OCP Global Summit: NVIDIA Opens Vera Rubin Rack Designs and Joins Google, Meta and Cisco on Open AI Factory Standards

On October 12, 2026, at the Open Compute Project (OCP) Global Summit in San Jose, NVIDIA announced a set of open contributions and partnerships around what it calls AI factories. NVIDIA's argument is that as AI factories grow from megawatts to gigawatts, building them takes more than advanced computing chips, and open standards help the industry scale AI infrastructure faster, more efficiently and more reliably. The news spans three layers: the AI factory itself, the utility grid that powers it, and the rack-scale architecture at the heart of the compute. The headline item is that NVIDIA contributed key Vera Rubin rack and compute tray designs to OCP through its open NVIDIA MGX architecture, giving system builders a common foundation for next-generation AI factory equipment. NVIDIA also joined industry leaders on the third-generation MGX data center reference design and paired it with open liquid-cooling technologies called MQD and MQDB. The aim is to turn proven AI factory designs into standards that others can adopt, manufacture and scale. The MGX 1.2 specification described by NVIDIA includes rack-scale blind-mate liquid cooling, interoperable rack-scale interconnect, a liquid-cooled 5,000A busbar, native 19-inch equipment support and higher-density rack designs. It also adds configurable I/O bays, fanless liquid-cooled compute trays and standardized power and cooling interfaces, with rack deployments up to 250kW. A separate Vera CPU rack design specification extends MGX beyond accelerated computing. Vera Rubin NVL72 serves large-scale training and inference, while the Vera CPU rack targets tool calling, orchestration and sandboxed execution, so agents can interact with applications, data and services in real time.

The second item is software. NVIDIA's DSX AI factory platform designs computing, networking, power, cooling, facilities and software as one system, and its open operating layer, DSX OS, gained several open source components. Config Manager automates switch configuration at scale. Cluster Readiness Engine stress-tests GPU clusters with real workload simulations before training begins, to avoid costly job crashes. Rack Management Service treats racks as first-class units for firmware management. Cluster API Provider for NVIDIA Infrastructure Controller shortens bare-metal deployment and tenant onboarding from days to hours. NVIDIA says the components come from its experience running DGX Cloud, and developers can explore and contribute on the DSX AI Factory GitHub page. The third item is standards published with the industry. NVIDIA joined AMD, Google and Meta to publish the Open Data Center Specification v0.7, a unified blueprint for AI factory facility design that covers power architecture, cooling infrastructure and physical layout. A separate OCP Grid Resilience Specification, published by Google, Meta, Microsoft and NVIDIA, defines how large AI infrastructure should behave during utility grid disturbances: stay connected, respond automatically and return to near-full load in under two seconds. With the standards open, equipment makers, colocation providers and operators have a common target and need not wait for one customer's proprietary specification. At the summit, Visionbay also unveiled plans for K-1, a platform built around an 800 VDC architecture with power and thermal technology from Vertiv, in preparation for future Vera Rubin deployments.

Networking is a focus too. Meta and NVIDIA announced that Meta is adopting NVIDIA Spectrum-6 102.4Tb/s Ethernet switches to build larger Vera Rubin AI factories on open networking. NVIDIA says Spectrum-6 delivers twice the switching capacity and per-lane bandwidth of Spectrum-4, at 200G per lane, with a lossless Ethernet AI fabric. Meta built a switch on it, N-IceCube, which slots into IcePack, its 100T liquid-cooled rack and runs FBOSS, Meta's open source network operating system. With Spectrum-X Multiplane, GPU traffic is spread in hardware across up to eight independent network planes, so AI factories can scale beyond 100,000 GPUs with fewer switch tiers and lower cost. If a component fails, traffic shifts to the remaining planes and GPUs keep working at reduced bandwidth. NVIDIA and Cisco also proposed USFP, an optical module form factor with about the footprint of OSFP that leads to 400Gb/s per lane, with 16 full-duplex lanes now and plans for 32, and a claimed four times the bandwidth density and roughly 1.3x better power efficiency. And NVIDIA and Google said future ConnectX-10 SuperNICs and BlueField DPUs will support Google's Falcon transport protocol, so NICs from both companies can talk over a common transport while applications keep standard RDMA interfaces.

Our reading is that the common thread is opening the interfaces. NVIDIA remains the largest supplier of compute, yet it chose to put the specifications for racks, power, cooling, networking and operations software into an open community, so that suppliers up and down the chain build against the same interfaces. For data center operators the benefit is that procurement and construction can start earlier. For NVIDIA it makes Vera Rubin easier to land in the ecosystem. One caveat: the source is NVIDIA's own blog, and every performance and efficiency figure here is a vendor statement. The real test is how many manufacturers adopt these specifications and when Vera Rubin systems actually ship.