Meta introduces liquid-cooled switches and open networks for larger AI clusters

Meta’s latest AI networking package tackles two limits at once: moving more data and cooling the equipment that carries it. At the Open Compute Project Global Summit on October 12, 2026, Meta introduced liquid-cooled switching hardware and a new GPU connection architecture, alongside open-source transport implementations and a joint effort to standardize switch management.

More bandwidth inside a liquid-cooled rack

IcePack is Meta’s first liquid-cooled network rack, built on OCP’s ORV3-HPRv1 rack standard. It houses IceCube switch blades, each with 102.4 terabits per second of switching capacity across 128 connections rated at 800 gigabits per second. The non-blocking design lets those connections operate at their rated capacity simultaneously.

The blades come in two versions: one uses Broadcom’s Tomahawk6C switch chip, the other NVIDIA’s Spectrum-6. Both run FBOSS, Meta’s open-source network operating system, through OCP’s Switch Abstraction Interface. Meta says the liquid-cooled generation is designed for the power and thermal density of gigawatt-scale AI clusters.

Eight paths out of each GPU

The new Flattened Planar Fabric changes how GPUs enter the network. Each GPU connects through eight independent 100-gigabit links to eight different switches, one for each separate network plane. Rather than funneling traffic into a single plane, the network interface card distributes individual packets across all eight in hardware.

Meta says this architecture supports larger non-blocking GPU clusters and removes a switch tier, moving from three tiers to two. Fewer stages mean fewer network hops, which the company says reduces latency and infrastructure cost. Hosts also participate in network control, allowing them to steer traffic away from failed planes.

Meta developed the fabric with NVIDIA and is sharing its lessons with joint cloud partners. That work serves a practical capacity problem: Meta says data-center construction takes years, while demand moves in months. It supplements its own campuses with leased facilities and public cloud capacity, where network designs and transport behavior differ.

To make those environments behave more consistently, cloud partners run a common performance-validation suite before handing over clusters. Meta says the tests cover direct transfers and coordinated GPU communication, scaling from tens to more than a thousand GPUs. Its MetaRoCE AI transport protocol also now has open-source implementations.

Ethernet within racks, common management across vendors

For connections among accelerators within a rack or pod—called scale-up networking—Meta introduced Tahan. The Ethernet switch uses Broadcom’s 51.2-Tbps Tomahawk 5 chip for Meta’s MTIA 300 accelerators. Four switches interconnect 16 accelerators in a non-blocking layout, using the same lossless RoCE transport as Meta’s larger scale-out network, with both managed under FBOSS.

Meta also previewed a denser design: eight switches built on 102.4-Tbps silicon connecting 72 accelerators per rack. A future generation is planned to adopt OCP’s Ethernet for Scale-Up Networking data plane. That standards transition is a plan, not a completed feature of the design introduced today.

The Open Switch Platform Architecture, or OSPA, addresses a different source of friction. Initiated by Meta, Microsoft and Alibaba, it defines a vendor-neutral architecture for switch management. OCP’s October 8 announcement identifies the revision 1.0.0 candidate as under Foundation review, rather than an approved specification.

OCP says functions such as identifying devices, controlling temperature and updating firmware are still rebuilt for each hardware variant. OSPA aims to make that software reusable across vendors. Its scope includes power, hardware access and inventory, environmental control, firmware recovery, debugging and security—not routing protocols or packet-forwarding management. The effort will proceed through open meetings and public review.