Nvidia contributed core rack and compute-tray designs for its forthcoming Vera Rubin AI systems to the Open Compute Project at the OCP Global Summit in San Jose on October 12.
The contribution, made through Nvidia's MGX 1.2 reference architecture, extends a modular blueprint to OEMs, ODMs and cloud operators for assembling high-density accelerated computing infrastructure. The move arrives as AI clusters scale from tens of megawatts toward gigawatt-scale facilities, where common mechanical and electrical interfaces can cut engineering time and speed deployment.
Nvidia said the MGX 1.2 specification adds rack-scale blind-mate liquid cooling, interoperable scale-out interconnects and a liquid-cooled 5,000 A busbar, along with native 19-inch equipment support, according to the company's newsroom. The architecture supports rack deployments up to 250 kW and includes configurable I/O bays and fanless liquid-cooled compute trays designed to streamline manufacturing and field service.
The company also contributed a separate Vera CPU rack specification built on the same MGX principles, positioning it for agentic AI workloads such as tool calling, orchestration and sandboxed code execution. It shares the modular cooling, power and networking foundation of the GPU-centric NVL72 systems, Nvidia said.
Open-source software components for DSX OS
Nvidia published four open-source software components for its DSX AI Factory operating layer, drawn from internal DGX Cloud operations. Nvidia Config Manager automates switch fabric configuration; Nvidia Cluster Readiness Engine stress-tests GPU clusters with workload simulations to detect hardware faults before training jobs begin; Nvidia Rack Management Service treats whole racks as manageable units for firmware and operations; and a Cluster API Provider for the Nvidia Infrastructure Controller accelerates bare-metal provisioning and tenant onboarding, reducing the time from raw hardware to production-ready clusters from days to hours, the company said.
Alongside the hardware and software contributions, Nvidia joined AMD, Google and Meta to publish the Open Data Center Specification and, with Google, Meta and Microsoft, the OCP Grid Resilience Specification, which defines how large-scale AI infrastructure should behave during utility grid disturbances. Both documents are presented by Nvidia as open standards for AI factory facility design and grid interaction.
Nvidia reports early ecosystem adoption
Nvidia said Meta will use its Spectrum-6 102.4 Tb/s Ethernet switches in liquid-cooled IcePack racks, combined with Spectrum-X Multiplane technology that distributes GPU traffic across up to eight independent network planes—a design Nvidia says can scale beyond 100,000 GPUs while reducing switch-tier costs. According to Nvidia, Meta's switch, called N-IceCube, runs the Facebook Open Switching System on the OCP Switch Abstraction Interface.
Nvidia also said Visionbay unveiled plans for K-1, an 800 V DC AI factory platform built around Nvidia's 800 V DC reference architecture and power components from Vertiv, and that it worked with Cisco to announce a new small form-factor pluggable standard for 400 Gb/s networking fabrics. Additionally, Google and Nvidia said future ConnectX-10 SuperNICs and BlueField DPUs will support Falcon Transport.
The announcement does not include pricing, general availability dates for Vera Rubin systems, or independent third-party performance results. All partner claims referenced derive solely from Nvidia's OCP summit statement.
Automatica reported in July that the MGX architecture was central to Nvidia's strategy for a multi-vendor supply chain. The new OCP contributions extend that foundation with rack-scale power and cooling specifications aimed at standardising deployments across system builders.