Meta Unveils Liquid-Cooled IcePack Rack and Open MetaRoCE Protocol at OCP 2026

Meta presented its next-generation open networking infrastructure at the 2026 Open Compute Project Global Summit, focusing on scale-up, scale-out, and scale-across architectures for AI. The announcements include the IcePack, Meta’s first liquid-cooled network rack, and the IceCube switch platform, which delivers 102.4 Tbps of bandwidth. These innovations aim to support gigawatt-scale AI clusters by improving thermal efficiency and density.

Scale-Out and Scale-Across Innovations

The IceCube blades offer 128×800G non-blocking switching capacity and come in two variants using Broadcom’s Tomahawk6C and NVIDIA’s Spectrum-6 ASICs, both running Meta’s open-source FBOSS network operating system. This generation represents a 2x bandwidth increase over previous 51.2T platforms.

Meta also introduced the Flattened Planar Fabric (FPF), a multiplane backend scale-out fabric that replaces traditional single-plane uplinks. In FPF, each GPU connects via eight independent 100G links to eight different switches, enabling larger non-blocking clusters and reducing latency by moving from a 3-tier to a 2-tier Clos topology. Load balancing is handled at the NIC level, spraying packets across all eight planes in hardware.

For multi-site clusters, Meta detailed its Backend Aggregation (BAG) layer, a petabit-scale super-spine that interconnects fabrics across regions. BAG uses SRv6 regional backbones and measures actual cable length to optimize buffer headroom, addressing the challenges of lossless Ethernet over long distances.

Scale-Up and Open Standards

Meta demonstrated Tahan, an Ethernet-based scale-up switch for its MTIA 300 accelerators, built on the 51.2 Tbps Tomahawk 5 ASIC. Tahan delivers 6.4 Tbps per accelerator and unifies scale-up and scale-out fabrics under FBOSS. A next-generation design is previewed to support 72 accelerators per rack on 102.4 Tbps silicon, adopting the Ethernet for Scale-Up Networking (ESUN) standard.

The ESUN initiative, co-led by Meta, Microsoft, and Alibaba, has released the ESUN 1.0 specification and is working on ESUN 1.1 to further optimize Ethernet for low-latency scale-up use cases. Additionally, Meta launched the Open Switch Platform Architecture (OSPA) initiative to standardize the switch platform layer, reducing onboarding time for new hardware.

Open Transport and Ecosystem

Meta announced open-source implementations of MetaRoCE, its RDMA transport protocol optimized for AI cluster networks. This move allows the broader community to adopt a production-proven transport stack for collective communication. Meta also highlighted its collaboration with cloud partners to standardize performance validation and normalize heterogeneous fabrics, ensuring consistent behavior across on-prem and public cloud environments.