Nvidia contributes Vera Rubin rack designs to OCP with updated MGX 1.2 specification

Nvidia contributed its Vera Rubin rack and compute-tray designs to the Open Compute Project on October 12 at the OCP Global Summit in San Jose, folding the blueprints into an updated MGX 1.2 specification that supports deployments up to 250kW. The contribution gives hyperscalers and colocation providers a reference architecture for AI clusters that require rack-scale blind-mate liquid cooling, a liquid-cooled 5,000A busbar, and interoperable interconnect infrastructure - the physical backbone needed to stand up 100,000-GPU factories without custom-engineering every rack.

Alongside the mechanical design, Nvidia and its partners released an Open Data Center Specification v0.7, which the company called "a unified blueprint for next-generation AI factory facility design." A separate OCP Grid Resilience Specification defines how AI infrastructure should behave during utility disturbances, with a target of restoring near-full load in under two seconds. For data center operators, that specification translates directly into uptime guarantees and generator-battery sizing decisions.

Meta commits to Spectrum-6 switching for Vera Rubin factories

Meta said it is adopting Nvidia Spectrum-6 102.4Tb/s Ethernet switches to build larger Vera Rubin AI factories on open networking. Nvidia's Spectrum-X Multiplane technology, layered on top of the switching fabric, is sized for "agentic AI factories of more than 100,000 GPUs," the company said. Meta also introduced an N-IceCube switch running its FBOSS software, signaling continued investment in disaggregated hardware-software networking stacks at scale.

The networking push comes as hyperscalers confront the bandwidth demands of training runs that span tens of thousands of accelerators. Spectrum-6 doubles the per-port throughput of the previous generation, and Meta's adoption means the switch will be tested inside one of the largest production AI environments on the planet. For network engineers planning multi-tenant AI clusters, the open-sourced rack designs and switch specifications reduce the guesswork on thermal and power budgets.

New form factors and interconnect standards emerge

With Cisco, Nvidia unveiled a pluggable form factor called USFP that the companies said "will provide a path to 400Gb/s-per-lane networking" at four times the bandwidth density of current optics. Google added Falcon Transport support to future ConnectX-10 SuperNICs, and Visionbay's K-1 platform contributed an 800 VDC architecture to the open stack. These moves point toward a rapid densification of the physical networking layer - more bits per lane, fewer cables per rack, and higher voltage distribution to cut copper losses.

DSX software stack goes open source

Nvidia open-sourced four DSX software components: switch configuration, cluster readiness, rack management, and bare-metal provisioning. Its Config Manager "automates switch configurations at scale, allowing customers to deploy, validate and maintain consistent AI factory network fabrics with minimal to zero manual intervention," according to Nvidia. For IT and operations teams, that means fewer touchpoints where a mistyped VLAN or BGP setting can delay a cluster bring-up by days.

The announcements landed during a crowded OCP week. AMD staked out its Helion open rack with 32,768 EPYC cores at the same summit, underscoring that the open-hardware race for AI infrastructure now spans both x86 CPU density and GPU-accelerated networking. Professionals responsible for AI Network Engineering Courses that cover fabric design at this scale will find the MGX 1.2 and Spectrum-6 blueprints becoming default reference material.

Why this matters for Finance, IT, Operations, and Real Estate professionals

For finance teams modeling AI infrastructure capex, the 250kW rack specification and 5,000A busbar create a clear per-rack power ceiling that feeds directly into facility cost projections. IT and network engineers get open, validated switch configurations and a 400Gb/s-per-lane optics roadmap that shortens design cycles. Operations staff gain a grid-resilience standard with a hard two-second recovery target - a number that directly shapes UPS and generator procurement. Real estate and construction leads can use the Open Data Center Specification v0.7 as a pre-engineered facility brief, reducing the architectural unknowns that typically inflate first-of-kind AI data center builds.