OCP 2026: Meta's Liquid Cooling and the 102.4T IceCube for Gigawatt AI Clusters

At OCP Summit 2026, Meta unveiled its next generation of open networking technology for gigawatt-scale AI infrastructure. The announcements span the full spectrum: from in-rack accelerators (scale-up), through buildings and campuses (scale-out), to regional scale (scale-across). These innovations directly support Meta's AI models: Muse Wonder, Muse Spark, and the personal AI agent Muse.

Liquid Cooling for Petabit Switching: IcePack and IceCube

Meta introduced its first liquid-cooled networking rack, IcePack (built on the OCP ORV3-HPR v1 standard), alongside the IceCube switches — 102.4-terabit fabrics with a non-blocking 128×800G architecture. IceCube ships in two variants: one built on the Broadcom Tomahawk6C and one on the NVIDIA Spectrum-6 ASIC. Both versions run on Meta's open network operating system, FBOSS, using OCP SAI (Switch Abstraction Interface).

This generation delivers a twofold bandwidth increase over 51.2T platforms while achieving the thermal density required for next-generation AI data centers. The architecture is designed to serve gigawatt-scale clusters with high energy efficiency.

Flattened Planar Fabric: Multi-Plane GPU Connectivity

Flattened Planar Fabric (FPF) is a scaling architecture that fundamentally changes how GPUs connect to the network. In FPF, each GPU connects via eight independent 100G links to eight different switches — one per network plane. This multi-plane approach replaces the traditional single-plane architecture with four key advantages: larger non-blocking GPU clusters for the biggest AI training runs; a simplified 2-tier Clos topology instead of 3-tier, reducing network hops and latency; hardware-level load balancing at the NIC that spreads packets across all eight planes; and a unified control plane that lets hosts reroute traffic around failed planes.

Meta developed FPF jointly with NVIDIA, building on Spectrum-X Multiplane innovations to deliver high performance across a range of failure scenarios. Meta and NVIDIA are now working with cloud partners to standardize multi-plane fabrics based on FPF, ensuring consistent network behavior regardless of where GPUs are physically located.

Petabit-Scale Expansion: BAG and Regional Interconnection

Meta is building AI compute at an unprecedented scale: the Prometheus cluster, at roughly 1 GW in 2026, and Hyperion, scaling toward 5 GW. Backend Aggregation (BAG) is a centralized Ethernet super-spine that stitches together numerous spine-layer fabrics into a single mega-cluster spanning buildings, campuses, and regions, operating at petabit scale — with bandwidth between a pair of regions reaching tens of petabits per second.

Meta has deployed regional backbones on Segment Routing over IPv6 (SRv6) using FBOSS, running on open platforms with shallow buffering based on Cisco SiliconOne. At these distances, physical distance itself becomes a design parameter: FBOSS measures actual cable length by synchronizing probe traffic over an idle queue to determine round-trip distance, then sizes buffer headroom based on that measurement. Training remains resilient to failures: if a replica goes down, it's automatically excluded, and the remaining replicas continue synchronizing through BAG.

Ethernet for

In-Rack Scaling: Tahan and the 72-Accelerator Design

In 2026, Ethernet is proving itself in the scale-up domain — the realm of high bandwidth and ultra-low latency that once required proprietary interconnects. Tahan is Meta's Ethernet scale-up switch for its MTIA 300 accelerators, built on the Broadcom Tomahawk 5 ASIC with 51.2 Tbps of bandwidth and a total capacity of 64×800G. Four Tahan switches connect 16 accelerators in a non-blocking topology, delivering 6.4 Tbps per accelerator.

Meta is already developing the next generation: 72 accelerators per rack powered by 102.4T silicon, alongside a shift to ESUN (Ethernet for Scale-Up Networking) — an OCP initiative to standardize Ethernet-based solutions for in-rack scaling.

Open Switch Platform Architecture and Open Protocols

Meta is co-leading a new OCP initiative — the Open Switch Platform Architecture (OSPA) — together with Microsoft and Alibaba. The initiative standardizes the switch platform layer to ensure compatibility across different silicon vendors; the v1.0 specification has already been submitted to OCP. MetaRoCE, Meta's open RDMA transport protocol for AI networking, is now open-sourced and designed for reliable data delivery during training on large-scale clusters.

Meta is inviting engineers, silicon vendors, system integrators, and operators to join the OCP community in building open, energy-efficient, and sustainable AI infrastructure.