The contribution runs through MGX 1.2, which supports rack deployments up to 250kW and adds blind-mate liquid cooling, configurable I/O bays and native 19-inch equipment support. Connectors inside the trays allow for cold plates to be swapped out more easily without any unwanted spillage. A separate spec covers the Vera CPU rack, which Nvidia revealed has started shipping. That rack is aimed at tool calling, orchestration and sandboxed execution for AI agents. Nvidia previously did the same with Blackwell. At the summit, the grid resilience spec it co-published calls for recovery to near-full load in less than two seconds after utility disturbances. Meta is adopting Spectrum-6 102.4Tb/s switches for Vera Rubin AI factories past 100,000 GPUs. For AI companies, a shared rack design gives more manufacturers the same target to build for. That could shorten buildouts and widen supplier choice, while keeping its compute hardware directly intertwined and therefore the center of attention. The 250kW and two-second figures are spec targets, neither figure represents measured results at deployed sites, and which system builders will adopt MGX 1.2 is not stated.