Nvidia Opens Vera Rubin Rack Designs to Open Compute Project

Nvidia used the Open Compute Project’s Global Summit in San Jose to push its Vera Rubin generation of AI hardware toward becoming an industry standard rather than a proprietary design. The company contributed Vera Rubin rack and compute tray designs to OCP through its MGX architecture, meaning system builders outside Nvidia can now manufacture compatible racks rather than wait for a single-source part. The updated MGX 1.2 specification adds blind-mate liquid cooling, a 5,000A busbar and support for rack deployments up to 250kW, alongside native 19-inch equipment support for operators who still run older gear alongside new Vera Rubin systems.

Several partners filed in alongside the core contribution. Meta said it is adopting Nvidia’s Spectrum-6 Ethernet switches, which run at 102.4Tb/s, to build larger Vera Rubin AI factories on open networking, pairing them with its own FBOSS operating system. Nvidia and Cisco separately proposed USFP, a new pluggable optics form factor meant to replace the widely used OSFP standard with roughly four times the bandwidth density and about 1.3 times the power efficiency. Google and Nvidia announced that future ConnectX-10 SuperNICs will support Google’s Falcon transport protocol, letting the two companies’ networking hardware talk over a shared transport layer. Visionbay unveiled an 800 VDC AI factory design built around Nvidia’s reference architecture, aimed at handling higher rack power densities without a full facility rebuild.

Nvidia, Google, Meta, AMD and Microsoft also co-published two facility-level standards: an Open Data Center Specification covering power and cooling layout, and a Grid Resilience Specification describing how large AI sites should behave during utility disturbances, including restoring near-full load within seconds of a disruption.

The numbers

Meta’s adopted Spectrum-6 switches run at 102.4Tb/s, roughly twice the per-lane bandwidth of the prior Spectrum-4 generation. The proposed USFP optical module promises about 1.3 times the power efficiency of current pluggable optics while packing more lanes into a similar footprint. MGX 1.2 racks are rated to support up to 250kW of load, a sharp increase over earlier rack generations. Spectrum-X Multiplane technology is described as letting agentic AI factories scale beyond 100,000 GPUs using fewer network switch tiers.

Taken together, the announcements describe infrastructure decisions, rack dimensions, power busbars, optical connectors and grid behaviour, made roughly a generation ahead of the chips that will sit inside them, which is the stage at which standards either take hold across an industry or get locked behind one vendor’s part numbers.