🌐Meta Open-Sources MetaRoCE and Shows 102.4T Switch at OCP 2026

Your AI cluster networking playbook just went public

Meta just open-sourced MetaRoCE, its RDMA transport, and put its entire AI networking playbook on the table at the 2026 OCP Global Summit. Talks and slide decks are now free for anyone to read. The catch? Meta is betting that open, multi-vendor gear beats locking into one stack. If you're building or operating GPU clusters, this is the most detailed public look yet at how a hyperscaler wires up AI at scale. The Flattened Planar Fabric connects each GPU to eight network planes, and the 3-tier to 2-tier Clos change is the kind of design decision you'd normally pay consultants to figure out. Teams running NCCL-heavy training jobs should read the scale-out and scale-across talks first. The hardware numbers are big. Meta's IcePack and IceCube rack runs 128 800G ports with 102.4T switch blades, and its Tahan scale-up switch puts 72 accelerators per rack on Broadcom Tomahawk5 silicon. The BAG super-spine links buildings, campuses, and regions in the gigawatt-scale Prometheus cluster. ESUN is moving from 1.0 to 1.1, and the v1.0 spec is already submitted to OCP.