Meta Engineering Opens 102.4t AI Networking Stack at OCP

Executive Summary

• At the 2026 Open Compute Project Global Summit, Meta Engineering detailed a next-generation open networking stack for AI data centers, spanning scale-up, scale-out, and scale-across fabrics (Meta Engineering).
• The company introduced IcePack, its first liquid-cooled network rack, and IceCube, a 102.4-Tbps switch blade delivering 128×800G non-blocking capacity in two variants built on Broadcom's Tomahawk6C and NVIDIA's Spectrum-6 switch ASICs (Meta Engineering).
• Flattened Planar Fabric (FPF), a multiplane backend scale-out fabric, connects each GPU through eight independent 100G links to eight separate switches and moves to a two-tier Clos topology (Meta Engineering).
• Meta Engineering launched the Open Switch Platform Architecture (OSPA) initiative within OCP, co-led with Microsoft and Alibaba, with specification v1.0 submitted to OCP (Meta Engineering).
• Open-source implementations of MetaRoCE, the company's RDMA transport protocol for AI networks over Ethernet, were announced at the summit (Meta Engineering).

Key Takeaways

• Meta Engineering's 102.4T generation doubles bandwidth over its 51.2T platforms while adding liquid cooling, a response to power and thermal limits at gigawatt-scale AI clusters.
• FPF shifts load balancing to the NIC and merges switch and host control planes, aiming at larger non-blocking GPU clusters with fewer network tiers.
• The scale-across story is about distance: backend aggregation interconnects regional fabrics at petabit scale, and FBOSS now measures cable length to allocate buffer headroom.
• Tahan, a 51.2-Tbps Ethernet scale-up switch for MTIA 300 accelerators, plus a previewed 72-accelerator-per-rack design, test whether Ethernet can displace proprietary scale-up interconnects.

Meta Engineering Pushes 102.4T Switching Into Liquid-Cooled Racks

The headline hardware development is IcePack, described by Meta Engineering as its first-generation liquid-cooled network rack, based on OCP's ORV3-HPRv1 rack standard. Inside each rack sit IceCube switch blades: 102.4-Tbps fabric switches with 128×800G non-blocking switching capacity. Meta Engineering says IceCube ships in two variants, one on Broadcom's Tomahawk6C switch ASIC and one on NVIDIA's Spectrum-6 Switch ASIC, both running the open-source FBOSS network operating system with OCP's Switch Abstraction Interface.

The company frames the generation as a 2x bandwidth increase over its 51.2T platforms, achieved alongside the thermal density it says next-generation AI data centers require. The stated design goal is gigawatt-scale AI clusters, where Meta Engineering argues switches must deliver more bandwidth inside increasingly dense power and cooling envelopes. That is a vendor claim about intended capability rather than a measured deployment result. What is verifiable from the source is the specification set, the silicon choices, and the open software stack underneath.

Meta Engineering Rewrites Scale-Out With Flattened Planar Fabric

Flattened Planar Fabric is the architectural change with the widest implications. Under FPF, each GPU connects via eight independent 100G links to eight different switches, one per network plane. Meta Engineering says this replaces a single-plane uplink architecture and yields four benefits: larger non-blocking GPU clusters, a two-tier Clos fabric instead of the three-tier design in its NSF fabric, NIC-level hardware load balancing that sprays packets across all eight planes, and a combined control plane spanning switches and hosts so traffic can be steered away from downed planes.

Removing a switch tier is the cost argument: fewer hops, lower latency, less infrastructure. The load-balancing shift is the operational argument, moving a function traditionally handled by switches down to the NIC. Meta Engineering states it collaborated closely with NVIDIA on FPF inside its own infrastructure, drawing on NVIDIA Spectrum-X Multiplane work, and that the two companies are working with joint cloud partners to standardize multiplane fabric implementations based on those learnings. Those partner implementations are described as work in progress, not shipped results.

Meta Engineering Confronts Distance, Buffer Headroom, and Multi-Region Training

The scale-across section is the most technically specific part of the disclosure. Meta Engineering's backend aggregation layer is an Ethernet-based super-spine that interconnects many spine-layer fabrics into one cluster spanning buildings, campuses, and regions. It says inter-BAG capacity between a pair of regions reaches tens of petabits per second, and that BAG is what lets Prometheus, described as roughly 1 gigawatt in 2026, operate as a single cluster, with Hyperion scaling toward 5 gigawatts.

Distance becomes a first-order constraint. Lossless Ethernet requires buffer headroom proportional to cable length, so Meta Engineering says FBOSS now measures actual cable length in place by timing probe traffic on an uncongested queue and allocates headroom against the measurement rather than provisioning for worst-case distance. BAG uses eBGP link-bandwidth signaling to distribute traffic according to available capacity, with deliberate damping to limit routing churn after a single link failure. Because inter-fabric latency runs several times intra-fabric latency, collective communication libraries that select algorithms without awareness of network distance can hit a performance cliff; Meta Engineering says it is addressing this with adaptive routing and scheduling that reduces how much collective traffic crosses BAG. It also describes SRv6 regional backbones running on FBOSS over open shallow-buffer platforms powered by Cisco SiliconOne.

Related: Freecash Data Harvesting App Banned After App Store Fraud 2026

Meta Engineering Tests Ethernet in the Scale-Up Domain

Scale-up, the high-bandwidth, ultra-low-latency interconnect between accelerators inside a rack or pod, has traditionally required proprietary technologies. Meta Engineering introduced Tahan, an Ethernet-based scale-up switch built on Broadcom's 51.2-Tbps Tomahawk 5 ASIC with 64x800G total capacity, half used for scale-up connectivity to accelerators via the backplane. Four Tahan switches interconnect 16 accelerators in a non-blocking topology. Notably, Tahan reuses the same lossless RoCE transport as the scale-out network, routed IPv6, PFC and AI-tuned QoS, unifying both fabrics under FBOSS.

A next-generation scale-up blade is previewed at 216×400G on 102.4-Tbps Tomahawk6C silicon, with eight switches interconnecting 72 accelerators per rack. Meta Engineering says a future generation is planned to adopt OCP Ethernet for Scale-Up Networking as its data plane. The ESUN initiative the company founded within OCP last year, working with AMD, Arista, ARM, Broadcom, Cisco, HPE, Marvell, Microsoft, and NVIDIA, has released the ESUN 1.0 Network Operator Requirements Base Specification and is close to ESUN 1.1, which targets switch profiles and inter-packet gap reduction. The 72-accelerator design and ESUN 1.1 are roadmap items, not shipping products.

Meta Engineering Targets the Switch Platform Onboarding Tax

OSPA addresses a layer beneath SAI: the switch platform layer, which Meta Engineering describes as proprietary and fragmented. The company says every new hardware platform requires months of duplicated custom engineering to bring up, and that the lack of standard platform design multiplies long-term maintenance and troubleshooting cost. OSPA standardizes hardware and software interfaces for power management and sequencing, sensor and environment control including thermal and fan management, hardware inventory and device enumeration, out-of-band interfaces such as management Ethernet, serial console, USB and LED indicators, firmware maintenance, and platform diagnostics. Specification v1.0 has been submitted to OCP.

For deeper context, see our Quantum AI analysis: "Quantum AI Power Map Redrawn as AWS, IBM and IonQ Spark Deals and Hybrid Stack Race".

Meta Engineering also extended FBOSS so switch vendors can self-onboard platforms without Meta engineers in the loop, aiming at a self-service model where multiple new platforms develop concurrently. Alongside this, it announced open-source MetaRoCE implementations, describing the RDMA transport as optimized for AI cluster networks over Ethernet with the congestion control and reliability semantics distributed training demands. The company positions its cloud convergence work around a standardized performance validation framework, with cloud partners running common point-to-point and collective benchmarks scaling from tens to over a thousand GPUs before a cluster is handed over.

What This Means for Practitioners

For infrastructure buyers and operators, the practical signal is that Meta Engineering is pushing open specifications into layers that have been vendor-specific, including the switch platform itself and the scale-up interconnect. Operators weighing OSPA-aligned hardware, FBOSS onboarding, or an Ethernet scale-up path should treat Meta Engineering's benefits as design intent, not measured outcomes: liquid cooling, FPF, Tahan, and the 72-accelerator rack are described at specification or preview stage. The near-term procurement question is whether silicon and system vendors converge on the same interfaces. Evidence to watch is ESUN 1.1 and partner multiplane deployments.

Meta Engineering Signals Reference

Geographic detail is not specified in the supplied source; no location rows have been inferred.

Additional coverage: CodeWords $9M Seed Round 2026: How Cody AI Targets No-Code Workflow Gap

Meta Engineering Implementation Risks

Three risks follow from the disclosure itself. First, heterogeneity: Meta Engineering states that fabric topologies differ by provider, buffering is often shallower, RDMA transports are provider-specific, and some clouds still run InfiniBand rather than Ethernet and RoCE, so every divergence becomes engineering work. Second, distance: lossless Ethernet buffer headroom scales with cable length, and Meta Engineering warns that getting buffer management wrong risks congestion spreading well beyond the link that caused it. Third, roadmap dependence: ESUN 1.1, the 72-accelerator-per-rack blade, and cloud-partner multiplane standardization are described as in progress, so outcomes remain unproven.

Editorial independence disclosure: this article was prepared independently from the cited source and reflects only details present in that source. Source note: all facts are drawn from Meta Engineering's OCP Summit 2026 networking post (Meta Engineering).