Meta Details AI Networking Innovations at OCP Global Summit 2026

Meta's AI Networking Innovations

At the 2026 Open Compute Project (OCP) Global Summit, Meta showcased its latest networking innovations designed to support its AI models, including Muse Wonder and Muse Spark, and its Prometheus cluster. These advancements focus on scaling AI infrastructure across various dimensions.

Key Hardware and Fabric Developments

Meta introduced IcePack & IceCube, its first liquid-cooled network racks featuring 102.4T switch blades with Broadcom Tomahawk 6C and NVIDIA Spectrum-6 variants. For scale-out, Meta detailed its Flattened Planar Fabric (FPF), a multi-plane fabric where each GPU connects to eight network planes, transitioning from a 3-tier to a 2-tier Clos architecture. The company also presented BAG, a petabit-scale super-spine for scale-across connectivity, linking buildings, campuses, and regions within its gigawatt-scale Prometheus cluster.

Ethernet and Open Standards

Ethernet advancements include Tahan, an Ethernet scale-up switch for MTIA 300 based on Broadcom Tomahawk5, and a preview of 72 accelerators per rack on 102.4T silicon. Meta also co-led the Open Switch Platform Abstraction (OSPA) initiative with Microsoft and Alibaba, aiming to standardize the switch platform layer, with v1.0 specifications submitted to OCP. Additionally, Meta announced the open-sourcing of MetaRoCE, its RDMA transport for AI networks.

Addressing AI Infrastructure Challenges

Meta's Head of Infrastructure, Santosh Janardhan, highlighted the need to rethink network, facility, and system design for gigawatt-scale AI infrastructure. Omar Baldonado, Senior Director of Data Center & AI Networking, emphasized that AI progress now depends significantly on networking, advocating for a full-stack, open approach. Vipul Deokar, VP, Core Network Infrastructure, discussed the necessity of open scale-across networking to connect AI capacity across data centers and regions, as training compute grows faster than infrastructure can scale.