Nvidia backs open standards for AI factory builds

NVIDIA has contributed Vera Rubin rack and compute tray designs to the Open Compute Project through its MGX architecture, as part of a broader push to establish open standards for AI infrastructure.

It also unveiled new open source software for its DSX AI Factory platform and announced separate networking and power infrastructure efforts with partners including Meta, Cisco, Google and Visionbay.

Open designs

NVIDIA said its latest MGX 1.2 specification extends its open architecture for AI factories with rack-scale blind-mate liquid cooling, interoperable rack-scale interconnect infrastructure and a liquid-cooled 5,000A busbar for higher-power AI systems. It also adds support for native 19-inch equipment and denser rack designs.

The updated design is intended to give system builders a common framework for AI data centre infrastructure based on the Vera Rubin generation of products. NVIDIA said the architecture supports rack deployments of up to 250kW.

NVIDIA also outlined a Vera CPU rack design based on the same third-generation MGX rack architecture. It said the system is aimed at workloads including tool calling, orchestration and sandboxed execution for AI agents, alongside larger-scale training and inference systems built on Vera Rubin NVL72.

Software layer

Alongside the hardware contribution, NVIDIA released open source software components for DSX OS, the operating layer within its DSX AI Factory platform. The additions cover network configuration, cluster validation, rack management and bare-metal provisioning.

The tools include NVIDIA Config Manager, Cluster Readiness Engine, Rack Management Service and a Cluster API Provider for NVIDIA Infrastructure Controller. NVIDIA said the software draws on its experience operating DGX Cloud and is intended for cloud providers and software partners running multi-tenant platforms.

Industry standards

NVIDIA also joined AMD, Google and Meta in publishing Open Data Centre Specification v0.7, which sets out a common approach to facility design for large AI installations, including power architecture, cooling infrastructure and physical layout.

A separate Grid Resilience Specification, published by Google, Meta, Microsoft and NVIDIA, sets expectations for how large AI facilities should respond during utility grid disturbances. According to NVIDIA, it covers staying connected to the grid, responding automatically and returning to near-full load in less than two seconds.

NVIDIA argued that open specifications would allow equipment manufacturers, colocation providers and data centre operators to design and build infrastructure without relying on customer-specific proprietary designs.

Power systems

Power distribution was another focus of the announcements. Visionbay introduced plans for K-1, an AI factory platform based on an 800 VDC architecture and designed with future Vera Rubin deployments in mind.

NVIDIA said the design combines its 800 VDC reference architecture with power and thermal systems from Vertiv. The approach centres on cluster-level power centres that convert AC to DC and distribute 800 VDC closer to compute resources, as operators contend with higher rack power densities.

The design is intended to reduce the need for additional copper, cabling and power-conversion equipment associated with lower-voltage power distribution. It would also allow operators to introduce 800 VDC in stages rather than through a full redesign.

Networking push

In networking, Meta said it is adopting NVIDIA Spectrum-6 102.4Tb/s Ethernet switches for larger Vera Rubin AI factory deployments. NVIDIA said the switches provide double the switching capacity and per-lane bandwidth of Spectrum-4.

Meta's switch, called N-IceCube, will run the Facebook Open Switching System, or FBOSS, the company's open source network operating system. NVIDIA said FBOSS is built on the Switch Abstraction Interface, an Open Compute Project standard it helped develop.

The collaboration also centres on Spectrum-X Multiplane technology, which distributes GPU traffic across up to eight independent network planes. NVIDIA said this is intended to support AI factory deployments of more than 100,000 GPUs with fewer switch tiers.

Meta provided the only direct quote in the announcements. "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," said Omar Baldonado, Senior Director of Data Centre and AI Networking at Meta. "Extending FBOSS to Spectrum ASICs means we can run that same software stack across our fabrics, and Meta's new Spectrum-based switch, N-IceCube, brings that into Icepack - our liquid-cooled 102.4T network rack for scale-out that connects upcoming generations of accelerators in our fleet."

NVIDIA and Cisco also announced the Ultra-Dense Small Form Factor Pluggable standard, or USFP, for optical transceivers. They said the format is designed to provide a path to 400Gb/s-per-lane networking while keeping a footprint similar to current OSFP modules.

According to NVIDIA, USFP packs 16 full-duplex lanes, with plans to scale to 32, offering four times the bandwidth density of conventional 200Gb/s-per-lane pluggable optics. The companies are working with TE Connectivity and Amphenol on connectors, cages and modules.

Separately, NVIDIA and Google said future ConnectX-10 SuperNICs and BlueField DPUs will support Falcon Transport, Google's transport protocol for Ethernet networks. NVIDIA said the aim is to improve efficiency and interoperability for large-scale AI networking while allowing NVIDIA and Google network interface cards to communicate over a common transport.