Supermicro’s Vera CPU rack packs 256 chips for agent sandboxes

Supermicro announced a fully liquid-cooled rack of 256 Nvidia Vera CPUs on October 23, which the company says can host up to 22,500 concurrent AI agent sandboxes. It is open for quote requests; no price or shipping date was given.

What is inside the rack

According to Supermicro, the rack integrates 256 Vera CPUs for a total of 22,528 cores and 45,056 threads. Nvidia BlueField-4 DPUs handle networking and what the company calls silicon-level security. Cooling comes from Supermicro’s DLC-2 direct liquid cooling stack, and the rack follows the third-generation Nvidia MGX design, the same infrastructure approach used for Vera Rubin NVL72.

Customers can power it from conventional AC feeds or from an 800 VDC sidecar power rack. The sidecar option lets battery backup units replace a traditional UPS, which Supermicro says frees floor space, can improve power conversion efficiency and lowers both capital and operating costs. Those savings are the company’s claim and come without figures.

The rack can stand alone or sit next to Vera Rubin NVL72 racks. For the combined setup, Supermicro offers an optional design with a defined Spectrum-X Ethernet topology, hot aisle containment and redundant coolant distribution units. A single Supermicro team covers site survey, design, L11 and L12 integration and testing, on-site deployment and support. Air-cooled 1U and 2U Vera systems are also part of the portfolio, CEO Charles Liang said.

The agent security layer is built in

Supermicro says the rack supports Nvidia’s OpenShell Reference Architecture. OpenShell runs on the Vera CPUs and applies least-privilege and policy controls to agents at runtime. Nvidia Sentry, running on the BlueField-4 DPUs, acts as an out-of-band monitor that enforces policy, supplies attested telemetry and can quarantine an agent within milliseconds, according to the release. Because it sits on the DPU, the company says it uses no host CPU capacity.

“NVIDIA Vera CPUs and BlueField-4 DPUs, integrated into Supermicro’s liquid-cooled DCBBS architecture, give customers the dedicated compute and hardware-based security needed to scale agentic workload execution within their AI factories.”

That is Chris Marriott, Nvidia’s vice president of enterprise platforms. For the design itself, see our report on Nvidia’s OpenShell blueprint.

A crowded race for agent CPU racks

Supermicro’s pitch is that GPUs keep generating tokens while CPUs run the actions, tools and checks around them, so large clusters need dedicated CPU capacity. The same logic drove AMD’s Helion, which our earlier coverage described as an open rack with up to 32,768 CPU cores for agents. On raw core count, AMD’s blueprint is about 45 percent larger than Supermicro’s 22,528 cores. Supermicro appears there too: it plans to sell Helion through its own building blocks, so it is hedging across both chip vendors.

The sandbox figure is the number to watch. In that same coverage, CoreWeave reported more than 11,000 concurrent agent environments per rack on Nvidia Vera, compared against an unnamed x86 chip. Supermicro’s 22,500 is roughly double that, but the release does not say what each sandbox contains, how much memory or CPU it gets, or what workload was assumed. The two figures may not be comparable.

What the release leaves unanswered

The announcement gives no price, no delivery date and no power draw per rack, and it offers no benchmark against x86 or Arm alternatives. Availability is limited to requesting a quote, so only large buyers building AI factories are realistically in scope. The 22,500 sandbox figure, the millisecond quarantine claim and the efficiency gains from the battery option all come from the company and have not been independently tested.

For readers outside data centers, the practical effect is indirect. If agent workloads keep growing, as Supermicro and Nvidia argue, the CPU share of AI spending rises, and rack vendors want to sell that capacity in the same liquid-cooled format as GPUs. Whether customers buy it as a standalone product or only as an add-on to Vera Rubin NVL72 will show how real that demand is.