, which is nice but could limit future collector valuations. Already this year we have seen the official Nvidia GeForce social media channels offer up several historical GPUs accompanied by the CEO's scrawl.

We also reported on Der8auer buying a Jensen-autographed Asus ROG Matrix GeForce RTX 4090 (with proceeds to charity) last year. That raised $16,000, which might help draw a baseline expectation on the sum the headlining 'Golden ROG RTX 5090 Astral' from the GTC 2025 event could achieve.

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.

At GTC 2025, former Intel CEO Pat Gelsinger reiterated his oft-repeated claim that Nvidia CEO Jensen Huang 'got lucky' with the AI revolution but explained his rationale more in-depth.

As GPUs have taken center stage in AI innovation, Nvidia is now one of the world's most valuable companies, while Intel is struggling. But it was not always this way. Fifteen to twenty years ago, Intel CPUs were the dominant force in computing as they handled all major workloads. During this time, Intel missed its AI and HPC opportunities with Larrabee, a project that attempted to build a GPU using the x86 CPU ISA. In contrast, Nvidia bet on purebred GPUs, said Pat Gelsinger, former CTO and CEO of Intel, while speaking at Nvidia's GTC 2025.

"The CPU was the king of the hill [in the mid-2000s], and I applaud Jensen for his tenacity of just saying, 'No, I am not trying to build one of those; I am trying to deliver against the workload starting in graphics," said Gelsinger. "You know, it became this broader view. And then he got lucky with AI, and one time I was debating with him, he said, 'No, I got really lucky with AI workload because it just demanded that type of architecture.' That is where the center of application development is [right now].

One of the reasons why Larrabee was canceled as a GPU in 2009 was that it was not competitive as a graphics processor against AMD's and Nvidia's graphics solutions at the time. To some extent, this was due to Intel's desire for Larrabee to feature ultimate programmability, which led to its lack of crucial fixed-function GPU parts such as raster operations units. This affected performance and increased the complexity of software development.

"I had a project that was well known in the industry called Larrabee and which was trying to bridge the programmability of the CPU with a throughput oriented architecture [of a GPU], and I think had Intel stay on that path, you know, the future could have been different," said Gelsinger during a webcast. "I give Jensen a lot of credit [as] he just stayed true to that throughput computing or accelerated [vision]."

Unlike GPUs from AMD and Nvidia, which use proprietary instruction set architectures (ISAs), Intel's Larrabee used the x86 ISA with Larrabee-specific extensions. This provided an advantage for parallelized general-purpose computing workloads but was a disadvantage for graphics applications. As a result, Larrabee was reintroduced as the Xeon Phi processor, first aimed at supercomputing workloads in 2010. However, it gained little traction as traditional GPU architectures gained general-purpose computing capabilities via the CUDA framework, as well as the OpenCL/Vulkan and DirectCompute APIs, which were easier to scale in terms of performance. After the Xeon Phi 'Knights Mill' failed to meet expectations, Intel dropped the Xeon Phi project in favor of data center GPUs for HPC and specialized ASICs for AI between 2018 and 2019.

To a large extent, Larrabee and its successors in the Xeon Phi line failed because they were based on a CPU ISA that did not scale well for graphics, AI, or HPC. Larrabee's failure was set in motion in the mid-2000s when CPUs were still dominant, and Intel's technical leads thought that x86 was a way to go. Fast forward to today, and Intel's attempts at adopting a more conventional GPU design for AI have largely failed, with the company recently canceling its Falcon Shores GPUs for data centers. Instead, the company is pinning its hopes on its next-gen Jaguar Shores that isn't slated for release until next year.

At GTC 2025, Nvidia introduced its DGX Station workstation platform that packs its upcoming GB300 Desktop Superchip that combines a Grace CPU with a Blackwell GPU for AI. The machine is aimed at software developers, researchers, and data scientists and will be available later this year from various workstation OEMs. 

Nvidia's DGX station workstation platform carries the GB300 Desktop Superchip (the first mention of a desktop-grade 'Superchip' that we see) that comprises a Grace CPU that connects using NVLink C2C interface with Nvidia's Blackwell Ultra GPU (which comes in an SXM form-factor) featuring the latest-generation Tensor Cores with enhanced FP4 precision. The machine is set to feature 784 GB of unified memory between the CPU's LPDDR5X and GPU's HBM3E, which will be handy for AI workloads. 

By launching its DGX Station based on the GB300 Desktop Superchip platform as well as the DGX Spark powered by the GB10 Grace Blackwell Superchip, Nvidia sets the stage for its Arm-based desktop workstation platforms that will be aimed at broader market segments beyond data scientists, researchers, and software developers. 

Nvidia has yet to disclose the 'Desktop Superchip' specifications. For now, it is reasonable to assume that the company calls 'Desktop Superchip' a combination of Grace CPU and Blackwell Ultra GPU components in configurations optimized for desktop PCs. In particular, we are talking about power consumption. Speaking of which, the motherboard has regular ATX + EPS12V power connectors for the CPU and other components and three 12V-2×6 (H++) connectors that can theoretically deliver up to 1800W to the GPU. 

In addition, the motherboard has three PCIe x16 slots for add-in-boards, three M.2 slots for SSDs, audio connectors, and USB connectors. 

For connectivity, DGX Station is equipped with Nvidia's ConnectX-8 SuperNIC, a networking component that supports speeds of up to 800 Gb/s to link multiple DGX Stations for collaborative AI projects. The high-speed networking also ensures smooth scaling for users working with extensive AI models or distributed computing tasks. 

Nvidia didn't disclose the recommended pricing of its DGX Station, which will be sold by Asus, Boxx, Dell, HP, Lambda, Lenovo, and Supermicro. Keeping in mind that each compute GPU in an SXM form factor costs tens of thousands of dollars, the DGX Station will likely cost a five-digit sum.

Nvidia CEO Jensen Huang was spotted hanging out at Denny’s mobile diner hours before giving his keynote at GTC 2025. The billionaire was said to be trying out the restaurant’s limited-time special, ‘Nvidia Breakfast Bytes,’ which honors Huang, who once worked at the chain as a dishwasher, busboy, and waiter.

The menu item consists of four sausage links served with four buttermilk silver dollar pancakes and some syrup. According to the menu poster at Denny’s food truck near the GTC 2025 site (shared by The Mercury News), you should wrap the sausage with the pancake and then pour some syrup on it, like pigs-in-a-blanket style. The signboard also quotes the Nvidia CEO, saying the sausage and pancake combo helped him through long work days.

Huang also unexpectedly appeared at the Live at Nvidia GTC with Acquired broadcast when the hosts were interviewing Jon Peddie about his memories with Jensen before Nvidia became the giant it is today. During that unplanned and somewhat chaotic scene, Jensen walked up the set and asked, “Did somebody order Denny’s?” He then started serving Nvidia Breakfast Bytes to everyone at the table while talking about his time at the diner.

“I came to the United States when I was quite young, and I didn’t really know what American food was. One day, I got a job at Denny’s when I was 15 years old, and there’s all of American food — all you can eat. It was the best job in the world because they gave you two free meals a day. Unbelievable, right?” Jensen said. “And you could eat anything you want except for the New York steak. So, my favorite was sausages and pancake, and you just roll it up like a corn dog, you know what I’m talking about? You dip it with some syrup. Back in the old days, we used to call it pigs in a blanket.”

This isn’t the first time that Jensen visited a regular restaurant and dined with some of his high-profile buddies. Just last January, he pulled some of the biggest industry leaders in Taiwan into his favorite place, with the local media calling it the ‘trillion-dollar banquet.’ He’s also been spotted several times going around food markets at night just before major events like Computex, making him quite popular among the common people and giving rise to the ‘Jensanity’ phenomenon.

Nvidia announced updates to its data center roadmap for 2026 and 2027 at the company's GTC 2025 conference today, showcasing the planned configurations for the upcoming Rubin (named after astronomer Vera Rubin) and Rubin Ultra.

Even though the company has just finished bringing Blackwell B200 into full production, and has Blackwell B300 slated for the second half of 2025, Nvidia is already looking forward to the next two years and helping its partners plan for the upcoming transitions.

One of the interesting points made is that "Blackwell was named wrong." In short, Blackwell B200 actually has two dies per GPU, which CEO Jensen Huang says changes the NVLink topology.

So even though the company calls the current solution Blackwell B200 NVL72, Huang says it would have been more appropriate to call it NV144L. Which is what Nvidia will do with the upcoming Rubin solutions.

Above we have the Rubin NVL144 rack that will be drop-in compatible with the existing Blackwell NVL72 infrastructure. We have the same configuration data for the Blackwell Ultra B300 NVL72 in the second slide for comparison. Where B300 NVL72 offers 1.1 EFLOPS of dense FP4 compute, Rubin NVL144 — that's with the same 144 total GPU dies — will offer 3.6 EFLOPS of dense FP4.

Rubin will also have 1.2 ExaFLOPS of FP8 training, compared to only 0.36 ExaFLOPS for B300. Overall, it's a 3.3X improvement in compute performance.

Rubin will also mark the shift from HBM3/HBM3e to HBM4, with HBM4e used for Rubin Ultra. Memory capacity will remain at 288GB per GPU, the same as with B300, but the bandwidth will improve from 8 TB/s to 13 TB/s. There will also be a faster NVLink that will double the throughput to 260 TB/s total, and a new CX9 link between racks, with 28.8 TB/s (double the bandwidth of B300 and CX8).

The other half of the Rubin family will be the Vera CPU, replacing the current Grace CPUs. Vera will be a relatively small and compact CPU, with 88 custom ARM cores and 176 threads. It will also have a 1.8 TB/s NVLink core-to-core interface to link with the Rubin GPUs.

Rubin Ultra will land in the second half of 2027, and while the Vera CPU will remain, the GPU side of things will get another massive boost. The full rack will be replaced by a new layout, NVL576. Yes, that's up to 576 GPUs in a rack, each with an unspecified power consumption.

The inference compute with FP4 will rocket up to 15 ExaFLOPS, with 5 ExaFLOPS of FP8 training compute. It's about 4X the compute of the Rubin NVL144, which makes sense considering it's also four times as many GPUs. The GPUs will feature four GPU dies per package this time, in order to boost the compute density.

Where the NVL144 Rubin solution has 75TB total of "fast memory" (for both CPUs and GPUs) per rack, Rubin Ultra NVL576 will offer 365TB of memory. The GPUs will get HBM4e, but here things are a bit curious: Nvidia lists 4.6 PB/s of HBM4e bandwidth, but with 576 GPUs that works out to 8 TB/s per GPU. That's seemingly less bandwidth per GPU than before.

Perhaps it's a factor of how the four GPU dies are linked together? There will also be 1TB of HBM4e per four reticle-sized GPUs, with 100 PetaFLOPS of FP4 compute.

The NVLink7 interface will be 6X faster than on Rubin, with 1.5 PB/s of throughput. The CX9 interlinks will also see a 4X improvement to 115.2 TB/s between racks — possibly by quadrupling the number of links.

Obviously, there's plenty we don't yet fully know about Rubin and Rubin Ultra, but those details will get fleshed out in the future. Data centers need a lot more planning than consumer GPUs, so Nvidia has shared full details well in advance of the products being ready to ship. And it's not quite done...

After Rubin, Nvidia's next data center architecture will be named after theoretical physicist Richard Feynman. Presumably that means we'll get Richard CPUs with Feynman GPUs, if Nvidia keeps with the current pattern.

At GTC 2025, Nvidia introduced its Spectrum-X Photonics and Quantum-X Photonics networking switch platforms for exascale datacenters that use silicon photonics. The new networking switch platforms up data transfer speed to 1.6 Tb/s per port, and 400 Tb/s in the aggregate, allowing millions of GPUs to operate together seamlessly. Nvidia says the new switches offer higher bandwidth, lower power loss, and superior reliability compared to conventional networking solutions.

The Spectrum-X Photonics Ethernet and Quantum-X Photonics InfiniBand platforms deliver speeds of 1.6 Tb/s per port (twice the maximum of current top-tier copper Ethernet solutions) and total bandwidths reaching 400 Tb/s through various port configurations. Nvidia’s Spectrum-X Photonics switches come in several configurations, offering 128 ports at 800 Gb/s or 512 ports at 200 Gb/s, reaching a total bandwidth of 100 Tb/s. A higher-capacity model provides 512 ports at 800 Gb/s or 2,048 ports at 200 Gb/s, achieving 400 Tb/s throughput. The Quantum-X Photonics series features 144 ports at 800 Gb/s InfiniBand, using 200 Gb/s SerDes for efficient data transmission.

Compared to previous-generation networking solutions, Quantum-X doubles performance and increases AI compute scalability fivefold, making it suitable for high-intensity workloads and building even bigger AI clusters.

The Quantum-X InfiniBand switches include a liquid cooling system, ensuring the onboard silicon photonics chips operate at peak efficiency without overheating. As a result, the new networking platforms promise 3.5 times better energy efficiency, 10 times greater network reliability, and 63 times stronger signal integrity, reducing power consumption and improving long-term performance. Additionally, deployment speeds increase by 1.3 times, making these switches a more effective solution for hyperscale AI data centers, according to Nvidia.

Nvidia’s Spectrum-X Photonics Ethernet and Quantum-X Photonics InfiniBand platforms use TSMC’s silicon photonics platform called Compact Universal Photonic Engine (COUPE) that combines a 65nm electronic integrated circuit (EIC) with a photonic integrated circuit (PIC) using the company's SoIC-X packaging technology. Nvidia also worked with Coherent, Corning, Foxconn, Lumentum, Senko, and others to establish its own silicon photonics ecosystem with a steady supply chain that enables Nvidia and its partners to build AI clusters and data centers that were impossible before using proprietary hardware.

Nvidia expects Quantum-X InfiniBand switches to be released later in 2025, while Spectrum-X Photonics Ethernet switches will arrive in 2026.

While these advancements promise significant improvements, integrating silicon photonics at such a large scale is complex, and widespread adoption depends on organizations being willing to upgrade their existing networking infrastructure. Despite these hurdles, Nvidia’s silicon photonics technology is a major step forward in AI networking. While Nvidia yet has to disclose future plans for its networking gear, TSMC’s COUPE has a very promising roadmap, which will perhaps be followed by Nvidia.

TSMC's next iteration of silicon photonics will incorporate COUPE technology within CoWoS packaging, combining optics directly with a switch. This will allow optical connections at speeds reaching 6.4 Tb/s. The third version aims to push speeds to 12.8 Tb/s, integrating directly into the processor package.

The Nvidia Blackwell Ultra B300 data center GPU was announced today during CEO Jensen Huang's keynote at GTC 2025 in San Jose, CA. Offering 50% more memory and FP4 compute than the existing B200 solution, it raises the stakes in the race to faster and more capable AI models yet again. Nvidia says it's "built for the age of reasoning," referencing more sophisticated AI LLMs like DeepSeek R1 that do more than just regurgitate previously digested information.

Naturally, Blackwell Ultra B300 isn't just about a single GPU. Along with the base B300 building block, there will be new B300 NVL16 server rack solutions, a GB300 DGX Station, and GB300 NV72L full rack solutions. Put eight NV72L racks together, and you get the full Blackwell Ultra DGX SuperPOD: 288 Grace CPUs, 576 Blackwell Utlra GPUs, 300TB of HBM3e memory, and 11.5 ExaFLOPS of FP4. These can be linked together in supercomputer solutions that Nvidia classifies as "AI factories."

While Nvidia says that Blackwell Ultra will have 1.5X more dense FP4 compute, what isn't clear is whether other compute have scaled similarly. We would expect that to be the case, but it's possible Nvidia has done more than simply enabling more SMs, boosting clocks, and increasing the capacity of the HBM3e stacks. Clocks may be slightly slower in FP8 or FP16 modes, for example. But here are the core specs that we have, with some inference of other data (indicated by question marks).

We asked for some clarification on the performance and details for Blackwell Ultra B300 and were told: "Blackwell Ultra GPUs (in GB300 and B300) are different chips than Blackwell GPUs (GB200 and B200). Blackwell Ultra GPUs are designed to meet the demand for test-time scaling inference with a 1.5X increase in the FP4 compute." Does that mean B300 is a physically larger chip to fit more tensor cores into the package? That seems to be the case, but we're awaiting further details.

What's clear is that the new B300 GPUs will offer significantly more computational throughput than the B200. Having 50% more on-package memory will enable even larger AI models with more parameters, and the accompanying compute will certainly help.

Nvidia gave some examples of the potential performance, though these were compared to Hopper, so that muddies the waters. We'd like to see comparisons between B200 and B300 in similar configurations — with the same number of GPUs, specifically. But that's not what we have.

By leveraging FP4 instructions, using B300 alongside its new Dynamo software library to help with serving reasoning models like DeepSeek, Nvidia says an NV72L rack can deliver 30X more inference performance than a similar Hopper configuration. That figure naturally derives from improvements to multiple areas of the product stack, so the faster NVLink, increased memory, added compute, and FP4 all factor into the equation.

In a related example, Blackwell Ultra can deliver up to 1,000 tokens/second with the DeepSeek R1-671B model, and it can do so faster. Hopper, meanwhile, only offers up to 100 tokens/second. So, there's a 10X increase in throughput, cutting the time to service a larger query from 1.5 minutes down to 10 seconds.

The B300 products should begin shipping before the end of the year, sometime in the second half of the year. Presumably, there won't be any packaging snafus this time, and things won't be delayed, though Nvidia does note that it made $11 billion in revenue from Blackwell B200/B100 last fiscal year. It's a safe bet to say it expects to dramatically increase that figure for the coming year.

Asus has lifted the curtains of the Ascent GX10, the company's rendition of Nvidia's Project Digits, at GTC 2025. Leveraging the chipmaker's GB10 Grace Blackwell Superchip, the Ascent GX10 offers up to 1,000 TOPS of AI performance.

Like Project Digits, the Ascent GX10 is a mini-PC that can be placed on your desk. You just need to connect a keyboard, mouse, and monitor to it to have a powerful AI supercomputer at your disposal. Asus hasn't liberated the product page for the Ascent GX10, so all the specifications we have on the mini-PC come from the press release.

The GB10, the heart of the Ascent GX10, combines Nvidia's Grace CPU and Blackwell GPU. However, the GB10 is a shrunk-down version of Nvidia's Grace Blackwell Superchip.

The Grace CPU features a 20-core Arm design comprising 10 Cortex-X925 and 10 Cortex-A725 cores. It is connected to Nvidia's latest Blackwell GPU through a high-performance NVLink-C2C interconnect. In unison, the GB10 delivers up to 1 PFLOP (1,000 TFLOPS) of FP4 performance.

Ascent GX10, Bringing AI Power To Developer's Fingertips

“AI is transforming every industry, and the ASUS Ascent GX10 is designed to bring this transformative power to every developer’s fingertips,” said KuoWei Chao, General Manager of ASUS IoT and NUC Business Group in the press release. “By integrating the NVIDIA Grace Blackwell Superchip, we are providing a powerful yet compact tool that enables developers, data scientists, and AI researchers to innovate and push the boundaries of AI right from their desks.”

The Ascent GX10 also has 128GB of unified system memory, which allows the device to handle AI models with up to 200 billion parameters. While Asus didn't reveal the memory's specifications, it should use the same LPDDR5x as Project Digits and up to 4TB of M.2 NVMe storage with self-encryption.

The device has Nvidia's ConnectX network interface as part of its networking capabilities, meaning you can hook up to Ascend GX10 systems for larger AI models, such as Llama 3.1, which flaunts up to 405 billion parameters.

Without a product page or additional renders, we can't know for sure what kind of connectivity the Ascend GX10 will offer. For reference, Project Digits has four USB4 Type-C ports, Wi-Fi, an Ethernet port, and one HDMI 2.1a port for the display.

Asus hasn't revealed the Ascent GX10's availability or pricing. Project Digits will hit the market in May, starting at $3,000, and we expect the AScent GX10 to have similar availability and pricing.

Nvidia’s annual GPU Technology Conference (GTC) is happening today, and Jensen Huang is set to give the keynote address this morning. The multi-day event focuses on artificial intelligence, computer graphics, and other technologies that rely on GPUs' specialized computational power. The keynote address will happen live at the SAP Center in San Jose, California at 10 am Pacific Time, but it will also be live-streamed to a global audience via YouTube.

A pre-broadcast livestream, Live at Nvidia GTC with Acquired, will start on YouTube at 8 am Pacific Time. The company says this event will feature speakers who will dive into Nvidia’s over 30-year history to see how it became the AI giant it is today.

But what’s more exciting for everyone is that Huang is expected to unveil the Blackwell Ultra GPU, which has since been renamed the B300 series, that is expected to deliver more performance and have upgraded memory configurations. Huang said he will also show off next-generation Rubin AI GPUs and more at GTC.

The B300 series AI GPUs are expected to be available in the latter half of this year, while the next-generation Rubin is scheduled for 2026. Many people are anticipating the arrival of these more powerful chips, especially as tech giants and startups alike are battling for supremacy in the AI space.

Nvidia’s competitors, like AMD and Intel, also have their own AI GPU offerings. However, they are miniscule compared to Team Green, which currently owns around 92% of the entire data center GPU market. Its near-monopoly on AI GPUs, plus the hype around AI models, allowed it to become the most valuable company in the world practically overnight.

It has since dropped to third place after some market corrections, with the company losing more than half a trillion dollars in market cap after the release of DeepSeek AI. But as long as there’s demand for powerful AI GPUs, it’s unlikely that Nvidia will go away anytime soon.

It’s just a shame that many gaming enthusiasts, which was Nvidia’s primary market before AI exploded into the scene, feel that they’re being left behind by the company. While it’s understood that the company will prioritize its AI cash cow, the pricing and availability (or lack thereof) of its recently launched RTX 50-series GPUs has disappointed millions of its core fan base.

Hard drives will remain the most cost-effective storage solution for data centers for years to come, but to make them more suitable for AI data centers, Seagate is developing HDDs that use a common PCIe interface and the NVMe 2.0 protocol. At GTC 2025, Seagate demonstrated a proof-of-concept system running NVMe HDDs, NVMe SSDs, Nvidia's BlueField 3 DPU, and AIStore software to show how NVMe transforms hard drives for AI workloads. Seagate is ahead of its rivals, which are also working on NVMe HDDs.

New protocol, new performance

Hard disk drives have traditionally used specialized interfaces, such as SCSI, Parallel ATA, Serial ATA (SATA), and SATA, which are fine but are reaching their limits for modern, high-performance data environments, specifically in AI and large-scale data centers. Both SATA and SAS rely on serialized protocols developed in the 1980s that carry legacy protocol layers not suited for modern high-speed data processing. Also, SAS and SATA setups require host bus adapters and additional controller layers, adding complexity, potential points of failure, and latency. As a result, these architectures are not suitable for AI workloads, which require high-throughput, low-latency access to massive datasets.

Compared to SAS/SATA, NVMe paired with PCIe offers significantly higher bandwidth, lower latency, and better scalability. Unlike SAS/SATA, which are limited to 6-12 Gbps speeds and rely on complex layers like HBAs and expanders, NVMe operates over an industry-standard PCIe interface, supporting speeds up to 128 GB/s (one HDD is not going to need more than 1 TB/s for quite a while, but on a system level, the more bandwidth, the merrier), and greatly reducing system complexity as well as simplifying scalability. NVMe also enables direct GPU-to-storage (GPUDirect) access through DPUs bypassing CPUs (thus reducing CPU bottlenecks), and supports 64K queues with 64K commands per queue, vastly improving parallel processing, which is important for AI systems.

Proof-of-concept system

To validate this architecture, Seagate built a proof-of-concept system integrating eight NVMe HDDs, four NVMe SSDs for caching, an Nvidia Bluefield 3 DPU, and AIStore software, all running in a Seagate NVMe hybrid array enclosure.

The test machine demonstrated that direct GPU-to-storage access minimizes latency in AI workflows. Also, removing legacy SAS/SATA infrastructure simplified system architecture and increased storage efficiency. AIStore software dynamically optimizes data caching and tiering, which greatly enhances AI model training performance.

Also, Seagate says that the system can scale to exabyte levels when using NVMe-over-Fabric (NVMe-oF). Specifically, NVMe-oF integration enables seamless expansion of multi-rack AI storage clusters, which is crucial to scale efficiently (perhaps more importantly, seamlessly) across large data centers.

The trial confirmed that NVMe HDDs could support high-performance AI environments without requiring fully flash-based storage solutions, reducing costs while maintaining performance.

AI systems are driving exponential growth in data storage needs, with Existing architectures facing significant constraints. SSDs deliver high-speed performance but are financially unsustainable for long-term storage at this scale. SAS and SATA HDDs offer affordability but introduce complexity and latency due to reliance on host bus adapters (HBAs), proprietary silicon, and controller systems not optimized for AI’s high-throughput, low-latency requirements. Cloud storage options further complicate infrastructure with high WAN data transfer costs, unpredictable retrieval times, and latency spikes, which hinder AI processing efficiency. These limitations result in complex, costly, and inefficient architectures, slowing down AI adoption and performance.

Future AI storage needs

Nowadays, enterprises manage petabyte to exabyte-scale datasets for AI model training and inference. Going forward, their needs will increase, and this is where Seagate's solution to bring NVMe connectivity to HDDs and create a unified and efficient data center storage platform will shine.

From an HDD complexity point of view, adding NVMe to hard drives is not that expensive as the HDDs retain SAS/SATA physical connectors and their traditional 3.5-inch form factor. The only things that change are the addition of the NVMe protocol support and a PCIe interface to the controller (which probably costs pennies), as well as the development of firmware that supports NVMe features and capabilities like GPUDirect. Keeping in mind that the transition to NVMe/PCIe connectivity eliminates HBAs and complexity, small HDD cost price hikes will hardly be noticed by the industry.

However, as HDDs gain capacity, their IOPS-per-TB performance drops, and this may affect performance when working in AI clusters going forward (or will require more flash to mitigate). To that end, it is possible that in the future, dual-actuator HDDs, such as Seagate's Mach.2, will be preferable for AI clusters. Such drives are, of course, more expensive than regular single-actuator HDDs, but each such drive still costs less than two single-actuator HDDs, so this will not create any significant disruptions.

When?

One of the things that an avid reader would ask after learning the benefits of NVMe for HDDs is when such hard drives are set to hit the market. Unfortunately, it is hard to tell. Large companies prefer to have dual source supply for products like HDDs, which is why NVMe hard drives were developed as part of the OCP project. However, while Seagate already has NVMe hard drives, its rivals yet have to introduce their devices. When that happens, and all makers of hard drives can produce such products in volume, cloud service providers that address AI companies and workloads will start to adopt such products.

It's been a very busy year so far for GPUs, with Nvidia launching the RTX 5090, RTX 5080, RTX 5070 Ti, and RTX 5070 in the span of just two months. AMD also launched its RX 9070 XT and RX 9070, while Intel released the Arc B580 late last year and the Arc B570 in mid-January. The only problem? Outside of the Arc B570 (kind of but not really), every GPU launched so far has ended up being sold out or severely overpriced. But have no fear, because Nvidia claims it has shipped twice as many Blackwell GPUs as Ada during the first five weeks of each product series. Except that comparison is questionable, at best. 

The above chart looks pretty good, right? And then you go check the data and discover that the RTX 4090 launched on October 12, 2022, and was the only Ada GPU for the first five weeks — the RTX 4080 arrived on November 16, 2022, exactly five weeks and one day later. By comparison, the RTX 5090 and RTX 5080 both launched on January 30, 2025; the RTX 5070 Ti arrived on February 20; and the RTX 5070 just came out on March 5. So, in the first five weeks, Nvidia appears to be comparing the sole halo card from its Ada Lovelace and RTX 40-series GPUs to the first four Blackwell and RTX 50-series GPUs — or at least, that's how we read it. Maybe the launch supply of the 4080 was also factored in.

If we were to get a more direct comparison, we'd need to look at the first five weeks of the 4090 and 4080, plus the first two weeks of the 4070 Ti, plus the first day of the 4070. We think it's a safe bet that all those added together would easily eclipse the number of RTX 50-series GPUs shipped so far. Even at the most simplistic level, Blackwell had two GPUs (5090 and 5080) launch on the same day compared to one GPU for Ada (4090), so shipping twice as many would be the baseline.

This was a 'great' way to obfuscate the number of units shipped and claim to have shipped more inventory than in the past. It also completely neglects the fact that RTX 30-series GPUs were still relatively available right up to the launch of the 40-series, while the 40-series cards have been disappearing from shelves since last fall. We're told that more RTX 50-series GPUs are being produced, and Nvidia and its add-in board (AIB) partners are making them as fast as possible, but concrete numbers are not something anyone is willing to disclose.

However you want to slice it, Nvidia's latest GPUs are hard to come by. The current lowest prices online have just one RTX 5070 for $739 at Newegg, with third-party scalpers listing RTX 5070 cards at Amazon for $900 or more. RTX 5070 Ti starts at $1,149, the 5080 costs $1,609 or more, and the RTX 5090... you don't even want to know. (None are listed at PC Part Picker, but on eBay during the past 30 days the average sold 5090 at auction cost nearly $4,500, with 194 units sold.) AMD's RX 9070 series GPUs aren't doing much better, with an $853 RX 9070 on Amazon and a $939 RX 9070 XT also at Amazon. It could be a long wait for supply to catch up to demand, needless to say.

Besides making some misleading, at best, claims about Blackwell RTX availability, Nvidia also discussed the advances it's working with developers to enable DLSS 4 MFG and upscaling, Neural Shading, and ACE. RTX Remix officially left beta today as well, and there will be a public Half-Life 2 RTX demo available on March 18, offering vastly improved visuals compared to the original game that's now over 20 years old. [Ed: Where's my cane?]

You can see the rest of the announcements in the above slide deck. There's more marketing hype around the performance boost offered by DLSS 4 and Multi Frame Generation (MFG), which, as we discussed in our own in-depth testing, tends to be a highly inflated way of looking at performance. It's not that MFG is bad, per se, but even Nvidia's own numbers should cause some raised eyebrows. Like this performance result from Portal RTX:

Updates to the RTX Remix toolset combined with neural rendering features like the neural radiance cache deliver more than double the performance compared to the Remix beta. And then we look at the details.

Base RTX Remix beta performance was 125 FPS — that's with DLSS 3.5 and frame generation. Runtime optimizations and texture streaming account for a 22% increase, and there's another 18% improvement from the neural radiance cache. Together, those would result in performance of 175 FPS, an impressive 40% improvement. Then DLSS 4 MFG4X, on top of that, "increases" performance by 73%. But what it really does is drop the base non-framegen performance from ~88 FPS to ~76 FPS.

Either way, it's still running fast enough that the result should be very playable and look incredibly smooth, particularly if you have a 240 Hz 4K monitor. But that's also with an RTX 5090, which in our testing provides up to 60% higher performance than the RTX 5080 for demanding ray traced games, 75% higher performance than the RTX 5070 Ti, and 143% more performance than the RTX 5070.

By those metrics, 4K with performance upscaling and MFG4X on an RTX 5070 might only get around 125 FPS in Portal RTX on an RTX 5070, and it would feel more like 31 FPS in terms of input sampling and latency. That's still playable, but nowhere near what the MFG numbers might suggest in terms of user experience. 120 FPS via MFG isn't the same feel as a native 120 FPS, or even 120 FPS with regular framegen. And how will Half-Life 2 RTX run, given it's presumably an even more demanding game? We'll find out next week.

As we've noted in so many of our recent GPU reviews, the fundamental problem is supply and demand. The demand — and record $130 billion in revenue — for AI and data center GPUs and hardware dwarfs what Nvidia or any other company might make on consumer GPUs for gaming right now. Until that changes, we're not likely to see sufficient supply to meet demand for gaming GPUs, never mind finding GPUs available at their ostensible MSRPs.

One thing Nvidia didn't talk about yet is the widely rumored impending announcement of the RTX 5060 Ti and RTX 5060. Prices have started popping up online, and the full specs have been leaked multiple times. The TLDR: They're both 128-bit memory interfaces, and it looks like they'll launch with 8GB configurations, with an optional 16GB 5060 Ti as an upgraded solution. Prices and availability are still unknown factors.

A design flaw delayed Nvidia's rollout of Blackwell GPUs for data centers last year, prompting the company to redesign the silicon and packaging. However, this did not affect Nvidia's work on mid-cycle refresh Blackwell 300-series (Blackwell Ultra) GPUs for AI and HPC and next-generation Vera Rubin GPUs. Nvidia can already share some details about Rubin GPUs and its post-Rubin products.

"Blackwell Ultra is [due in the] second half," Jensen Huang, chief executive of Nvidia, reaffirmed analysts and investors at the company's earnings conference call. "The next train [is] Blackwell Ultra with new networking, new [12-Hi HBM3E] memory, and, of course, new processors. […] We have already revealed and have been working very closely with all of our partners on the click after that. The click after that is called Vera Rubin and all of our partners are getting up to speed on the transition to that. […] [With Rubin GPUs] we are going to provide a big, big, huge step up."

Later this year, Nvidia plans to release its Blackwell B300-series solutions for AI and HPC (previously known as Blackwell Ultra) that will offer higher compute performance as well as eight stacks of 12-Hi HBM4E memory, thus providing up to 288GB of memory onboard. Unofficial information indicates that performance uplift enabled by Nvidia's B300-series will be around 50% compared to the comparable B200-series products, though the company has yet to confirm this.

To further improve performance, B300 will be offered with Nvidia's Mellanox Spectrum Ultra X800 Ethernet switch, which has a radix of 512 and can support up to 512 ports. Nvidia is also expected to provide additional system design freedom to its partner with its B300-series data center GPUs.

Nvidia's next-generation GPU series will be based on the company's all-new, codenamed Rubin architecture, further improving AI compute capabilities as industry leaders march towards achieving artificial general intelligence (AGI). In 2026, the first iteration of Rubin GPUs for data centers will come with eight stacks of HBM4E memory (up to 288GB). The Rubin platform will also include a Vera CPU, NVLink 6 switches at 3600 GB/s, CX9 network cards supporting 1,600 Gb/s, and X1600 switches.

Jensen Huang plans to talk about Rubin at the upcoming GPU Technology Conference (GTC) in March, though it remains to be seen what he plans to discuss. Surprisingly, Nvidia also intends to talk about post-Rubin products at the GTC. From what Jensen Huang announced this week, it is unclear whether the company plans to reveal details of Rubin Ultra GPUs or its GPU architecture that will come after the Rubin family.

Speaking of Rubin Ultra, this could indeed be quite a breakthrough product. It is projected to come with 12 stacks of HBM4E in 2027 once Nvidia learns how to efficiently use 5.5-reticle-size CoWoS interposers and 100mm × 100mm substrates made by TSMC.

"Come to GTC and I will [tell you about] Blackwell Ultra," said Huang. "There are Rubin, and then [we will] show you what is one click after that. Really, really exciting new product."