Meta Unveils Liquid-Cooled 102.4T Switches and Multiplane Fabrics at OCP Summit
Open networking innovations target gigawatt-scale AI infrastructure across scale-up and scale-out domains.
1Open AI Infrastructure Expansion
At the 2026 OCP Global Summit, Meta introduced its next-generation open networking infrastructure designed for gigawatt-scale AI clusters. The portfolio includes liquid-cooled 102.4T switch platforms, multiplane scale-out architectures, open transport protocols, and standardized scale-up interconnects. These systems support large training models, including Meta's proprietary Muse series architectures.
2Liquid-Cooled Switching at 102.4T
Meta launched IcePack, its first liquid-cooled network rack compliant with the OCP ORV3-HPRv1 standard. Operating inside the rack are IceCube blades providing 102.4 Tbps non-blocking bandwidth across 128 800G ports. The hardware utilizes Broadcom Tomahawk6C and NVIDIA Spectrum-6 ASICs, running Meta's open-source FBOSS operating system via SAI.
3Flattened Planar Fabric Architecture
Meta introduced Flattened Planar Fabric to overhaul how GPUs connect to backend networks. Instead of standard single-plane uplinks, each accelerator connects across eight independent 100G links to distinct network planes. This transition reduces network tiers from a 3-tier Clos to a 2-tier design, cutting latency, hops, and hardware costs.
4Hardware-Level Traffic and Fault Management
In the FPF design, packet spraying and load balancing move directly to the network interface card in hardware across all eight planes. The control plane spans hosts and switches simultaneously, allowing endpoints to bypass disabled planes dynamically. Meta co-developed this architecture with NVIDIA using Spectrum-X Multiplane capabilities.
5Standardizing Hybrid Cloud Training Fabrics
To supplement data center lead times, Meta combines owned campuses with leased facilities and public cloud capacity. To ensure operational consistency across heterogeneous networks, Meta deploys standardized benchmarking up to thousands of GPUs. Incompatible traffic-class schemes and congestion parameters are mapped directly onto uniform lossless RDMA profiles across vendors.
6Petabit-Scale Backend Aggregation Super-Spine
Meta deployed an Ethernet-based Backend Aggregation layer to interconnect distributed fabrics across regions into single mega-clusters like Prometheus and Hyperion. Inter-BAG bandwidth reaches tens of petabits per second between geographic regions. This super-spine abstracts physical location constraints, enabling synchronized AI training across multiple physical facilities and campuses.
7Distance-Aware Buffer and Routing Optimizations
Long-distance lossless Ethernet requires dynamic buffer sizing proportional to link length. FBOSS actively measures live round-trip cable distance using uncongested probes to allocate buffer headroom accurately. Meanwhile, BAG implements eBGP link-bandwidth signaling to distribute traffic across degraded paths proportionally without generating widespread routing instability across networks.
8Regional Backbones Using FBOSS SRv6
For regional wide-area backbones interconnecting campuses and public cloud instances, Meta deployed Segment Routing over IPv6. Managed completely through FBOSS via OCP SAI, the backbone runs on shallow-buffer platforms using Cisco SiliconOne ASICs. The design delivers flexible traffic engineering while simplifying multi-region cluster scaling without proprietary protocols.
9Tahan Ethernet Scale-Up Switch
Meta announced Tahan, a 51.2 Tbps liquid-cooled scale-up switch dedicated to MTIA 300 accelerators. Built on Broadcom Tomahawk 5 silicon, four switches interconnect 16 accelerators non-blockingly, delivering 6.4 Tbps per chip. Tahan uses routed IPv6, PFC, and AI-tuned QoS under FBOSS, unifying scale-up and scale-out transports.
10Next-Gen 72-Accelerator High-Density Racks
Meta previewed a next-generation scale-up switch blade packing 216 400G ports on Broadcom Tomahawk6C 102.4T silicon. Eight integrated switches connect 72 accelerators per rack, drastically expanding scale-up density. Future iterations will adopt the open Ethernet for Scale-Up Networking data plane standard to replace proprietary domain interconnects.
11Progress on Open ESUN Specifications
The OCP ESUN initiative, founded by Meta alongside major semiconductor and system partners, released its 1.0 base specification. ESUN focuses exclusively on network fabric requirements, preserving GPU implementation flexibility while coordinating with IEEE and UEC. The upcoming 1.1 revision targets switch profiles and inter-packet gap reduction mechanisms.
12Open Standardization Across AI Stacks
Alongside network hardware, Meta submitted OSPA v1.0 with Microsoft and Alibaba while releasing open-source MetaRoCE implementations. Standardizing switching layers, scale-up fabrics, and RDMA transport protocols shifts the industry away from proprietary lock-in. Open Ethernet architectures now prove capable of driving both scale-up and scale-across AI workloads sustainably.