Supermicro unveils 256-CPU liquid-cooled Vera rack for agentic AI

A CPU rack for the agent side of AI factories

Super Micro Computer, Inc. (NASDAQ: SMCI) announced the NVIDIA Vera CPU rack on Oct. 22, 2026, as the latest addition to its Data Center Building Block Solutions (DCBBS) portfolio. The rack is meant to provide dense CPU capacity for agentic AI, reinforcement learning (RL) workloads and data processing.

According to Supermicro, the rack integrates 256 NVIDIA Vera CPUs, delivering 22,528 cores and 45,056 threads. NVIDIA BlueField-4 DPUs handle networking and in-silicon security, while Supermicro's DLC-2 direct liquid cooling stack makes the rack fully liquid-cooled.

The company's reasoning is that agentic AI and RL shift from linear inference chains to branched, and often looped, execution with many intermediate steps that depend on CPU availability and performance. GPUs remain the primary engine for generating tokens, code and queries, Supermicro says, while CPU-based sandboxes execute actions, evaluate results, run tools and orchestrate workflows. With hundreds or thousands of GPUs deployed, dedicated CPU compute is needed to host thousands or millions of parallel execution environments for agents.

Common infrastructure with Vera Rubin NVL72

The rack complements the Supermicro NVIDIA Vera Rubin NVL72, which Supermicro describes as now shipping. It can run standalone or alongside NVL72 in an AI factory. Both use the third-generation NVIDIA MGX design, part of the DLC-2 stack, which creates a common infrastructure architecture and supports standardized deployment.

For power, customers can choose a traditional AC feed to the compute rack or an 800 VDC sidecar power rack. The sidecar option allows battery backup units (BBU) to replace a conventional UPS for power loss protection. Supermicro says this approach helps free data center floor space, can improve end-to-end power conversion efficiency, and lowers CapEx and OpEx.

Customers building on the Supermicro DCBBS Blueprints for NVIDIA Vera Rubin NVL72 can choose an optional design that pairs NVL72 racks with Vera CPU racks. It includes a defined network topology built on NVIDIA Spectrum-X Ethernet, hot-aisle containment layouts, and redundant in-row coolant distribution units and rear door heat exchangers. One Supermicro team handles site survey, design, L11 and L12 integration and testing, on-site deployment and ongoing support.

Agent security built in

The rack natively supports the NVIDIA OpenShell Reference Architecture to monitor and govern agent behavior. According to the release, NVIDIA OpenShell runs on the Vera CPUs and applies least-privilege, policy-based controls to agents at runtime. 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 moves outside its permitted boundaries in milliseconds.

Supermicro says this lets companies keep observability and auditability in production agentic systems without consuming host CPU resources.

Standalone use and availability

Used on its own, the rack can serve as a general CPU platform. Supermicro lists agent orchestration and sandboxing, database operations, cluster management, schedulers such as Slurm and Kubernetes, data analytics and large-scale software build pipelines.

Charles Liang, president and CEO of Supermicro, noted that customers who prefer air cooling can use the company's NVIDIA Vera CPU portfolio of air-cooled 1U and 2U systems, in addition to rack-scale liquid cooling options. Chris Marriott, vice president of enterprise platforms at NVIDIA, said supporting thousands of concurrent agent workflows requires "a new class of high-density CPU infrastructure."

The Supermicro NVIDIA Vera CPU rack is available to quote now. The release gives no pricing or delivery dates.

For operators of AI factories, the announcement points to CPU capacity becoming a planned part of cluster design next to GPUs, with the option to buy it as a matched rack or as a separate platform.