NVIDIA Opens Vera Rubin Architecture to OCP as Tech Giants Unite on Gigawatt-Scale AI Factories

At the Open Compute Project (OCP) Global Summit in San Jose, NVIDIA announced that it is opening key designs from its next-generation Vera Rubin architecture to the industry. As data centers expand from megawatt footprints toward gigawatt-scale operations, the company is contributing rack and compute tray designs through the open NVIDIA MGX architecture to establish standardized foundations for AI factories worldwide.

The move comes alongside a sweeping set of infrastructure standards and hardware collaborations involving hyperscalers and networking leaders, including Meta, Google, Microsoft, Cisco, and Visionbay.

Open MGX 1.2 and Purpose-Built CPU Racks

Through the OCP Open DC Initiative, NVIDIA contributed to the third-generation MGX AI data center reference design. The updated MGX 1.2 specification incorporates rack-scale blind-mate liquid cooling, standardized power interfaces, fanless compute trays, and a liquid-cooled 5,000A busbar engineered for power-intensive deployments.

The architecture supports rack power densities reaching up to 250kW while adding native 19-inch equipment compatibility. In tandem with the Vera Rubin NVL72 platform designed for heavy training and inference workloads, NVIDIA introduced the Vera CPU rack design specification. Built on the MGX foundation, this platform is specifically tailored for agentic AI tasks, handling tool calling, orchestration, and sandboxed execution in real time.

Hyperscalers Standardize Facility Layouts and Grid Resilience

To prevent fragmentation in AI facility construction, NVIDIA collaborated with industry peers on two major open specifications:

• Open Data Center Specification v0.7: Developed alongside AMD, Google, and Meta, this document provides a standardized blueprint covering power architecture, cooling systems, and physical facility layout.
• OCP Grid Resilience Specification: Published with Google, Meta, and Microsoft, the standard dictates how mega-scale AI clusters handle utility grid disruptions. It establishes protocols for systems to remain connected, automatically adapt, and return to near-full power load in under two seconds.

Addressing escalating power densities, Visionbay unveiled plans for K-1, an 800V DC AI factory platform developed using NVIDIA reference designs and Vertiv power and thermal systems. K-1 utilizes cluster-level power centers to convert AC to 800V DC closer to compute trays, cutting down on copper cabling and conversion losses compared to traditional low-voltage distribution.

Next-Generation Networking: Spectrum-6, USFP, and Falcon

Networking emerged as another primary focus at the summit, highlighted by multiple architectural milestones:

• Meta Adopts Spectrum-6: Meta is deploying NVIDIA Spectrum-6 102.4Tb/s Ethernet switches within its liquid-cooled IcePack rack. The switch, dubbed N-IceCube, runs Meta's open-source FBOSS network operating system. Utilizing Spectrum-X Multiplane architecture with ConnectX-9 SuperNICs, the configuration distributes GPU traffic across up to eight network planes, supporting clusters exceeding 100,000 GPUs.
• Cisco and NVIDIA Introduce USFP: The two companies revealed plans for the Ultra-Dense Small Form Factor Pluggable (USFP) transceiver module. Retaining a footprint similar to the current 200Gb/s-per-lane OSFP standard, USFP offers 16 full-duplex lanes (with roadmaps extending to 32 lanes) for 400Gb/s-per-lane links. The design achieves four times the bandwidth density and roughly 1.3 times greater power efficiency, with ecosystem support from Amphenol and TE Connectivity.
• Google Falcon Transport Integration: NVIDIA and Google confirmed support for Google's Falcon Transport protocol on upcoming NVIDIA ConnectX-10 SuperNICs and BlueField DPUs, extending reliable hardware-accelerated multipath Ethernet transmission across large AI clusters.

Open-Source Software for DSX AI Factories

To orchestrate these physical systems, NVIDIA also released new open-source software tools for its DSX AI Factory operating layer. Derived from operational practices used on NVIDIA DGX Cloud, the modules include the NVIDIA Config Manager for automated switch deployment, the Cluster Readiness Engine for stress-testing GPU clusters prior to training jobs, the Rack Management Service for unified firmware operations, and the Cluster API Provider for bare-metal multi-tenant provisioning.