
Nvidia is no longer content with owning the GPU side of the data centre. At SC26 in Chicago today, the company unveiled the Vera CPU rack-scale platform — a single liquid-cooled cabinet holding 192 of its first fully custom data centre CPUs, aimed squarely at the scientific supercomputing work that has belonged to Intel and AMD’s x86 chips for decades.
This is not a GPU server with a CPU helper inside. Vera is the headline act, and Nvidia built the whole rack around keeping those processors fed.
192 CPUs, 300TB of Memory, One Cabinet
Each Vera CPU carries 88 custom Nvidia Olympus cores and 176 threads, with SOCAMM2 LPDDR5X memory delivering up to 1.2TB/s of bandwidth per chip. Stack 192 of them into one MGX rack and you get roughly 300TB of memory, 1 petabyte of storage and more than 200TB/s of aggregate peak memory bandwidth, all linked by Nvidia’s Quantum-X800 InfiniBand networking and fully liquid-cooled, with eight CPUs per 1U tray.
In Nvidia’s own tests, Vera delivered “1.9x the performance and 2.4x the performance per watt of x86 systems” on OpenFOAM’s motorBike Large airflow simulation — Nvidia’s figures, not independent benchmarks, so treat them as the opening claim in an argument Intel and AMD will happily contest. Its projected gains on the Weather Research and Forecasting model are 1.6x performance and 2x efficiency against x86.
“Scientific simulations need CPUs that can keep large datasets moving and sustain performance as applications scale,” said Ian Buck, Nvidia’s vice president of hyperscale and high-performance computing, in the announcement. “Vera brings custom CPU cores, high-bandwidth memory and InfiniBand networking together to address those bottlenecks.”
The platform already has a flagship customer: DKRZ, the German Climate Computing Center, has selected Vera CPUs for its HLRE-5 supercomputer, due in early 2028 alongside GB200 Grace Blackwell systems, for global kilometre-scale climate simulations. Nvidia’s HPC SDK 26.11, with compilers and libraries optimised for Vera, is available now, along with a public beta of a new Fortran compiler.
Why It Matters
Vera lands in a week when the custom-silicon pressure on Nvidia itself is the story — OpenAI and Anthropic’s huge Google TPU orders and a J.P. Morgan estimate putting Nvidia at just 39% of future frontier-lab compute deals — both underline how fast the AI hardware map is being redrawn. Vera is Nvidia redrawing a different part of that map: the CPU socket, on its own silicon, in the labs that buy supercomputers.
Vera CPU racks are expected from Dell Technologies, HPE, Lenovo and Supermicro beginning in early 2027. No pricing was announced — and with Nvidia’s earnings due Tuesday, Vera gives the company a brand-new CPU story to tell investors right when it needs one.
Source: NVIDIA via GlobeNewswire — https://investingnews.com/nvidia-expands-vera-cpu-platform-to-accelerate-scientific-discovery/
- Nvidia’s Rubin HPC Packs 50,000+ CUDA Cores Into One GPU — and It’s Built for AI Agents Doing Science - November 16, 2026
- OpenAI’s AI Agents Can Now Click Through Hospital Software — and Doctors Rated 99% of Their Work Safe - November 16, 2026
- Nvidia’s Vera Rack Packs 192 Custom CPUs Into One Cabinet — and It’s Coming for Intel and AMD’s Science Crown - November 16, 2026