Nvidia's new CPU rack is sized at exactly one agent sandbox per core
Vera has 88 Olympus cores, and the 256-socket rack advertises 22.5K concurrent environments, which is the core count almost to the digit.
Ian Buck has been hand-delivering processors to customers, which is not how server CPUs normally arrive.
Nvidia's vice-president for hyperscale carried the first Vera systems to Oracle Cloud Infrastructure and to three AI labs (Anthropic, OpenAI and SpaceXAI), and then to AWS this week. The company published the photographs on 27 August.
Skip the photographs. The rack arithmetic is the news.
Vera has 88 custom Olympus cores, with 1.2TB/s of LPDDR5X bandwidth and up to 1.5TB of memory per socket. Nvidia's Vera CPU Rack takes up to 256 of them, and is advertised as running "over 22.5K concurrent environments."
Multiply the core count by the socket count and you get 22,528, which is probably not a coincidence.
One sandbox per core, rounded down for the marketing copy. So Nvidia has decided that the unit of agentic compute is a core, not a container and not a virtual machine, and has priced a rack accordingly.
The comparison Nvidia does not make
Grace is the fair one, roughly speaking. It has 72 Arm Neoverse V2 cores and up to 500GB/s of LPDDR5X, so Vera brings about a fifth more cores and more than twice the bandwidth.
Which works out at roughly double the memory bandwidth per core.
Why would a CPU need that? Because the work an agent generates is not floating point. Agents compile generated code, run Python toolchains, hold long-context state, shuffle cache around, and start and kill thousands of short-lived sandboxes. Small, branchy, latency-bound work, which a core-count-optimised server chip likely handles badly.
Buck's framing is that agentic AI "is creating a new CPU moment in the AI factory." On the evidence of what the rack is sold to do, he means it fairly literally.
Be careful with the headline speed-up
Nvidia's blog says Vera "delivers up to 1.8x faster per-core performance on agentic AI workloads." Its own product page says the same figure applies to three named workloads "over leading x86 CPUs," with a footnote that the number is "relative performance based on measured data, and subject to change."
So it is a claim against Intel and AMD, not against Grace, and it is a ceiling rather than an average. The Grace comparison is left for the reader to work out, which is what we did above.
The counter-argument, and Oracle's answer to it
Hyperscalers have been designing their own Arm server CPUs for years and are not short of them. So why would any of them buy this one?
OCI plans to deploy hundreds of thousands of NVIDIA Vera CPUs beginning in 2026 because agentic AI demands sustained performance at massive scale. Vera's architecture is purpose-built for high-throughput reasoning workloads, delivering the efficiency, density and footprint OCI needs to power the next generation of enterprise AI.
Nobody buys hundreds of thousands of anything for a host-processor role beside a GPU. If you are buying at that scale you are not buying an accessory. That is a standalone fleet, ordered as one.
Where the volume actually comes from
Our read is that it arrives through the rack rather than through the socket next to a Rubin GPU, and that reinforcement-learning environments are the workload driving it. Anthropic's James Bradbury calls Vera "a promising part of the ecosystem when solving for agentic workloads," which is the sound of a compute lead who has already modelled it.
We would also expect Nvidia never to publish a CPU revenue line, for the same reason it stopped publishing a networking one.
But hold on to 22,528 until GTC. If the next Vera rack advertises a sandbox count that is no longer a multiple of the core count, something in the software layer has changed.
