Networking is the bottleneck

At the 2026 Open Compute Project (OCP) Global Summit, Meta shared the networking innovations supporting its latest AI models and its gigawatt-scale Prometheus cluster. The message: AI progress now depends on the network as much as on compute.

Hardware

  • IcePack and IceCube: Meta's first liquid-cooled network rack. The 102.4T switch blades offer 128×800G capacity, with Broadcom Tomahawk 6C and NVIDIA Spectrum-6 variants, running Meta's FBOSS network OS.
  • Tahan: An Ethernet scale-up switch for MTIA 300 accelerators, based on Broadcom Tomahawk5.
  • BAG: A petabit-scale super-spine connecting buildings, campuses and regions.

Open standards and software

MetaRoCE, Meta's RDMA transport for AI networks, is now open source. Meta, Microsoft and Alibaba are also co-leading OSPA, an initiative to standardize the switch platform layer; the v1.0 spec has been submitted to OCP.

Scaling architecture

For scale-out, Meta uses a flattened multi-plane fabric where each GPU connects to eight network planes. The three-tier Clos design was reduced to two tiers, with packet spraying moved to the NIC level. The work was built with NVIDIA and shared with cloud partners.