Meta's summit material is organized around the layer of the stack each effort touches, from the switch and rack upward to the fabric and the clusters that run on it. The switch platform talks cover the physical build; the FBOSS sessions cover the software that drives it.

Switch Hardware and Open Platform Layer

IceCube is Meta's 102.4T switch built for the liquid-cooled Icepack rack. IceCube and N-IceCube share a 2OU, hybrid liquid/air-cooled, PHY-less design and carry either the Broadcom Tomahawk 6C or the NVIDIA Spectrum-6 ASIC, each offering 128x800GE of non-blocking capacity.

The OSPA session starts from a gap left by routing: SAI made data plane software portable, but the switch platform layer stayed proprietary and fragmented, costing months of duplicated bring-up per platform. OSPA proposes a unified, multi-vendor reference design for that layer, and a 1.0 specification has been contributed to OCP.

Drivers and bring-up

Meta's pairing of the NVIDIA Spectrum-X switch, SDK, SAI implementation, and simulator with FBOSS was presented as a template for production AI fabrics, with the CI/CD and test-driven workflow credited for fast joint integration. A separate session walked the FBOSS stack itself — routing, forwarding, platform services, and board support — and argued for a config-driven route from board bring-up to a platform qualified for the OCP ecosystem.

Optics and Short-Reach Interconnect

Integrated optics were framed around rack-scale AI constraints: power, density, disaggregation, and reliability. Meta's validation work on 51.2T co-packaged optics was presented as the basis for what could carry over at 102.4T.

Linear pluggable optics attack a different line item. At AI scale the per-module retimer DSP is a repeated cost in power, heat, and expense across millions of links; removing it pushes equalization and triage onto the host. The talk asks whether linear optics can hold retimed-level reliability at fleet scale.

Short-reach links set the economics, which motivates Meta's large-scale 800Gb/s active electrical cable deployment and the component quality work behind it. A further session examined the requirements for a near-package optics (NPO) ecosystem to address AI scaling needs that pluggable modules cannot meet.

Running Scale-Out and Scale-Across in Production

The operational view covered the full training stack — GPUs and scale-up fabrics, FBOSS, drivers, communication libraries, and the scheduler — as the scope of running Meta's largest AI clusters reliably.

On BAG (Backend Aggregation), presenters covered purpose, architecture, and performance benefits, then drew on deployment in the Prometheus cluster: platform fit, distance management, reliability, and capacity planning under a fast-moving AI ecosystem.

Backbone and hybrid capacity

Meta chose SRv6 for a new backbone layer, running data plane and control plane in production on FBOSS with shallow buffers at WAN distances on a timeline of months from concept to deployment. Complementing on-premises work, another session described how Meta uses cloud GPU capacity and which practices from its own deployments it carries into clouds to make them performant and reliable for AI workloads.

A cross-industry scale-across panel compared what routing and optical transport must supply for scale-across AI clusters, drawing on hyperscaler, silicon, and systems-vendor perspectives.

Scale-Up: Tahan and the Move to ESUN

Meta's rack-scale MTIA 300 AI system is served by Tahan, an open Ethernet-based scale-up switch. The talk on Tahan walks through its bandwidth and latency characteristics and its scale-up feature set, arguing Ethernet's case for scale-up workloads and pointing toward ESUN as the direction for future scale-up designs.

ESUN itself — Ethernet for Scale-up Networking — was launched with the community at last year's OCP Global Summit. Since then the 1.0 specification has been released, and a 1.1 revision is upcoming; the session covers that progress.

Validating the Ecosystem

A separate session proposes a validation methodology for the ESUN ecosystem, placing scale-up alongside the other AI scaling domains: scale in, up, out, across, edge, and above.

Copper Scale-Up as a Statistical Problem

Copper scale-up domains complicate testing. With thousands of connections per rack, even a 99.9% yield leaves roughly 10 bad links, and blind-mate cartridges offer no per-link serviceability. Meta treats this as a statistical problem rather than a collection of isolated debugging cases, and the talk presents the resulting test methodology for multi-accelerator copper scale-up interconnect.

RDMA: MetaRoCE

The network now sits in the critical path of AI clusters: all-reduce and all-to-all performance is gated by the slowest node across hundreds of thousands of accelerators. MetaRoCE is Meta's answer at the transport layer, and the summit covered its MetaRoCE transport specification.

On the implementation side, Meta is open-sourcing its MetaRoCE software, beginning with a user-space version on GitHub, libsoftmetaroce. A DPDK-based version is coming soon.

Where to Find the Talks

All recorded talks from the event are available on the 2026 OCP Global Summit website.