Meta unveils liquid-cooled network rack and 102.4T switch at OCP 2026

At the Open Compute Project Global Summit 2026, Meta presented a package of open networking technologies for AI clusters. It includes IcePack, the company's first liquid-cooled network rack; IceCube, a 102.4T switch platform; Flattened Planar Fabric (FPF), a multi-plane backend fabric; and MetaRoCE, an open transport protocol. Meta pitches all of it as infrastructure for three levels of scaling: scale-up, scale-out and scale-across.

The announcement, published by Engineering at Meta, lists the components and what each is for. The portion available to us does not disclose detailed specifications, so below we separate what Meta has claimed from what is worth checking in the full documentation.

Three network layers for one cluster

The summit framing describes three directions in which a training cluster's network grows. This is standard industry terminology.

• Scale-up connects accelerators within a single domain: a node or a rack. This layer needs maximum bandwidth and minimum latency.
• Scale-out joins racks and domains into one cluster through the backend network.
• Scale-across links sites and data centers, for cases where a single training run does not fit into one building because of power or floor space.

Meta presents the announcement as a package covering all three layers. Hardware, topology and transport are shown together, which sets it apart from a standalone switch release.

Hardware: IceCube and IcePack

IceCube is a switch platform with 102.4T of bandwidth. That figure conventionally denotes aggregate throughput in terabits per second. For scale, you could reach that capacity with, for example, 64 ports at 1.6T each, but Meta does not name the specific port configuration in the available text.

IcePack is the company's first liquid-cooled network rack. Liquid cooling has already become the norm for compute racks with accelerators, and now it has reached networking gear as well. The reason is clear: switches with this much bandwidth density, plus the optics next to them, give off a lot of heat, and air cooling is coping less and less well.

Fabric and protocol: FPF and MetaRoCE

Flattened Planar Fabric is a multi-plane backend fabric for scale-out. In Meta's words, it "fundamentally changes how GPUs connect to the network." In multi-plane topologies, each accelerator gets several independent network paths, or planes. That provides more parallel routes and makes it easier to isolate failures. The word "flattened" in the name points to a reduction in the number of switching tiers. How exactly the GPU connection works is only visible from the full article.

MetaRoCE is an open transport protocol for AI traffic. The name refers to RoCE, that is, RDMA over Ethernet. Meta says it is open source, and that matters more than the other items: the transport determines how a cluster behaves under congestion and packet loss, and closed implementations usually tie an operator to a single vendor.

Key points of the announcement

• IcePack is Meta's first liquid-cooled network rack.
• IceCube is a 102.4T switch platform.
• FPF is a multi-plane backend fabric for scale-out that changes how GPUs connect to the network.
• MetaRoCE is an open transport protocol for AI workloads.
• The venue for the announcement was the Open Compute Project Global Summit 2026.

What to check next

For cluster operators and network engineers, the main task now is to wait for the materials Meta publishes through OCP. Four things are needed:

1. IceCube specifications: number and speed of ports, power consumption, type of optics. 2. IcePack's liquid-loop requirements: whether it is compatible with compute racks that are already built. 3. FPF topology: number of planes, number of tiers and the estimated cluster size. 4. The license and repository for MetaRoCE, as well as its compatibility with existing network cards and with other open initiatives, above all Ultra Ethernet.

If the documentation confirms the claims, companies outside the hyperscalers will get a ready-made reference design for an AI network that does not have to be built from scratch around a single vendor. Until the details are published, this remains a claim, not a verified architecture.