2 weeks Ago
NVIDIA CEO Jensen Huang with a DGX Spark in hand during the GTC 2025 keynote I have had the opportunity to attend every GTC since Nvidia launched the annual conference in 2009. This year’s show built on GTC 2024’s absurd scale and hype with a full year of Nvidia Blackwell product releases and customer wins. CEO Jensen Huang once again delivered his keynote address to a packed SAP Center in San Jose, California and continued many of the themes we’ve seen from the company in the last year, holding steady rather than introducing many new concepts.
If anything, this event was a clear demonstration of Nvidia’s confidence in its own vision and how it plans to bring the world along with it. (Note: Nvidia is an advisory client of my firm, Moor Insights & Strategy.) Part of Nvidia’s role as a leading provider of chips for AI infrastructure is to become more predictable for its own customers.
This means not only delivering products that perform up to expectations, but also delivering them on a dependable timeframe with good reliability. That’s why it came as no surprise when, just like clockwork, Nvidia announced the Blackwell Ultra chip, also known as the B300. Nvidia also announced Dynamo , the company’s replacement for its Triton open-source libraries that helps accelerate inference and improve scalability while lowering deployment costs.
Looking beyond Blackwell, Nvidia also gave more details about the Rubin GPU and Vera CPU that are expected in 2026 and 2027. This was part of a broader disclosure of its compute, NVLink, networking and system roadmap. The base-model Rubin is slated for early 2026 release, complete with HBM4 high-bandwidth memory, while the Rubin Ultra is set to follow later in 2026.
Nvidia’s CPUs will also get an upgrade to the Vera architecture, which will enable the Vera Rubin series of servers based on those CPUs and GPUs. Nvidia also plans to upgrade its NVLink to sixth- and seventh-generation NVSwitches with 3600 GB/s of bandwidth paired with Spectrum6 and CX9 networking chips. The Spectrum6 is also expected to be Nvidia’s first silicon photonics product.
Moving into 2028, Nvidia plans to introduce the Feynman GPU, which will pair again with Vera CPUs but leverage “next-gen HBM” — without further specification since there hasn’t been much public disclosure of HBM standards beyond HBM4. Feynman will also pair with Nvidia’s eighth-generation NVSwitch and Spectrum7 (including its second generation of silicon photonics) for scale-out networking. Nvidia has been teasing DGX Spark, formerly known as Project Digits, since the CES show at the start of 2025, but it has finally gotten to the point of launching this product, which it calls a “desktop supercomputer.
” It is not only Nvidia’s smallest-scale Blackwell product, but it is also its most accessible platform for developing AI. The GB10 Superchip that powers it pairs a Blackwell GPU with a 20-core Arm-based CPU from MediaTek to create a very small-scale system that basically fits in the palm of your hand but is scalable and powerful. I believe that this is Nvidia’s way of becoming more accessible for inference and enabling developers to tap into the Nvidia ecosystem at any scale while still offering scalability, given that multiple DGX Spark machines can be connected to improve performance.
I also believe that it can and will be deployed as a sort of AI appliance for running Nvidia AI workloads at the edge when a user needs compute, but not something as big as a GB300. Nvidia was taking pre-orders during GTC, but it was also clear that ASUS, Dell and HP will be coming to market with their own GB10-based solutions, which I believe will enable much better reach for Nvidia. Speaking of the GB300, both Dell and HP also showed off workstations based on the DGX Workstation that Nvidia announced at GTC.
DGX Workstations are basically GB300 nodes, with a single Grace CPU and dual Blackwell GPUs running on a single liquid-cooled desktop. This puts maximum power in the hands of the data scientist or AI developer without requiring constant access to the cloud or an investment in a complete rack solution. I happened to witness Huang and Michael Dell sign the Dell GB300 DGX Workstation prototype machine on the GTC 2025 expo floor.
Michael Dell pictured with Jensen Huang in front of the Dell GB300 DGX Workstation One of Nvidia’s other big announcements from the show was the last of the Blackwell line, the RTX Pro series of GPUs. The RTX Pro series helps to clarify Nvidia’s product categories, with the Pro series specifically representing professional graphics and aimed at empowering the laptop, desktop and server categories. On top of its dominant position in AI, Nvidia remains the world leader in professional graphics and is the provider of the graphics that power The Sphere in Las Vegas, among many other venues and specialized graphics implementations.
The RTX Pro series is based on the GB200 series of chips and comes in five performance levels, with the top-end RTX Pro 6000 featuring a 600-watt, 300 Max-Q configuration. I got to see Dell’s Pro workstation running the RTX Pro 6000 with a 600-watt card on the show floor, which indicates to me that Nvidia is keeping Dell very close on AI — and for more than just datacenter use cases. HP and Lenovo are also partners with Nvidia for the RTX Pro series, and HP announced Nvidia-based systems the same week as GTC at its own Amplify conference in Nashville.
A table with the three different versions of the RTX Pro 6000 The RTX Pro 6000 also comes with the full 20,064 CUDA cores that the Blackwell architecture supports (the RTX 5090 ships with a mere 20,000 cores) and comes with a whopping 96GB of GDDR7 with ECC error correction. It also offers 4,000 TOPS of AI performance (FP4) and 380 TFLOPS of ray tracing performance. Additionally, because this is a PCIe Gen5 GPU, Nvidia has enabled VR-SLI for high-end XR solutions, enabling dual-GPU configurations that drive one GPU per eye for maximum photorealistic experiences.
The RTX Pro series will extend into laptops but will start with the RTX Pro 5000 and work all the way down to the RTX Pro 500. The RTX Pro 5000 for laptops will cap out at 24GB of VRAM, while the desktop and server versions of the RTX Pro 6000 will reach up to 96GB. One interesting detail is that these laptops will be available this summer from Dell, HP and Lenovo — as one would expect — but also from Razer, which sells gaming laptops and may be looking to move into the workstation market.
I believe that Nvidia made a mistake by announcing these RTX Pro GPUs at GTC 2025, because there is already so much going on at the show and it felt like this announcement didn’t get the attention it deserved. In fact, since these GPUs won’t be available until the summer anyway, I think it would have made much more sense to wait until August to make the announcement at the Siggraph 2025 conference. This is where Nvidia has traditionally announced its professional graphics products.
As it stands, there are now so many details that the industry wants to know about, but I guess Siggraph will be the place to ask those questions. During the GTC 2025 Keynote, Huang rapid-fired his appreciation to many of the company’s partners in making Blackwell a success. One key partner cited was Cadence, of which Huang said last year: “Blackwell would not have been possible without Palladium.
” Palladium is Cadence’s high-powered emulation platform. (For more, see this analysis of Cadence that I wrote in mid-2024 .) Cadence slide during Nvidia’s GTC 2025 keynote Nvidia is not only a customer of Cadence, but also a partner in accelerating Cadence’s own design products, including the Voltus, Celsius, Clarity, Fidelity and Reality tools.
Cadence has said that its Fidelity CFD (computational fluid dynamics) tool can perform multi-billion cell simulations in under 24 hours using B200 GPUs, versus a Top 500 Supercomputer cluster with hundreds of thousands of CPU cores requiring several days. Fidelity LES was also optimized for Blackwell — specifically for digital twins — and the comparative performance numbers running Nvidia are similarly stark. Using the highly complex mathematics of large eddy simulations, Fidelity LES can perform a sophisticated automotive simulation in 11 minutes using eight B200s, as compared to 16 hours on 1,000 CPUs, or a turbofan simulation in 50 minutes on eight B200s, versus 69 hours on 1,000 CPUs.
Blackwell also accelerated 3-D-IC multiphysics design, seeing improvements on a single Blackwell ranging from 7.7x to 8.2x compared to 32 CPUs.
Nvidia and Cadence are also expanding their partnerships on digital twinning of datacenters with the Cadence Reality platform. Part of that means integrating Cadence Allegro and Cadence Reality into Nvidia Omniverse; it also includes Cadence embracing OpenUSD and joining the AOUSD promoter group. This will bring SoCs, PCBs, blades, racks, infrastructure and environments all into OpenUSD, which should make it easier than ever to build a digital twin of a datacenter for anyone.
Nvidia is also releasing an Omniverse Blueprint for datacenter digital twins to make it easier for anyone to use when building their own Blackwell-based datacenter. If GTC 2025 reinforced any message, it’s that Nvidia is far from done. For years now, the company has been building many partnerships and growing its influence in the enterprise with its many AI tools — and there’s no end in sight.
Nvidia understands its role within the industry, especially now that we are seeing the company become more prominent in 5G networks with partners like Ericsson and carriers like T-Mobile and Verizon. While I do believe that AI RAN will become more prominent, I wouldn’t write Nvidia off, especially when leaders like Cisco are now joining forces with Nvidia for 6G. Nvidia’s GPU products are industry-leading across the board, and GTC is clearly a key venue for the company to show off its importance — but also to connect some of its closest partners to potential new customers.
I was really impressed with what I saw from the RTX Pro series, and even though I think the timing was off, these are some incredibly powerful GPUs that are driving a lot of excitement within the industry, especially for XR applications. I believe that AMD will have to do a lot more work to compete with Nvidia’s latest products, and that GTC was a tour de force even if there weren’t many surprises. Moor Insights & Strategy provides or has provided paid services to technology companies, like all tech industry research and analyst firms.
These services include research, analysis, advising, consulting, benchmarking, acquisition matchmaking and video and speaking sponsorships. Of the companies mentioned in this article, Moor Insights & Strategy currently has (or has had) a paid business relationship with Arm, Cadence, Cisco, Dell, Ericsson, HP, Lenovo, MediaTek, Nvidia, T-Mobile and Verizon..
