Meta Engineering said on 12 October 2026 that it is showing new AI networking designs at the Open Compute Project (OCP) summit. IcePack is its first liquid-cooled network rack, made for tight power and cooling limits. Its IceCube switch blades move 102.4 Tbps, twice Meta's 51.2T platforms. Flattened Planar Fabric links each GPU to eight switches over eight 100G links. Meta also named Tahan, an Ethernet scale-up switch for its MTIA 300 accelerators. Four of them connect 16 accelerators in a rack.
Meta said it is co-leading the Open Switch Platform Architecture (OSPA) with Microsoft and Alibaba. It is a shared design for the switch platform layer, which Meta says needs months of custom work for each new platform. Meta also announced open-source MetaRoCE code, a data transport protocol for AI traffic over Ethernet. The post gives no date for the 72-accelerator switch blade or for the next Ethernet for Scale-Up Networking version. It does not give costs.
Meta says standard designs make mixed cloud setups affordable, because each difference adds engineering work. Buyers could ask vendors and clouds whether their hardware runs shared open software and passes common performance tests before handover.