NVIDIA contributed rack and compute tray designs for its Vera Rubin architecture to the Open Compute Project at the OCP Global Summit in San Jose, NVIDIA announced. The contribution uses the third-generation NVIDIA MGX architecture to offer open blueprints for data centers scaling from megawatts to gigawatts.
Under the updated MGX 1.2 specification, the designs introduce rack-scale blind-mate liquid cooling, a liquid-cooled 5,000A busbar, and native 19-inch equipment support. The architecture accommodates deployments up to 250 kilowatts per rack while standardizing power and cooling interfaces. NVIDIA also released design specifications for a standalone Vera CPU rack built to manage tool calling, orchestration, and sandboxed execution for autonomous AI agents.
Data center and power blueprints
Hyperscale operators joined NVIDIA to publish facility guidelines through the OCP Open DC Initiative. AMD, Google, Meta, and NVIDIA published the Open Data Center Specification v0.7, establishing physical layouts, cooling infrastructure, and power distribution rules.
A related document, the OCP Grid Resilience Specification, came from Google, Meta, Microsoft, and NVIDIA. It requires large-scale computing infrastructure to remain connected during utility disturbances, respond automatically, and restore operations to near-full electrical load within two seconds.
Infrastructure provider Visionbay detailed plans for K-1, a platform using an 800-volt direct-current architecture built for Vera Rubin deployments. Developed with power and thermal vendor Vertiv, the design uses centralized power centers for alternating-current conversion to reduce downstream copper cabling.
Networking and optical hardware
Meta adopted NVIDIA Spectrum-6 102.4-terabit-per-second Ethernet switches to expand infrastructure beyond 100,000 graphics processors. Meta integrated the hardware into its liquid-cooled IcePack rack using a custom switch named N-IceCube running the Facebook Open Switching System.
The deployment relies on Spectrum-X Multiplane technology to split network traffic across up to eight independent planes using ConnectX-9 SuperNICs. “Extending FBOSS to Spectrum ASICs means we can run that same software stack across our fabrics,” said Omar Baldonado, senior director of data center and AI networking at Meta.
Cisco and NVIDIA unveiled the Ultra-Dense Small Form Factor Pluggable transceiver standard alongside component makers Amphenol and TE Connectivity. USFP packages 16 full-duplex lanes into the footprint of existing 200-gigabit OSFP modules, delivering 1.3 times the power efficiency with an upgrade roadmap to 32 lanes for 400-gigabit fabrics.
Google and NVIDIA confirmed support for Google's Falcon Transport protocol on future ConnectX-10 network adapters and BlueField data processing units, enabling hardware loss recovery and congestion control across shared Ethernet fabrics.
To assist multi-tenant cluster management, NVIDIA released four open source tools on GitHub for its DSX AI Factory operating system: the Config Manager network configuration tool, the Cluster Readiness Engine validation engine, the Rack Management Service, and an automated bare-metal provisioning controller.
