NVIDIA opens Vera Rubin rack and compute tray designs to OCP (MGX 1.2, up to 250kW per rack) and adds a Vera CPU rack spec for AI agents; Meta adopts Spectrum-6 as four firms publish a grid resilience spec

Why it matters

【Key points】

At the Open Compute Project (OCP) Global Summit in San Jose on Oct 12, NVIDIA contributed its next-generation Vera Rubin rack and compute tray designs to OCP through the open MGX architecture, so system builders can build Vera Rubin racks from a common design.

The updated MGX 1.2 spec supports racks up to 250kW and adds rack-scale blind-mate liquid cooling, a liquid-cooled 5,000A busbar, fanless liquid-cooled compute trays and native 19-inch equipment. NVIDIA also published a separate Vera CPU rack spec aimed at AI agents' tool calling, orchestration and sandboxed execution, distinct from the Vera Rubin NVL72 used for training and inference.

Also announced: an Open Data Center Specification v0.7 with AMD, Google and Meta; an OCP Grid Resilience Specification from Google, Meta, Microsoft and NVIDIA (ride through grid disturbances, near-full load in under 2 seconds); four new open-source DSX OS components; Meta adopting Spectrum-6 102.4Tb/s Ethernet switches with multiplane networking for 100,000+ GPU clusters; a USFP optics form factor with Cisco for 400Gb/s per lane; and Google Falcon transport support on future ConnectX-10 and BlueField.

【Verifiable】

NVIDIA blog (Oct 12): https://blogs.nvidia.com/blog/ocp-global-summit-vera-rubin-scale-ai-factories/

The figures (250kW, 5,000A, 102.4Tb/s) are NVIDIA's own; no independent measurements yet.

【Our read】

1. An open Vera Rubin rack design lets ODMs and builders start ahead of any single customer's private spec; the next bottleneck is power, cooling and facilities, not just chips.

2. A dedicated Vera CPU rack for agents signals NVIDIA sees agent tool calls and sandboxes as a large enough workload to justify its own compute tier.

3. Four companies co-signing a grid resilience spec reflects that gigawatt AI factories now strain utility grids. With Microsoft scaling its own Maia chips to multiple gigawatts the same day (e/828), in-house silicon and NVIDIA's open standards will share the same buildings.