On October 12, 2026, Meta used its official engineering blog to unveil the next generation of open networking infrastructure prepared for the OCP Global Summit: IcePack, its first liquid-cooled network rack; the IceCube 102.4T switch platform; Flattened Planar Fabric (FPF), a multiplane backend fabric; the open MetaRoCE transport protocol; and a new Open Switch Platform Architecture initiative, OSPA. The switching layer rarely headlines the news, but Meta is positioning it as a precondition for gigawatt-scale AI clusters.
Two chip routes for IceCube, one constant: FBOSS
IcePack follows OCP's ORV3-HPRv1 rack standard and holds IceCube switch blades: 102.4 Tbps of fabric capacity with 128 non-blocking 800G ports, twice the bandwidth of the previous 51.2T generation. Notably, IceCube comes in two variants, one built on Broadcom's Tomahawk 6C switch ASIC and one on NVIDIA's Spectrum-6, yet both run Meta's open-source FBOSS network operating system through OCP's SAI abstraction layer. In other words, Meta keeps a dual-vendor strategy at the most critical networking layer, avoiding lock-in to a single chip supplier while smoothing over the differences in its own software. Alongside it sits Tahan, a 51.2T Ethernet-based scale-up switch serving Meta's MTIA 300 system at 6.4 Tbps per accelerator; the next design fits 72 accelerators per rack and moves to 102.4T silicon.
FPF flattens the fabric from three tiers to two, and pushes fault tolerance to the NIC
The architectural change runs deeper than raw bandwidth. Instead of a single uplink per GPU, FPF connects each GPU with eight independent 100G links to eight different switches across eight network planes, flattening the fabric from a three-tier Clos to two tiers: one fewer switch layer means fewer hops, lower latency and lower cost. Load balancing no longer lives only in the switches; it starts at the NIC, which sprays packets across all eight planes in hardware, and when a plane fails, the host-side control plane steers traffic around it. Meta says FPF was honed with NVIDIA around Spectrum-X Multiplane technology, and that it is sharing the lessons with joint cloud partners to standardize multiplane fabrics beyond its own data centers. One layer further out, Meta also disclosed a petabit-scale backend aggregation super-spine, using SRv6 regional backbones to support multi-site clusters spanning buildings and campuses.
Why a social and model company open-sources switches
OSPA, launched the same day, is co-led by Meta, Microsoft and Alibaba, and aims to standardize the switch platform layer; specification v1.0 has already been submitted to OCP, while MetaRoCE now has open-source implementations. Read together, the logic is plain: as clusters head toward 100,000 GPUs and gigawatt power draws, network power, cooling and failure rates decide whether a training run finishes at all, and the switching layer has become as much of a bottleneck as the GPUs themselves. By turning racks, fabrics and protocols into open specifications wherever possible, Meta buys supply-chain leverage and an ecosystem that follows its lead. For everyone else, the practical message is that the default network spec of the next AI factory generation is being written early: liquid cooling, 102.4T, multiplane fabrics and Ethernet scale-up. Latecomers either adopt this open route or have to prove a closed design is worth its higher cost.
