NVIDIA's OCP 2026 optics and rack contributions

At the OCP Global Summit, NVIDIA contributed Vera Rubin rack and compute tray designs to the Open Compute Project through an updated MGX 1.2 specification. The specification adds rack-scale blind-mate liquid cooling, a 5,000A liquid-cooled busbar, rack deployments up to 250kW, native 19-inch equipment support, and a new Vera CPU rack for agentic AI workloads such as tool calling and orchestration. The MGX architecture was first shared with OCP the prior year and has since been used across multiple server designs.

NVIDIA and Cisco announced the Ultra-Dense Small Form Factor Pluggable (USFP) standard, targeting 400Gb/s per lane in roughly the OSFP footprint. NVIDIA describes the module as delivering four times current bandwidth density and approximately 1.3x greater power efficiency versus 200Gb/s-per-lane pluggable optics, with 16 full-duplex lanes and plans to scale to 32. NVIDIA and Cisco are working with TE Connectivity and Amphenol to advance USFP as a community-led multi-source agreement; no sampling or availability date has been announced. NVIDIA also collaborated with Google on Ethernet specifications at the summit.

Meta announced it will use NVIDIA Spectrum-6 switches in its N-IceCube switch running FBOSS inside liquid-cooled IcePack racks. Spectrum-6 operates at 102.4 Tb/s and offers twice the switching capacity and per-lane bandwidth of Spectrum-4. Spectrum-X Multiplane distributes traffic across up to eight independent network planes, enabling GPU clusters beyond 100,000 GPUs with automatic rerouting on component failure. Omar Baldonado of Meta stated: 'We are excited about our close partnership with NVIDIA in making multiplane fabrics operationally scalable and performant, a critical requirement both for our own infrastructure builds and for our joint cloud partners.'

NVIDIA also released four open-source DSX AI Factory software tools derived from its experience running DGX Cloud: Config Manager for switch automation, the Cluster Readiness Engine for pre-training cluster stress tests, NVIDIA Rack Management Service for firmware management at rack scale, and Cluster API Provider for NVIDIA Infrastructure Controller, which the company says reduces bare-metal-to-production cluster setup from days to hours.