Super Micro Computer, Inc. announced a 256-CPU NVIDIA Vera rack Oct. 22, presenting it as dedicated compute for agentic AI, reinforcement-learning and data-processing workloads.
Each rack is listed with 22,528 cores and 45,056 threads. Supermicro says its fully liquid-cooled design uses the company’s DLC-2 cooling stack and NVIDIA’s third-generation MGX design. Customers can deploy it on its own or pair it with Vera Rubin NVL72 racks, according to Supermicro.
Supermicro’s rationale is that AI agent workflows can branch or repeat. The company says GPU systems produce model outputs while CPU sandboxes handle tool calls, check results and manage the next steps. It argues that dedicated CPU resources can support those tasks alongside GPU clusters.
The announced configuration includes BlueField-4 data processing units (DPUs), NVIDIA’s Open Agent Safety Platform and its OpenShell Reference Architecture, Supermicro says. NVIDIA describes OpenShell as the software element and Sentry as a reference setup using BlueField-4 DPUs.
Supermicro says a rack can host up to 22,500 concurrent agent sandboxes. NVIDIA’s product page gives a different measure, listing capacity for more than 22,500 concurrent CPU environments. The descriptions use different labels and do not explain the gap.
Supermicro lists agent orchestration and sandboxing, database operations, cluster management, analytics and software build pipelines as potential uses. The company has named no customers or orders for the CPU rack.
Supermicro is taking quote requests, but its announcement gives no price or delivery schedule. The company separately said its Vera Rubin NVL72 racks began shipping Sept. 23; that update applies to those systems, not the newly announced CPU rack.
