Through its MGX architecture, Nvidia has handed its Vera Rubin rack and compute tray designs to the Open Compute Project, handing system builders a shared basis for AI factory hardware. The news came out at the OCP Global Summit in San Jose.
A liquid-cooled 5,000A busbar for higher-power systems, rack-scale blind-mate liquid cooling, and interoperable interconnect infrastructure are all added by the updated MGX 1.2 specification. Racks reaching 250 kilowatts are supported, and native 19-inch equipment support is now included.
Cisco took the same stage to introduce the Silicon One G302, billed as the first 51.2T switching chip featuring 200G SerDes and aimed at top-of-rack deployments. Direct-to-chip liquid cooling is built into the G302, and shipping is expected before the end of 2026.
Microsoft has entered production on systems driven by the earlier Silicon One P200, Cisco also said. Alibaba reported deploying that same silicon inside its eCORE backbone routers, matching Cisco’s part with its own Quantum Cat systems and the SONiC operating system.
The DS6005 came from Celestica, a fully liquid-cooled switch pushing 102.4Tbps through co-packaged optics. Optical engines sit directly on the switch package rather than in front-panel pluggables, a design built on Broadcom’s Tomahawk 6 chip and housed in a 1.5OU enclosure for OCP ORv3 racks.
ASUS brought the XA P2N-E2, a 2U server built on Nvidia’s MGX architecture with dual Vera CPUs. Under Spatial Multithreading each socket carries 88 Olympus cores and 176 threads, aimed at sandboxed code execution and agent orchestration rather than model training.
