Nvidia contributes Vera Rubin rack and tray designs to OCP under MGX 1.2 spec

At the OCP Global Summit in San Jose on 12 October 2026, Nvidia contributed its Vera Rubin rack and compute tray designs to the Open Compute Project, releasing them through an updated MGX 1.2 specification^[1] that supports rack deployments of up to 250 kW^[1] - roughly double the density of the previous generation of AI-optimised racks.

What MGX 1.2 changes

The specification centres on eliminating the thermal and mechanical complexity that has made dense AI racks difficult to deploy at scale. Key additions include:

• Rack-scale blind-mate liquid cooling, which connects cooling interfaces automatically when a tray is pushed into position, removing the need to connect individual hoses by hand^[1]
• A liquid-cooled 5,000 A busbar designed for higher-power AI systems^[1]
• Fanless, cable-free compute trays that replace traditional connections with a central printed circuit board midplane
• Native 19-inch equipment support and configurable I/O bays
• Standardised power and cooling interfaces for interoperability across ecosystem partners

Nvidia said the same MGX rack footprint already supports the GB300 NVL72 and will carry forward to Vera Rubin NVL72 and Vera Rubin NVL72 CPX variants. A separate specification covers the Vera CPU rack, which Nvidia said has begun shipping for AI agent workloads.

This is not the first time Nvidia has opened a generation of rack designs to OCP; it did the same with Blackwell. The pattern is consistent: standardise the infrastructure layer around the chip, giving manufacturers a common target while keeping the compute hardware proprietary.

Power and grid standards alongside the rack

The rack contribution arrived alongside a broader set of facility and grid specifications. Nvidia joined AMD, Google, and Meta to publish Open Data Center Specification v0.7, a shared blueprint covering facility power architecture, cooling, and physical layout^[1]. Google, Meta, Microsoft, and Nvidia separately published an OCP Grid Resilience Specification defining how large AI sites should behave during utility disturbances - staying connected, responding automatically, and restoring to near-full load within two seconds.

On the power delivery side, Visionbay - an AI infrastructure company under Foxconn Technology - unveiled Project K-1, an AI factory platform built around Nvidia's 800 V DC reference architecture and Vertiv power and thermal equipment. The design moves AC-to-DC conversion into cluster-level power centres and distributes 800 V DC closer to compute, reducing copper dependency and conversion losses. More than 80 partners are now developing 800 V DC-compatible infrastructure through OCP.

The ecosystem moves

Meta committed to adopting Nvidia's Spectrum-6 102.4 Tb/s Ethernet switches for its Vera Rubin AI factories, with Nvidia's Spectrum-X Multiplane technology sized for clusters of more than 100,000 GPUs. Nvidia and Cisco also outlined the Ultra-Dense Small Form Factor Pluggable (USFP) standard, targeting 400 Gb/s per lane - though no sampling schedule or availability date was given.

What to watch

The Open Data Center Specification is at version 0.7, and Nvidia has not disclosed which system builders will adopt MGX 1.2 or on what timeline. The grid resilience specification's two-second recovery target is a meaningful commitment for utilities and grid operators managing large-load interconnection - watch for how transmission planners respond to facilities that claim near-instant autonomous grid response as a design standard rather than an aspiration.