NVIDIA contributes Vera Rubin rack designs to open AI infrastructure standards

NVIDIA is giving infrastructure builders more of the designs and software surrounding its Vera Rubin systems. At the Open Compute Project Global Summit on October 12, 2026, it contributed rack and compute-tray designs alongside a coordinated package of open facility standards and operating tools, aimed at making AI data centers easier to build and run.

A shared rack, with different computing roles

The hardware contribution runs through NVIDIA’s MGX architecture, a modular foundation for building computing systems. NVIDIA joined OCP’s Open DC Initiative in contributing to the third-generation MGX data-center reference design. The updated MGX 1.2 specification brings liquid cooling, rack interconnections and standardized power interfaces into that common framework.

Builders also get support for 19-inch equipment, configurable input/output bays and fanless, liquid-cooled compute trays. NVIDIA says the standardized design improves assembly, testing and servicing, while allowing partners to build interoperable systems. Those are the company’s stated benefits of the architecture, rather than quantified deployment results.

The same rack foundation serves two different jobs. Vera Rubin NVL72 handles large-scale model training and inference—the work of generating model responses. The Vera CPU rack is designed for the surrounding agent tasks: calling tools, coordinating work and executing code in isolated environments. Both use the third-generation MGX approach to combine computing, networking, power and cooling.

Up to 250 kWSupported rack deployment

NVIDIA specifies support for rack deployments up to 250 kilowatts.

5,000 ALiquid-cooled power conductor

The specification includes a liquid-cooled 5,000-amp busbar, the conductor that distributes power.

The building and grid get specifications too

Where MGX defines the rack, Open Data Center Specification v0.7 addresses the facility around it. Published by NVIDIA, AMD, Google and Meta, it covers power architecture, cooling infrastructure and physical layout. The shared blueprint gives equipment manufacturers and data-center operators a common design target instead of separate proprietary requirements from individual customers.

A second document specifies behavior rather than layout. Google, Meta, Microsoft and NVIDIA published the OCP Grid Resilience Specification, which calls for large AI facilities to stay connected during utility-grid disturbances, respond automatically and return to near-full load in less than two seconds. That timing is a requirement in the specification, not a reported measurement from an operating facility.

NVIDIA says both specifications are compatible with its DSX reference design, connecting facility layout and power architecture to software that monitors grid signals. The summit package also reaches into network connections: NVIDIA and Cisco proposed USFP, a removable optical-module format designed for 400-gigabit-per-second lanes. It remains a planned format, with 16 two-way lanes initially and plans to scale to 32.

Open software takes on the operating work

Physical compatibility is only one part of the package. NVIDIA also added four open-source components to DSX OS, the operating layer of its AI-factory platform. They address the work of configuring equipment, checking that clusters are ready and turning newly installed hardware into systems customers can use.

• Config Manager automates switch configuration so operators can deploy, validate and maintain consistent network settings across an AI facility.
• Cluster Readiness Engine stress-tests GPU clusters with simulated workloads before training starts, aiming to catch hardware problems before they interrupt jobs.
• Rack Management Service treats racks as operating units, providing a common hardware-management layer and tools for managing firmware at facility scale.
• Cluster API Provider for NVIDIA Infrastructure Controller handles bare-metal provisioning—preparing physical servers for use—and customer onboarding. NVIDIA says it can make clusters production-ready in hours rather than days.

NVIDIA says these components grew out of its experience operating DGX Cloud. Cloud providers and software partners can integrate them into platforms serving multiple customers. That gives the release a different role from the physical standards: the specifications define what builders should assemble, while the software supplies tools for preparing and managing it.