Supermicro ships 100% liquid-cooled rack with 256 Nvidia Vera CPUs

Supermicro has launched a fully liquid-cooled rack-scale system carrying 256 Nvidia Vera processors, available to purchase now. The product joins the company's Data Center Building Block Solutions (DCBBS) portfolio, its pre-integrated rack-to-data-center product line.

The rack delivers what Supermicro calls "dense CPU capacity" for three workloads: agentic AI (autonomous software agents that plan and act with minimal human input), reinforcement learning, and general data processing. The company puts capacity at 22,500 sandboxes — isolated runtime environments for testing and running AI agents — at what it terms "liquid-cooled AI factory scale."

What does the rack actually contain?

The building blocks, as announced:

• 256 Nvidia Vera CPUs per rack, liquid-cooled end to end
• Nvidia BlueField4 DPUs (data processing units that offload networking and security from the host CPU) integrated with Supermicro's Open Agent Safety Platform
• OpenShell Reference Architecture for securing agentic AI orchestration — the coordination layer that routes work between agents

The security pairing matters for the target market. Agentic AI workloads multiply the number of autonomous processes running in a data center, and each one is a potential attack surface. Supermicro is positioning the BlueField4 integration as the answer at the infrastructure layer rather than in software alone.

How does it fit existing data centers?

The rack uses the third-generation Nvidia MGX design, the modular chassis standard shared across Nvidia's partner ecosystem. That shared infrastructure means common mechanics, power and cooling with the Vera Rubin NVL72 platform — Nvidia's next-generation rack configuration pairing 72 GPUs with Vera CPUs.

The practical consequence: operators can deploy the Vera CPU rack standalone or alongside Vera Rubin NVL72 systems in the same facility without bespoke integration work. That cuts deployment time, which is the metric hyperscalers and neocloud operators watch most closely right now.

Power gets two options. Customers can stick with traditional AC distribution, or spec an 800VDC sidecar power rack — a direct-current power shelf mounted beside the compute rack. The 800VDC route supports backup battery units (BBUs) in place of a conventional UPS (uninterruptible power supply), trimming conversion losses and floor space. At rack densities like these, DC distribution is becoming the default choice; 800 volts is the emerging Nvidia-recommended standard for AI factories.

What does this change for buyers?

Air-cooled shops aren't locked out. Charles Liang, president and CEO of Supermicro, said the Vera CPU portfolio also includes air-cooled 1U and 2U systems alongside the rack-scale liquid options.

"As agentic AI transforms data center requirements, Supermicro is extending its DCBBS portfolio with fully integrated infrastructure," Liang said. "Together with Nvidia, we help accelerate time to deployment with a common purpose-built architecture for secure, scalable agentic AI."

The pitch, stripped of marketing, is straightforward: one architecture across air-cooled nodes and liquid-cooled racks, one power scheme shared with the NVL72 GPU platform, and agent security baked into the silicon. Buyers evaluating CPU-heavy inference, sandbox-scale agent testing or RL training now have a turnkey rack option that ships immediately — no pricing was disclosed.

For an industry where GPU racks grab headlines, a 256-CPU rack is a deliberate bet: the next wave of AI spend is orchestration and simulation, not just model training. Supermicro is first to market with that shape. Expect the usual suspects to follow.