Meta Unveils Next-Gen Open AI Networking at OCP Global Summit 2026
New liquid-cooled 102.4T switching, multiplane fabrics, and scale-up Ethernet drive gigawatt clusters.
1Next-Gen AI Networking Roadmap
At OCP Summit 2026, Meta introduced its open network architectures tailored for gigawatt-scale AI workloads like Muse Wonder and Muse Spark. The releases span scale-up, scale-out, and multi-region scale-across domains, demonstrating end-to-end Ethernet integration across training clusters. Meta shared blueprints for liquid-cooled hardware, unified fabrics, and protocol specifications with open industry partners.
2102.4T Liquid-Cooled Switching
Meta debuted IcePack, its first liquid-cooled network rack built on OCP ORV3-HPRv1 standards, housing IceCube 102.4-Tbps switch blades. Delivering 128x800G non-blocking capacity, IceCube doubles the bandwidth of 51.2T platforms. Available in Broadcom Tomahawk6C and NVIDIA Spectrum-6 variants, both configurations operate Meta’s FBOSS network operating system over Switch Abstraction Interface.
3Flattened Planar Fabric Architecture
Meta introduced Flattened Planar Fabric, shifting scale-out backend topologies from 3-tier to 2-tier Clos networks. Each GPU connects through eight independent 100G links directly to eight isolated switch planes. This multiplane design reduces network hops, lowers latency, and cuts overall infrastructure costs across ultra-large training clusters.
4NIC-Level Hardware Packet Spraying
Instead of relying on network switches for traditional load balancing, FPF initiates packet spraying directly at the host network interface card. The NIC distributes every packet across all eight planes in hardware. The combined host-and-switch control plane allows accelerators to dynamically route traffic away from downed network planes.
5Public Cloud Fabric Normalization
Because facility construction requires multi-year lead times, Meta leverages leased data centers and public cloud capacity to meet immediate compute demands. Meta standardizes heterogeneous multiplane fabrics by deploying a shared validation benchmark suite. The framework maps varying traffic classes and congestion-control parameters directly onto lossless RDMA classes across vendors.
6Petabit Backend Aggregation Super-Spine
To support clusters like Prometheus at 1 gigawatt and Hyperion approaching 5 gigawatts, Meta deploys Backend Aggregation super-spines. BAG interconnects multiple building-scale spine fabrics with tens of petabits per second inter-regional capacity. This enables geographically dispersed facilities, outdoor structures, and colocation sites to operate as unified mega-clusters.
7Dynamic Buffer and Link Management
Connecting regions across tens of kilometers requires active buffer headroom management for lossless Ethernet. FBOSS actively measures physical cable lengths using uncongested probe traffic instead of provisioning for worst-case distances. BAG also utilizes eBGP link-bandwidth signaling with deliberate damping, distributing traffic proportionately across degraded paths without routing churn.
8SRv6 Regional Backbone Deployment
Meta deployed Segment Routing over IPv6 across its regional backbone networks using FBOSS to simplify operations across geographically distributed AI clusters. Operating on open shallow-buffer switch platforms powered by Cisco SiliconOne, the architecture leverages OCP SAI to integrate additional silicon families into standard traffic engineering workflows.
9Tahan Ethernet Scale-Up Switch
Meta introduced Tahan, an open Ethernet scale-up switch purpose-built for MTIA 300 AI accelerators. Powered by Broadcom’s 51.2 Tbps Tomahawk 5 ASIC, four Tahan switches connect 16 accelerators in a non-blocking topology. Tahan utilizes routed IPv6, PFC, and tuned QoS under FBOSS, unifying scale-up and scale-out management.
10Dense 72-Accelerator Rack Preview
Meta previewed its next-generation scale-up architecture featuring eight 216x400G switch blades powered by 102.4 Tbps Broadcom Tomahawk6C silicon. This hardware interconnects 72 accelerators per rack, drastically expanding compute density. Future revisions are slated to adopt the OCP Ethernet for Scale-Up Networking data plane standard.
11Advancing Open ESUN Standards
Co-founded by Meta alongside AMD, Broadcom, Cisco, Microsoft, and NVIDIA, the ESUN initiative focuses on scale-up fabric specifications. Following the release of ESUN 1.0, the group is finalizing ESUN 1.1, which addresses switch profiles and inter-packet gap reduction while coordinating directly with IEEE and Ultra Ethernet Consortium efforts.
12Unified Open AI Infrastructure
Alongside open-source MetaRoCE implementations, Meta, Microsoft, and Alibaba submitted specification v1.0 of the Open Switch Platform Architecture to OCP. By standardizing platform layers, scale-up interconnects, and cross-region aggregation, Meta proves open Ethernet can eliminate proprietary networking bottlenecks across multi-gigawatt artificial intelligence clusters.