Meta Unveils Liquid-Cooled Networking Innovations for AI at OCP Summit
Meta announced new networking solutions at the 2026 Open Compute Project (OCP) Global Summit, designed to support its latest AI models and gigawatt-scale Prometheus cluster. These innovations include IcePack, the company's first liquid-cooled network rack with 102.4T switch blades, and IceCube, a 102.4T switch platform. Meta also introduced Flattened Planar Fabric (FPF), a multiplane scale-out fabric that connects each GPU to eight network planes. These advancements aim to address the growing demand for high-performance, scalable AI infrastructure by improving efficiency and reducing latency. MetaRoCE, an open-source AI transport protocol, was also highlighted as a key component of the new networking stack.

Meta unveiled several networking innovations at the 2026 OCP Global Summit to support its AI infrastructure, including the Muse Wonder and Muse Spark models and the Prometheus cluster. Key innovations include IcePack, the first liquid-cooled network rack with 102.4T switch blades, and IceCube, a 102.4T switch platform built with Broadcom Tomahawk 6C and NVIDIA Spectrum-6 variants. These technologies are designed to improve efficiency and scalability in large-scale AI operations.
Flattened Planar Fabric (FPF) was introduced as a multiplane scale-out fabric where each GPU connects to eight network planes. This approach reduces latency and improves performance by enabling more direct communication between GPUs and the network. Meta also highlighted the 3-tier to 2-tier Clos architecture, NIC-level packet spraying, and a combined host+switch control plane.
Meta's scale-across initiatives include BAG, a petabit-scale super-spine that connects buildings, campuses, and regions within the Prometheus cluster. SRv6 regional backbones on FBOSS + SAI were also demonstrated as part of the scale-across strategy. These innovations are intended to support the rapid growth of AI workloads by enabling seamless connectivity across data centers and regions.
In addition to hardware advancements, Meta introduced MetaRoCE, an open-source AI transport protocol, which is now available for use in AI networks. This protocol is designed to enhance performance and reduce latency in AI clusters by optimizing data transmission across the network. Meta also shared insights into its new OCP initiative, OSPA, which aims to standardize the switch platform layer.