Nvidia Opens Vera Rubin Rack Blueprints Via MGX 1.2 at OCP
TL;DR
- MGX 1.2 blind-mate liquid cooling removes all cables, cooling tubes, and fans from compute trays, enabling 250kW rack deployments without manual assembly.
- Meta's N-IceCube integrates Spectrum-6 switches into its Vera Rubin AI factory fabric, extending Nvidia's dependency from GPUs to the networking layer.
- The Grid Resilience facility spec published at OCP is version 0.7 draft status, meaning hyperscalers committing capex against it carry specification-change risk.
Nvidia used the Open Compute Project Global Summit in San Jose on October 12 to push its Vera Rubin rack blueprints into the open stack, contributing the rack and compute-tray designs to OCP through an updated MGX 1.2 specification that supports deployments up to 250kW and adds "rack-scale blind-mate liquid cooling, interoperable rack-scale interconnect infrastructure and a liquid-cooled 5,000A busbar designed for higher-power AI systems."
Alongside the rack, Nvidia and its partners published an Open Data Center Specification v0.7, framed as "a unified blueprint for next-generation AI factory facility design." A separate OCP Grid Resilience Specification sets how AI infrastructure should behave during utility disturbances, with the goal of "restoring to near-full load within less than two seconds."
Meta committed to the networking side of the stack. The company is "adopting NVIDIA Spectrum-6 102.4Tb/s Ethernet switches to build larger, higher-performance NVIDIA Vera Rubin AI factories on open networking," with Nvidia saying the Spectrum-X Multiplane technology layered on top is sized for "agentic AI factories of more than 100,000 GPUs." Meta is also introducing an N-IceCube switch that runs its FBOSS software.
With Cisco, Nvidia unveiled a pluggable form factor called USFP that the companies say "will provide a path to 400Gb/s-per-lane networking" at four times the bandwidth density of current optics. Google is adding Falcon Transport support to future ConnectX-10 SuperNICs, and Visionbay's K-1 platform contributes an 800 VDC architecture.
Nvidia also open-sourced four DSX software components covering switch configuration, cluster readiness, rack management, and bare-metal provisioning; the Config Manager, per Nvidia, "automates switch configurations at scale, allowing customers to deploy, validate and maintain consistent AI factory network fabrics with minimal to zero manual intervention." The announcement lands into a crowded OCP week on our AI infrastructure tracker, where AMD's Helion open rack with 32,768 EPYC cores was staked out at the same summit.
What others are reporting
-
StorageReview Read →
Connects hardware announcements to operational software (DSX OS) and hyperscaler deployment strategy, framing the specs as an integrated production platform rather than isolated product releases.
Building the world's largest AI factories starts with rethinking the smallest of connections.
-
Converge Digest Read →
Deep technical breakdown of power specs (250kW, 5,000A busbar), Meta's N-IceCube integration, all four DSX OS components, and a caveat that performance targets remain subject to production validation.
The initiatives extend OCP standardization from individual servers and switches to electrical distribution, liquid cooling, rack mechanics, software operations and network transport.
-
Futurum Group Read →
Analyst framing: manufacturing throughput has replaced silicon capability as the binding constraint, and Nvidia is converting dominant mindshare into structural dependency through standards leadership.
Building the world's largest AI factories starts with rethinking the smallest of connections.
-
MLQ News Read →
Critically examines what Nvidia did not open: GPU silicon, CUDA, and commercial networking are excluded, and OCP's royalty-free terms come with patent opt-out windows and member-only submission gates.
The contribution covers infrastructure specifications, not Rubin GPU silicon, CUDA or Nvidia's commercial networking products.
-
ExplainX Read →
Names component-level supply chain partners (Cisco, TE Connectivity, Amphenol), flags the Grid Resilience spec's v0.7 draft status, and breaks implications by buyer segment.
Meta is excited about its partnership with NVIDIA in making multiplane fabrics operationally scalable and performant.
-
HuggingNews Read →
Frames Meta's Spectrum-6 adoption as deepening Nvidia's position across Meta's full hardware stack even as Meta develops proprietary MTIA silicon.
These switches provide 102.4 Tb/s of switching capacity and allow Meta to link more than 100,000 GPUs while reducing the number of required network tiers.
Originally reported by blogs.nvidia.com
Read the original article →Original headline: Nvidia Opens Vera Rubin Rack Designs via MGX 1.2 at OCP Summit, Meta Adopts Spectrum-6 for 100K+ GPU Factories