Summary

NVIDIA says it has contributed Vera Rubin rack and compute-tray designs to OCP, alongside new AI-factory standards, software and networking work with industry partners.

At the OCP Global Summit on October 12, 2026, NVIDIA announced that it had contributed Vera Rubin rack and compute-tray designs to the Open Compute Project (OCP) through its MGX architecture. The announcement also covered open specifications for AI-factory facilities, new software components and networking work with partners including Meta, Cisco and Google.

The package reaches beyond server hardware: it sets out designs and standards for parts of an AI factory, from facility power and cooling to racks, network connections and cluster management. NVIDIA presents these as a common foundation that equipment makers and operators can build on.

Contents

Rack architecture and power

The updated MGX 1.2 specification adds rack-scale blind-mate liquid cooling, interoperable rack interconnects and a liquid-cooled 5,000A busbar. It also supports native 19-inch equipment and higher-density rack designs. Other features include configurable I/O bays, fanless liquid-cooled compute trays and standardised power and cooling interfaces. NVIDIA says the architecture supports rack deployments up to 250kW.

A blind-mate connection is designed to join components as they are brought together, without requiring a technician to make each connection separately. At rack scale, standardised interfaces can make it easier for different suppliers’ equipment to fit into a shared design. The specification’s cooling and power features are aimed at systems with higher energy and heat loads.

NVIDIA also described a Vera CPU rack specification based on MGX. While the company positions Vera Rubin NVL72 for large-scale training and inference, it says the CPU rack is designed for AI-agent tasks such as tool calling, orchestration and sandboxed execution. Both designs use common rack-scale principles for integrating computing, networking, power and cooling.

Facility standards and operations software

NVIDIA, AMD, Google and Meta published Open Data Center Specification v0.7, a blueprint covering AI-factory facility design, including power architecture, cooling and physical layout. A separate OCP Grid Resilience Specification, published by Google, Meta, Microsoft and NVIDIA, describes how large AI facilities should respond to utility-grid disturbances: remain connected, respond automatically and return to near-full load in less than two seconds. That is the behaviour the specification sets out for infrastructure to support.

NVIDIA also introduced open-source components for its DSX AI Factory platform. They include Config Manager for automating switch configuration; Cluster Readiness Engine for testing GPU clusters with workload simulations before training; Rack Management Service for rack-level hardware and firmware operations; and a Cluster API provider intended to help provision bare-metal infrastructure and onboard tenants.

These components address operational tasks that become more involved as clusters grow: configuring networks consistently, checking hardware readiness and preparing machines for workloads. NVIDIA says they draw on its experience operating DGX Cloud and are available through its DSX AI Factory GitHub project.

Networking for larger AI systems

Meta is adopting NVIDIA Spectrum-6 Ethernet switches, specified at 102.4Tb/s, for future Vera Rubin AI factories. NVIDIA says Spectrum-6 doubles the switching capacity and per-lane bandwidth of Spectrum-4. Meta’s N-IceCube switch uses Meta’s FBOSS network operating system and is designed for its liquid-cooled IcePack rack.

NVIDIA and Meta also described Spectrum-X Multiplane, which distributes GPU traffic across as many as eight network planes using Spectrum-6 switches and ConnectX-9 SuperNICs. The companies say multiple paths can improve network utilisation and allow traffic to shift if a component fails; they describe the design as supporting systems beyond 100,000 GPUs with fewer switch tiers.

NVIDIA and Cisco announced plans for USFP, a pluggable optical module form factor designed for 400Gb/s-per-lane networking. The proposed design has 16 full-duplex lanes, with plans to scale to 32; NVIDIA says it could provide roughly 1.3 times the power efficiency of traditional 200Gb/s-per-lane pluggable optics. Separately, NVIDIA and Google announced Falcon Transport support on future ConnectX-10 SuperNICs and BlueField DPUs, bringing Google’s Ethernet transport protocol to that hardware.

Sources