Meta opens its gigawatt AI network designs: a first liquid-cooled rack holds a 102.4T switch, and scale-out drops a tier
Synopsis
At the 2026 OCP Global Summit, Meta published the network designs behind its latest AI models and its gigawatt-scale Prometheus cluster. Scale-out had been organized as a three-tier Clos, and the switch platform layer stayed proprietary, with each new platform repeating months of bring-up work. Meta now connects each GPU to eight network planes and flattens Clos to two tiers, and shows IcePack, its first liquid-cooled network rack, holding IceCube, a 102.4T switch with 128 800G ports; MetaRoCE goes open source and the OSPA 1.0 specification goes to OCP.
Interpretation
Meta's scale-out fabric drops a tier: each GPU connects straight to eight network planes, the Clos runs in two tiers instead of three, NICs spray packets across paths, and hosts and switches share one control plane. Scale-out had been organized as a three-tier Clos, so traffic passed through one more stage of switching between GPU and peer. Meta built the flattened fabric together with NVIDIA and shared it with its cloud partners.
What carries those planes is a liquid-cooled switch: IceCube, a 102.4T switch with 128 non-blocking 800G ports, sits inside IcePack, Meta's first liquid-cooled network rack. The chassis is a 2OU, hybrid liquid/air-cooled, PHY-less design, using either Broadcom Tomahawk 6C or NVIDIA Spectrum-6 switch silicon and running FBOSS, Meta's own open switch operating system. These designs directly support Meta's latest AI models and its gigawatt-scale Prometheus cluster.
Perspective
What was released is the design and specification set presented at OCP 2026 for networks serving Meta's latest AI models and its gigawatt-scale Prometheus cluster: a two-tier flattened fabric with eight planes per GPU, IcePack as Meta's first liquid-cooled network rack with the 102.4T IceCube switch inside it, the BAG super-spine joining buildings and campuses, the OSPA 1.0 specification submitted to OCP, and open-source MetaRoCE. ESUN moves from a 1.0 to a 1.1 specification, and a DPDK-based MetaRoCE version is described as coming soon.
The 72 accelerators per rack remain a preview on 102.4T silicon. MetaRoCE is released first as a user-space implementation, with the DPDK version described only as coming soon. IceCube and its N-IceCube variant are designed as 2OU, hybrid liquid/air-cooled, PHY-less chassis.