Meta Networking at the 2026 OCP Global Summit: Open Innovation for Scale-Up, Scale-Out, and Scale-Across AI Networks

Meta unveiled networking advancements at the 2026 Open Compute Project (OCP) Global Summit, innovations crucial for supporting their massive AI models like Muse Wonder and Muse Spark, alongside the gigawatt-scale Prometheus cluster.

A key highlight was the IcePack & IceCube, Meta's inaugural liquid-cooled network rack featuring 102.4T switch blades, utilizing 128×800G Broadcom Tomahawk 6C and NVIDIA Spectrum-6 variants, running FBOSS.

For scaling out, Meta presented the Flattened Planar Fabric (FPF), a multi-plane fabric where every GPU connects to eight network planes, combining a 3-tier to 2-tier Clos architecture, NIC-level packet spraying, and a unified host+switch control plane, built with NVIDIA and shared with cloud partners.

To achieve scale-across, they detailed BAG, their petabit-scale super-spine connecting buildings, campuses, and regions within the Prometheus cluster, augmented by SRv6 regional backbones on FBOSS + SAI.

In terms of scale-up, Tahan, an Ethernet scale-up switch designed for MTIA 300 based on Broadcom Tomahawk5, was introduced, previewing 72 accelerators per rack on 102.4T silicon, while ESUN advanced from 1.0 to 1.1.

Meta also introduced OSPA, a new OCP initiative co-led with Microsoft and Alibaba to standardize the switch platform layer, with its v1.0 specification submitted to OCP. Furthermore, their RDMA transport for AI networks, MetaRoCE, is now open source.

Presentations covered the necessity of rethinking infrastructure at a gigawatt scale, as detailed by Santosh Janardhan, Head of Infrastructure at Meta, emphasizing how AI has reshaped design and operation.

The necessity of an open approach was stressed by Omar Baldonado, Senior Director, Data Center & AI Networking, Meta, asserting that AI progress now relies on the network as much as on compute, requiring a full-stack, open ecosystem.