At the 2026 Open Compute Project Global Summit, Meta disclosed a portfolio of open networking innovations designed to support large-scale AI workloads. Key disclosures include a liquid-cooled network rack named IcePack and 102.4-terabit per second (Tbps) switches called IceCube, along with a multiplane backend fabric and open transport protocols for AI networks.
Meta described these components as part of a broader strategy to enable scale-up, scale-out, and scale-across architectures that can span buildings, campuses, and regions.
ICEPACK AND ICECUBE: THE FIRST LIQUID-COOLED RACK AND HIGH-DENSITY SWITCHES
IcePack is Meta’s first liquid-cooled network rack, based on the Open Compute Project (OCP) ORV3-HPRv1 rack standard. Each IcePack rack houses IceCube switch blades that deliver 102.4 Tbps of fabric capacity and sit inside the rack, with variants built on Broadcom’s Tomahawk6C and NVIDIA Spectrum-6 Switch ASICs. TechStaged has also covered Microsoft Expands Azure AI Infrastructure With AMD for Inference and HPC.
IceCube provides 128×800G non-blocking switching capacity and runs Meta’s open-source FBOSS network operating system, leveraging OCP SAI for switch abstraction.
This generation reportedly represents a 2× bandwidth increase over prior platforms while targeting the thermal density required for next-generation AI data centers.
FLATTENED PLANAR FABRIC AND MULTIPLANE SCALABILITY
Meta introduced Flattened Planar Fabric (FPF), a multiplane backend scale-out fabric that fundamentally changes how GPUs connect to the network and how the network control plane operates. In FPF, each GPU connects via eight independent 100G links to eight different switches, one per network plane, enabling larger non-blocking GPU clusters and reducing network hops compared with traditional three-tier designs.
The design supports NIC-level load balancing and a unified control plane across switches and hosts for fault tolerance. Meta collaborated with NVIDIA on FPF, leveraging NVIDIA Spectrum-X Multiplane innovations and sharing learnings with cloud partners to standardize multiplane fabrics.
OPEN SWITCH PLATFORM ARCHITECTURE (OSPA) AND OPEN ECOSYSTEM
Meta also announced the Open Switch Platform Architecture (OSPA) within OCP, an initiative co-led by Meta, Microsoft, and Alibaba to standardize the switch platform layer. OSPA specification version 1.0 has been submitted to OCP.
The move aims to streamline how different vendors’ switches and software interoperate within open-networking stacks used for AI workloads.
ETHERNET AT SCALE-UP: TAHAN AND THE PATH TO ESUN
A defining milestone is Meta’s assertion that Ethernet can serve scale-up interconnects between accelerators within a rack. The Tahan switch, built on Broadcom’s Tomahawk 5 ASIC, offers 51.2 Tbps of backplane bandwidth and 64×800G capacity, connecting 16 accelerators in a non-blocking arrangement.
Meta plans to extend this with a next-generation scale-up blade delivering 216×400G on a 102.4 Tbps silicon, interconnecting 72 accelerators per rack. A future generation aims to adopt OCP’s Ethernet for Scale-Up Networking (ESUN) to unify the scale-up data plane across hardware generations.
The Ethernet-based RoCE transport is maintained to unify scale-up and scale-out networking within the FBOSS-managed stack.
SCALE-ACROSS, REGIONAL BACKBONES, AND SRV6
Meta described Scale-Across for multi-site clusters, including a petabit-scale BAG (backend aggregation) super-spine that interconnects spine-layer fabrics across buildings, campuses, and regions. The BAG layer operates at petabit-scale capacity between regions and enables a single cluster view across disparate facilities.
To manage long distances and heterogeneity, Meta uses Segment Routing over IPv6 (SRv6) with FBOSS for regional backbone networks, enabling flexible traffic engineering as AI clusters span geographies. The approach runs on open, shallow-buffer switch platforms powered by Cisco SiliconOne, with ongoing work to onboard additional silicon families under FBOSS using OCP SAI.
Meta also highlighted that large-scale training requires resilience to failures, with automatic replica exclusion and continued synchronization across the BAG when failures occur.
RELATED COVERAGE
- Microsoft Expands Azure AI Infrastructure With AMD for Inference and HPC
- Cloudflare reports 74% fastest-provider share across top networks in 2026 Birthday Week performance update
- GitLab Dependency Firewall blocks risky packages before the build (early access)
- Cloudflare closes Birthday Week 2026 with 46 launches, expanding agentic tools and post-quantum roadmap
- Developer Tools articles
