Meta Unveils Liquid-Cooled 102.4T Switches and Open Networking Stack at OCP 2026

Liquid-Cooled 102.4T Switching

At the 2026 Open Compute Project Global Summit, Meta presented IcePack, its first liquid-cooled network rack, based on the OCP ORV3-HPRv1 standard. The rack hosts IceCube switch blades delivering 102.4 Tbps of non-blocking capacity through 128×800G ports. IceCube comes in two variants, one built on Broadcom's Tomahawk6C ASIC and one on NVIDIA's Spectrum-6 ASIC, both running Meta's open-source FBOSS operating system on OCP SAI. Meta says the generation doubles the bandwidth of its 51.2T platforms while meeting the thermal density required for next-generation AI data centers.

Flattened Planar Fabric and Scale-Across Networking

Meta also introduced the Flattened Planar Fabric (FPF), a multiplane backend scale-out fabric in which each GPU connects through eight independent 100G links to eight different switches, one per network plane. The design moves from a 3-tier to a 2-tier Clos topology, reducing hops, latency and cost, and shifts load balancing into the NIC, which sprays packets across all eight planes in hardware. A combined control plane spanning switches and hosts lets traffic steer around failed planes. Meta developed FPF with NVIDIA, leveraging Spectrum-X Multiplane innovations, and is working with cloud partners to standardize multiplane implementations based on FPF learnings.

For multi-site clusters, Meta described its petabit-scale backend aggregation (BAG) super-spine, which interconnects spine fabrics across buildings, campuses and regions into a single mega cluster, with inter-BAG capacity between a pair of regions reaching tens of petabits per second. FBOSS now measures actual cable length in place to allocate buffer headroom, and eBGP link-bandwidth signaling distributes traffic in proportion to available path capacity. Meta has also deployed SRv6 on its regional backbone using FBOSS on Cisco SiliconOne switches.

Ethernet for Scale-Up and Open Standards

Meta introduced Tahan, an Ethernet-based scale-up switch for its MTIA 300 training accelerators, built on Broadcom's 51.2 Tbps Tomahawk 5 ASIC with 64x800G total capacity. Four Tahan switches interconnect 16 accelerators in a non-blocking topology, reusing the same lossless RoCE transport as the scale-out network under FBOSS. A next-generation 216×400G blade on 102.4 Tbps Tomahawk6C silicon will interconnect 72 accelerators per rack and is planned to adopt the OCP ESUN standard as its data plane.

The ESUN initiative, founded within OCP in 2025 with partners including AMD, Arista, ARM, Broadcom, Cisco, HPE, Marvell, Microsoft and NVIDIA, released its ESUN 1.0 Network Operator Requirements Base Specification and is nearing ESUN 1.1 with switch profiles and inter-packet gap reduction. Meta also co-leads the new Open Switch Platform Architecture (OSPA) initiative with Microsoft and Alibaba, standardizing the switch platform layer, with specification v1.0 submitted to OCP.

Open Transport and Ecosystem

Meta announced open-source implementations of MetaRoCE, its RDMA transport protocol optimized for AI cluster traffic over Ethernet. The company also said it has extended FBOSS so switch vendors can self-onboard their platforms without Meta engineers, with documentation and code available through the FBOSS documentation site and GitHub repository. Meta invited engineers, silicon vendors, system integrators and operators to join the OCP community in building open AI networking infrastructure.

Sources: Engineering At Meta