NVIDIA has contributed its Vera Rubin rack and compute tray blueprints to the Open Compute Project to standardize AI factory infrastructure. The company released these designs alongside shared hardware and networking specifications developed with Cisco, Google, Meta, Microsoft, and AMD. The move aims to accelerate enterprise data center buildouts while standardizing power, cooling, and network hardware across the industry.
The updated MGX 1.2 framework introduces rack level blind mate liquid cooling, standardized interconnections, and a liquid cooled 5000A busbar with native 19 inch chassis compatibility. Racks built on this standard can handle power loads up to 250kW. To speed up software installations, NVIDIA released 4 open source programs for the DSX AI Factory platform. These tools automate switch configuration, GPU stress testing, firmware management, and bare metal server provisioning, reducing setup timelines down to hours.
Hardware standardization extends beyond the chassis through collaborations with competing chipmakers and major cloud operators. NVIDIA, AMD, Google, and Meta published the Open Data Center Specification Version 0.7, unifying power delivery, thermal management, and physical floor layouts. A separate electrical grid stability guideline created with Microsoft, Google, and Meta requires high capacity facilities to withstand electrical disruptions and return to near peak load in less than 2 seconds.
Meta selected NVIDIA Spectrum 6 102.4Tb per second Ethernet switches to operate its upcoming infrastructure. The switch doubles the capacity and lane bandwidth of Spectrum 4, scaling clusters beyond 100000 GPUs using Spectrum X Multiplane routing. For physical connectivity, NVIDIA and Cisco developed the USFP transceiver standard, supplying 16 full duplex channels for 400Gb per second setups with 4 times the bandwidth density and 1.3 times the power efficiency of legacy optics. Future ConnectX 10 SuperNIC adapters and BlueField DPU chips will also incorporate the Google Falcon Transport protocol to reduce Ethernet packet loss.
