Supermicro launches liquid-cooled CPU rack for agentic AI

The shift to agentic AI

Artificial intelligence is moving beyond linear inference chains, where a model simply generates a single response to a prompt. The next phase involves agentic AI and reinforcement learning (RL), which utilize branched and often looped execution paths. These workflows require many intermediate steps that rely heavily on CPU performance rather than just GPU power.

While GPUs remain the primary engines for generating tokens, code, and queries, CPUs are required to manage the logic of the actions. Agents need environments to execute actions, evaluate the results of those actions, and orchestrate complex workflows. This creates a new demand for high-density CPU infrastructure that can handle thousands of parallel execution environments simultaneously.

Supermicro is addressing this shift by introducing the NVIDIA Vera CPU rack. This rack-scale solution is designed to provide the dedicated compute capacity needed for these specific workloads. It allows companies to build full AI factories capable of hosting millions of parallel execution environments for agents at scale.

Hardware and cooling specs

The NVIDIA Vera CPU rack is built with 256 NVIDIA Vera CPUs, which deliver a total of 22,528 cores and 45,056 threads. To manage the heat generated by such high-density compute, the system utilizes a 100% liquid-cooled architecture. Specifically, it uses the Supermicro DLC-2 direct liquid cooling stack to maintain performance during intensive workloads.

In addition to the CPUs, the rack integrates NVIDIA BlueField-4 DPUs (Data Processing Units). These units provide high-performance networking and in-silicon security, which are essential for managing large-scale AI deployments. The hardware is designed to be compatible with the third-generation NVIDIA MGX modular reference architecture, which helps builders create future-compatible systems faster.

Powering these racks offers flexibility for data center operators. Customers can choose between a traditional AC feed or an 800 VDC sidecar power rack. The latter allows for the use of battery backup units (BBU) instead of a conventional UPS, which can improve power conversion efficiency and free up valuable floor space in the data center.

Security and orchestration

Security is a primary focus for the new rack, as it natively supports the NVIDIA OpenShell Reference Architecture. This software allows for the continuous monitoring and governance of agent behavior by applying least-privilege, policy-based controls at runtime. It ensures that agents operate within defined boundaries without consuming significant host CPU resources.

Complementing this is NVIDIA Sentry, which runs on the BlueField-4 DPUs as an out-of-band monitor. This setup provides attested telemetry and can quarantine an agent in milliseconds if it moves outside of its permitted boundaries. By using the DPUs for monitoring, the system maintains high security without taking away processing power from the main CPUs.

These integrated software and hardware solutions help enterprises retain observability and auditability in production systems. As companies deploy more autonomous agents, having a hardware-based security layer becomes critical. This architecture provides a secure way to scale the execution of agentic workloads across large AI factories.

Deployment and integration

The Vera CPU rack is designed to be versatile, functioning as either a standalone platform or a companion to the NVIDIA Vera Rubin NVL72. It can be used to accelerate agent orchestration, database operations, cluster management, and large-scale software build pipelines. This makes it useful for a variety of tasks beyond just running AI agents.

For customers already building on Supermicro's blueprints for the Vera Rubin NVL72, an optional design is available. This design pairs the two rack types with a defined network topology built on NVIDIA Spectrum-X Ethernet. It also includes hot-aisle containment layouts and redundant in-row coolant distribution units from the DLC-2 stack.

Supermicro provides a unified team to support the entire lifecycle of these deployments. This includes site surveys, design, integration, testing, and on-site deployment. Whether a customer uses the rack on its own or as part of a larger cluster, the goal is to accelerate the time to deployment for secure, scalable AI infrastructure.