Supermicro has added an NVIDIA Vera CPU rack to its Data Center Building Block Solutions (DCBBS) portfolio, pitching it as a rack-scale pool of CPU compute for agentic AI, reinforcement learning, and data processing. The rack is designed to deploy standalone or alongside Supermicro’s NVIDIA Vera Rubin NVL72 systems, using a common rack and liquid-cooling infrastructure.
The NVIDIA Vera CPU rack integrates 256 NVIDIA Vera CPUs and is specified at 22,528 CPU cores and 45,056 threads. Supermicro also pegs it at support for up to 22,500 concurrent agent sandboxes per rack. Networking and hardware security are provided via NVIDIA BlueField-4 DPUs, and the rack is built as a 100% liquid-cooled system using Supermicro’s DLC-2 direct liquid-cooling stack.
For data center operators, the practical point is density: if agent workflows are CPU-bound (sandboxing, tool execution, orchestration, and monitoring), a dedicated CPU rack can keep those tasks from contending with GPU nodes and can simplify capacity planning. But it also hard-commits you to liquid cooling at the rack level, which changes commissioning, maintenance procedures, and spare strategy versus air-cooled 1U/2U CPU fleets.
On the mechanical and electrical side, the Vera CPU rack uses the third-generation NVIDIA MGX rack design as part of the DLC-2 stack, aiming for standardized deployment alongside Vera Rubin NVL72. Power options include either a “traditional AC feed to compute rack” or an 800 VDC sidecar power rack. Supermicro says the 800 VDC sidecar approach enables battery backup units (BBU) in place of a conventional UPS, with the stated goals of freeing floor space, improving end-to-end power conversion efficiency, and lowering CapEx and OpEx.
For agent safety and governance, Supermicro says the rack supports NVIDIA OpenShell Reference Architecture. In that setup, NVIDIA OpenShell runs on NVIDIA Vera CPUs to apply runtime least-privilege, policy-based controls, while NVIDIA Sentry runs on BlueField-4 DPUs as an out-of-band monitor that enforces policy, provides attested telemetry, and can quarantine an agent “in milliseconds,” according to NVIDIA.
“As enterprises scale agentic AI and reinforcement learning, supporting thousands of concurrent agent workflows requires a new class of high-density CPU infrastructure,” said Chris Marriott, vice president of enterprise platforms at NVIDIA.
Commercially, Supermicro says the NVIDIA Vera CPU rack is available to quote now.
Source: Supermicro
