Meta and AMD unveil AI networking hardware at OCP Summit

Meta and AMD unveiled AI networking hardware at the 2026 Open Compute Project Global Summit, held October 12–15, 2026 at the San Jose Convention Center. Meta shared its first liquid-cooled network rack, IcePack, and 102.4T switch blades called IceCube. AMD revealed Helion, a rack-scale AI system for agent workloads. Both companies presented designs meant to help operators build larger, denser AI clusters within tighter power and cooling limits.engineering.fb.com+1

The announcements matter because AI clusters now reach gigawatt scale, forcing switches to deliver far more bandwidth inside dense power and cooling envelopes. Meta's IcePack rack and IceCube blades answer that with liquid cooling and a 2x bandwidth increase over its 51.2T platforms. AMD's Helion adds a rack blueprint that mixes CPU-only and CPU/GPU configurations, giving operators a choice of how to run agent execution and inference within one rack.theregister.com

Meta's IcePack rack is built on OCP's ORV3-HPRv1 rack standard and holds IceCube 102.4T switch blades delivering 128×800G non-blocking switching capacity. IceCube ships in two variants: one on Broadcom's Tomahawk6C switch ASIC and one on NVIDIA's Spectrum-6 Switch ASIC. Both run Meta's open-source FBOSS network operating system, which uses OCP's Switch Abstraction Interface. Meta says this generation delivers 2x the bandwidth of its 51.2T platforms while meeting the thermal density next-generation AI data centers require.

Meta's Flattened Planar Fabric changes how GPUs reach the network. Each GPU connects via eight independent 100G links to eight different switches, one per network plane, replacing the traditional single-plane uplink. The design flattens the topology from a three-tier Clos fabric to a two-tier Clos, cutting a tier of switches to reduce network hops, latency, and infrastructure cost. Load balancing starts at the network interface card, spraying every packet across all eight planes in hardware rather than relying on switches.

The fabric's control plane spans both switches and hosts, letting a host steer traffic away from downed planes for fault tolerance. Meta collaborated with NVIDIA on developing Flattened Planar Fabric inside its infrastructure, using NVIDIA Spectrum-X Multiplane innovations to keep performance high under failure conditions. Meta and NVIDIA are also working with joint cloud partners to standardise multiplane fabric implementations. Meta says the fabric enables larger non-blocking GPU clusters than single-plane topologies, supporting its largest AI training workloads.

AMD's Helion rack, revealed at the summit in San Jose, offers an OCP-compliant blueprint for agentic AI infrastructure built around AMD CPUs and accelerators. Customers can choose a CPU-only hyperscale configuration or an enterprise configuration mixing CPU and GPU blades. AMD advertises up to 32,768 Zen 6 CPU cores via compute nodes with 256-core Venice Epyc processors, up to 16 Instinct MI350P GPU accelerators, and up to 102 Tbps of networking bandwidth. Dense deployments use EPYC 9006 chips rated at 400 W; performance blades use a 96-core version clocked at 5 GHz.

AMD SVP for Compute and Enterprise AI Dan McNamara said agentic AI is not a single workload, spanning enterprise applications and complex workflows with different compute requirements. He said AMD Helion brings Epyc performance, open software, and flexible rack designs together so partners can build systems around customer needs and use available power efficiently. AMD plans to contribute the Helion specifications to OCP for community development and adoption, with partner-built configurations scheduled for November 2026. AMD cites Cisco, Dell, HPE, Foxconn, Lenovo, and Supermicro as on board.

Wiwynn Corporation presented NVIDIA Vera Rubin NVL72 and AMD Helios rack platforms at Booth C74 of the San Jose Convention Center on October 12, 2026. Wiwynn and Wistron are among the first NVIDIA Pilot Build Partners for Vera Rubin NVL72, a liquid-cooled system integrating 72 NVIDIA Rubin GPUs and 36 NVIDIA Vera CPUs linked by sixth-generation NVLink. The AMD Helios rackscale solution, built on the Open Rack Wide standard, unifies Instinct MI450 Series GPUs, sixth-generation EPYC CPUs, Pensando networking, and ROCm software.

Wiwynn also demonstrated cooling and optical interconnect work. A double-sided cold plate with a cavity PCBA layout delivers 6kW of cooling capacity, a thermal performance gain of more than 15%, and a power delivery loss reduction of up to 80%. A lightweight cold plate and inner manifold design cuts cooling component weight by at least 30%. Wiwynn is presenting a co-packaged optics scale-up rack solution with Ayar Labs, Corning, GUC, and Senko, including a live demonstration of optical engine telemetry.

Meta also shared Ethernet for Scale-Up Networking details. Its MTIA 300 system uses Tahan, an Ethernet-based 51.2T scale-up switch delivering 6.4 Tbps per accelerator, with a preview of a next-generation 72-accelerator-per-rack design on 102.4T silicon. Meta launched the Open Switch Platform Architecture initiative within OCP, co-led with Microsoft and Alibaba, standardising the switch platform layer; specification v1.0 is now submitted to OCP. Meta's MetaRoCE RDMA transport protocol now has open-source implementations. The 2026 OCP Global Summit runs October 12–15, 2026 at the San Jose Convention Center. On October 13, 2026, Wiwynn Senior Director Lentis Pai gives an Expo Hall Talk titled Powering AI Infrastructure: Innovations in Compute, Power, Connectivity and Cooling. AMD's partner-built Helion configurations are scheduled for November 2026.