Supermicro announced the NVIDIA Vera CPU rack on October 22, a fully liquid-cooled, rack-scale system that packs 256 NVIDIA Vera CPUs into a single rack dedicated to running AI agents. The rack delivers 22,528 cores and 45,056 threads and supports up to 22,500 concurrent agent sandboxes per rack, positioning CPU capacity — not GPU capacity — as the bottleneck hardware for agentic AI at factory scale.
The system is the latest addition to Supermicro’s Data Center Building Block Solutions (DCBBS) portfolio and is available to quote now. It is designed to work standalone or alongside the now-shipping Supermicro NVIDIA Vera Rubin NVL72, sharing the same third-generation NVIDIA MGX rack infrastructure, battery backup units and DLC-2 liquid-cooling stack.
What Supermicro Announced
The Vera CPU rack exists because agentic workloads have changed shape. GPU clusters still generate the tokens, code and queries, but the work around them — executing actions, evaluating results, running tools and orchestrating workflows — runs on CPUs, in sandboxed environments that have to exist in parallel. Supermicro’s argument is that an AI factory with thousands of GPUs needs dedicated CPU compute to host thousands, in its design tens of thousands, of those execution environments at once.
“As agentic AI transforms data center requirements, Supermicro is extending its DCBBS portfolio with fully integrated infrastructure,” said Charles Liang, Supermicro’s president and CEO. Alongside the liquid-cooled rack, the company’s NVIDIA Vera CPU portfolio includes air-cooled 1U and 2U systems for customers who do not want rack-scale liquid cooling.
NVIDIA’s Chris Marriott, vice president of enterprise platforms, made the same point from the chip side: “As enterprises scale agentic AI and reinforcement learning, supporting thousands of concurrent agent workflows requires a new class of high-density CPU infrastructure.” The rack targets agentic AI, reinforcement learning and data processing workloads, and standalone deployments can also run agent orchestration, database operations, cluster management with Slurm or Kubernetes, data analytics and large-scale software build pipelines.
OpenShell And Sentry Keep Agents Inside Their Boundaries
The rack integrates NVIDIA Vera CPUs with BlueField-4 DPUs and natively supports the NVIDIA OpenShell Reference Architecture — the agent safety runtime NVIDIA launched at GTC Berlin this week. OpenShell runs on the Vera CPUs to apply least-privilege, policy-based controls to agents at runtime, while NVIDIA Sentry runs on the BlueField-4 DPUs as an out-of-band monitor that enforces policy, provides attested telemetry and can quarantine an agent that steps outside its permitted boundaries in milliseconds.
Supermicro says the pairing is meant to keep production agentic systems observable and auditable without consuming host CPU resources — the DPUs carry the monitoring load in silicon while the Vera CPUs carry the sandboxes. It is the same architecture NVIDIA is pushing across its partner systems, now in a rack whose entire purpose is the CPU side of agent execution.
One Infrastructure For Vera CPU And Vera Rubin Racks
The rack uses the third-generation NVIDIA MGX design inside Supermicro’s DLC-2 direct liquid-cooling stack, giving it a common architecture with the Vera Rubin NVL72. For power, customers can choose a traditional AC feed or an 800 VDC sidecar power rack, which allows battery backup units in place of a conventional UPS — a change Supermicro says frees floor space, improves end-to-end power conversion efficiency and lowers both capital and operating cost.
For customers building on Supermicro’s DCBBS Blueprints for NVIDIA Vera Rubin NVL72, an optional design pairs Vera Rubin NVL72 racks with Vera CPU racks in a defined topology built on NVIDIA Spectrum-X Ethernet, with hot-aisle containment and redundant in-row coolant distribution units and rear door heat exchangers. A single Supermicro team handles site survey, design, L11 and L12 integration and testing, on-site deployment and ongoing support, whether the Vera CPU rack ships with a Vera Rubin system or on its own — the same factory-assembly logic behind NVIDIA’s containerised DSX POD AI factories, applied at rack level.
Why It Matters
The Vera CPU rack is a sign of where AI infrastructure spending is moving. NVIDIA introduced Vera as a standalone CPU in March, and demand has followed the agents: Modal became a Vera CPU customer this month after tests showed its agent work doubling on the chip. Supermicro’s rack turns that demand into a unit of purchase — 256 CPUs, 22,500 sandboxes, one cooled and powered block — which lets an operator add agent-execution capacity in step with GPU capacity instead of discovering mid-deployment that thousands of agents are queuing for CPU time. With GTC Berlin’s announcements stacking up around Vera Rubin, the CPU rack fills the one slot the GPU racks left open.
Background
Supermicro, founded in 1993 and based in San Jose, built its position on modular Server Building Block Solutions and has ridden the AI data centre buildout as a first-wave supplier of GPU systems. NVIDIA’s Vera CPU, named in the Vera Rubin platform after astronomer Vera Rubin, is its first standalone data centre CPU line, aimed at the orchestration and sandbox work agentic AI generates around GPU inference. The Vera CPU rack, announced during GTC Berlin week, extends that platform from chips and GPU racks into the CPU-dense tier of the AI factory.
Source: Supermicro via PR Newswire
0 Comments