NVIDIA announced that it contributed Vera Rubin rack and compute tray designs to the Open Compute Project (OCP) through its open MGX architecture. The contribution gives system manufacturers a common foundation for building artificial intelligence infrastructure, with standardized power, cooling, and connectivity interfaces.
The announcement, presented at the OCP Global Summit, covers the infrastructure that houses and connects the equipment. New developments include MGX 1.2, a specification that supports deployments of up to 250 kilowatts per rack and incorporates liquid cooling, fanless compute trays, and support for 19-inch equipment.
NVIDIA also detailed a rack design for its Vera CPUs, aimed at AI agent tasks such as coordinating processes, using tools, and running code in isolated environments. It shares MGX’s modular foundation with the Vera Rubin NVL72 infrastructure, designed for training and running large-scale models.
Software and standards for operating the infrastructure
The open-source effort includes new components for DSX AI Factory. Their functions include automating network configuration, testing groups of GPUs before training, managing rack hardware, and preparing physical servers for use. NVIDIA presents these tools as a way to simplify operations; the time and efficiency improvements described are company claims.
In parallel, NVIDIA, AMD, Google, and Meta published Open Data Center Specification v0.7, which defines a shared foundation for the physical layout, power supply, and cooling of AI data centers. Another specification, published by Google, Meta, Microsoft, and NVIDIA, establishes how this infrastructure should respond to power grid disturbances, including recovery from a near-full load in less than two seconds as a requirement, not as a demonstrated operational result.
