A common blueprint for AI factories: Nvidia opens up its Vera Rubin rack designAI 工厂建设迎来统一蓝图：英伟达开放 Vera Rubin 机架设计

Nvidia announced on 12 October that, at the Open Compute Project (OCP) Global Summit, it opened up the Vera Rubin rack and compute tray design and published a set of open AI infrastructure standards together with Cisco, Google, Meta, Microsoft and Visionbay. Through the open NVIDIA MGX architecture, Nvidia contributed the Vera Rubin rack and compute tray designs to OCP so system builders can develop infrastructure for next-generation AI factories. The updated MGX 1.2 specification adds rack-level blind-mate liquid cooling, interoperable rack interconnects, a 5,000A liquid-cooled busbar and native 19-inch device support; the architecture supports rack deployments up to 250kW and simplifies assembly, testing, maintenance and deployment. Nvidia also released four open-source software components for DSX AI Factory covering switch configuration, GPU cluster stress testing, rack firmware management and bare-metal deployment — derived from DGX Cloud operations experience, they let cloud providers turn hardware into a production cluster in hours. It also published Open Data Center Specification v0.7 with AMD, Google and Meta to unify power, cooling and physical layout for AI factories, plus a grid-resilience specification with Google, Meta and Microsoft requiring large AI facilities to stay connected during grid disturbances and return to near-full load in under two seconds. Meta will build Vera Rubin AI factories on open networking with Nvidia’s Spectrum-6 102.4Tb/s Ethernet switch, which doubles the switching capacity and per-lane bandwidth of Spectrum-4 and, with Spectrum-X Multiplane, supports clusters of more than 100,000 GPUs. With Cisco, Nvidia released USFP, the industry’s first pluggable optical module for 400Gb/s network architectures, offering 16 full-duplex lanes in a form factor close to OSFP — four times the bandwidth density and about 1.3x the energy efficiency of traditional 200Gb/s single-lane pluggables. With Google, Nvidia said future ConnectX-10 SuperNICs and BlueField DPUs will support Falcon Transport to improve reliability and bandwidth utilisation in Ethernet AI networks.