Meta open-sources MetaRoCE RDMA transport and OSPA switch spec at OCP 2026

At the 2026 Open Compute Project Global Summit, Meta disclosed multiple networking technologies powering its gigawatt-scale Prometheus AI cluster and Muse Wonder and Muse Spark models. The company announced it is open-sourcing MetaRoCE, its RDMA transport for AI networks. Meta also co-led submission of the OSPA (Open Switch Platform Architecture) v1.0 specification to OCP alongside Microsoft and Alibaba, aiming to standardize the switch platform layer that remains proprietary despite SAI's data-plane portability. The announcements accompany detailed technical disclosures including IceCube 102.4T switches, Flattened Planar Fabric scale-out architecture, and Tahan Ethernet scale-up switches for MTIA 300.

Meta's networking stack runs on FBOSS, its open network operating system, which powers production AI fabrics in collaboration with NVIDIA's Spectrum-X SDK and SAI implementation. The company released video recordings and slide decks from more than a dozen technical talks covering switch hardware, integrated optics, linear pluggable optics, and operational experience at gigawatt scale.

What Meta open-sourced and standardized at OCP 2026

• MetaRoCE: Meta's RDMA transport for AI networks is now going open source, providing an alternative RDMA implementation for high-performance AI cluster networking.
• OSPA v1.0: A new OCP initiative co-led by Meta, Microsoft, and Alibaba to standardize the switch platform layer, with the v1.0 specification already submitted to OCP for review.
• FBOSS integration: Meta's network operating system runs on IceCube 102.4T switches (Broadcom Tomahawk 6C or NVIDIA Spectrum-6, 128x800GE) and integrates with NVIDIA's Spectrum-X SDK and SAI for production AI fabrics.
• Flatened Planar Fabric: A multi-plane scale-out architecture where each GPU connects to eight network planes, collapsing 3-tier to 2-tier Clos with NIC-level packet spraying and combined host+switch control plane.
• Scale-across backbone: BAG petabit-scale super-spine connects buildings, campuses, and regions using SRv6 regional backbones running on FBOSS plus SAI.

Why it matters

Open-sourcing MetaRoCE gives the Linux AI infrastructure ecosystem a production-proven RDMA transport alternative to proprietary implementations, potentially reducing vendor lock-in for large-scale training clusters. The OSPA standardization effort addresses a genuine friction point where SAI-enabled data-plane portability still leaves every switch platform requiring months of duplicated bring-up work, which if successful would accelerate multi-vendor switch deployment for cloud providers and enterprises building open networks.