NVIDIA Opens Vera Rubin Architecture and Advances AI Factory Open Standards at OCP Global Summit

Image: NVIDIA · Source

At the OCP Global Summit, NVIDIA unveiled its open Vera Rubin rack and compute tray designs within the MGX architecture, new open-source software components for AI factory operations, and collaborations advancing AI infrastructure and networking standards.

NVIDIA has made significant announcements at the Open Compute Project (OCP) Global Summit, presenting new open architecture contributions and partnerships aimed at accelerating and standardizing AI factory infrastructure worldwide. Central to these developments is the opening of the NVIDIA Vera Rubin rack and compute tray designs under the MGX (Modular GPU eXtension) architecture, enabling system builders to develop scalable AI infrastructure based on a common foundation.

The Vera Rubin AI factories represent a next-generation platform scaling from megawatt to gigawatt power levels, integrating accelerated computing, networking, power, and cooling into a modular system. The updated MGX 1.2 specification supports rack-scale blind-mate liquid cooling, interoperable interconnects, a liquid-cooled 5,000A busbar for high-power AI systems, and native 19-inch equipment compatibility, facilitating deployments up to 250kW per rack with improved serviceability and efficiency.

Complementing hardware innovations, NVIDIA launched open-source software components for the DSX AI Factory platform, designed to unify computing, networking, power, and cooling operations across AI factories. New tools introduced include NVIDIA Config Manager for automated network configuration, the Cluster Readiness Engine to validate GPU clusters with realistic workloads, Rack Management Service for hardware abstraction at rack scale, and a Cluster API Provider to speed bare-metal provisioning. These components are publicly available on NVIDIA's DSX AI Factory GitHub repository.

NVIDIA also collaborated with hyperscalers such as Google, Meta, and AMD to publish open standards, including the Open Data Center Specification v0.7 and OCP Grid Resilience Specification. These define unified blueprints for AI factory facility design and responses to utility grid disturbances, enabling manufacturers and data center operators worldwide to develop interoperable AI infrastructure without waiting on proprietary specs.

Power infrastructure was addressed by Visionbay.ai, which revealed an 800V DC AI factory design tailored for Vera Rubin-era deployments. This architecture centralizes AC-to-DC conversion at cluster-level power centers and distributes 800V DC closer to compute resources, simplifying power delivery and supporting future modular expansion.

In networking, Meta and NVIDIA showcased the adoption of Spectrum-6 Ethernet switches and Spectrum-X Multiplane technology, enabling AI fabrics supporting over 100,000 GPUs with enhanced bandwidth, performance isolation, and resilience. NVIDIA and Cisco jointly announced the Ultra-Dense Small Form Factor Pluggable (USFP) module standard, promising 400Gb/s lane speeds and 1.3x improved power efficiency for AI factory-scale fabrics.

Additionally, NVIDIA and Google revealed support for the Falcon Transport protocol on future NVIDIA ConnectX-10 SuperNICs and BlueField DPUs, improving network reliability and performance in large-scale AI data centers.

These announcements underscore NVIDIA’s strategy to foster an open ecosystem for AI factory infrastructure where hardware, software, networking, and power systems are co-designed and standardized for rapid, efficient scaling of AI workloads globally. Industry partners and developers can access these open-source resources to build interoperable, next-generation AI factories that meet the growing demands of agentic and accelerated computing workloads.