NVIDIA has contributed the rack and compute tray designs for its Vera Rubin platform to the Open Compute Project, published at the OCP Global Summit on 12 October as an update to its open MGX specification.
MGX 1.2 adds rack-scale blind-mate liquid cooling, a liquid-cooled 5,000-amp busbar, native support for 19-inch equipment, configurable I/O bays and fanless liquid-cooled compute trays, and covers rack deployments up to 250 kilowatts. A separate Vera CPU rack design extends MGX to agentic workloads such as tool calling, orchestration and sandboxed execution, sitting alongside the Vera Rubin NVL72 configuration aimed at training and inference.
Two things arrived with it. Meta said it will adopt NVIDIA’s Spectrum-6 Ethernet switches, rated at 102.4 Tb/s, for larger Vera Rubin build-outs, running them in its own liquid-cooled IcePack rack under its FBOSS network operating system. And NVIDIA co-published an OCP grid resilience specification with Google, Meta and Microsoft, covering how AI facilities should ride through grid disturbances and restore near-full load in under two seconds.
Opening the mechanical and power designs is less generosity than a bid to make NVIDIA’s rack the default shape of a data centre hall. The grid resilience work may matter more: utilities increasingly treat large AI sites as a stability problem, and a shared specification gives the industry something to point at.