NVIDIA contributes Vera Rubin rack designs to OCP and co-defines 400Gb/s optics with Cisco

NVIDIA is standardizing the physical layer of 'AI factories'. At the Open Compute Project Global Summit, it contributed rack-level and compute-tray designs for Vera Rubin through its MGX modular architecture. Hyperscalers and OEMs can now build compatible systems without bespoke engineering. NVIDIA and Cisco also announced USFP, a new pluggable optics form factor running 400Gb/s per lane for scale-out AI networking.

Opening rack designs lowers integration cost for data-center builders and pulls the ecosystem toward NVIDIA's reference designs at the moment rivals are pushing their own silicon. On the desktop, a 64GB DGX Spark configuration ships October 23 at $4,999 for developers running large models locally.

The timing is pointed. Microsoft revealed Maia 300 the same day with efficiency claims aimed squarely at GPUs, and Google and AWS keep expanding TPUs and Trainium. NVIDIA's answer is to make its systems the default standard: if racks, trays and optics follow NVIDIA reference designs, swapping in custom accelerators gets harder. Reports also say NVIDIA is in talks to deepen its $800M stake in open-weight coding lab Reflection AI.

Watch how fast OCP members adopt the MGX Vera Rubin designs, and whether USFP becomes an industry optics standard or stays a Cisco-NVIDIA format.