Meta opens up new networking designs for AI superclusters

Meta has unveiled new networking designs for AI data centres, including a liquid-cooled switch rack and a fabric intended to connect larger GPU clusters with fewer network hops. The announcement also includes open implementations and a new standards initiative, putting infrastructure choices that are usually buried in the server room onto the industry’s shared agenda.

Meta AI Watch analysis

What happened

At the 2026 Open Compute Project Global Summit, Meta outlined new networking hardware and software for AI workloads. Its IcePack rack houses IceCube switches with a stated capacity of 102.4 terabits per second. Meta says that is twice the bandwidth of its previous 51.2T platforms, with liquid cooling designed for the thermal demands of denser AI data centres.

The company also introduced Flattened Planar Fabric, a network design in which each GPU connects across eight independent network planes. Meta says the approach can support larger non-blocking clusters while reducing network hops, latency and infrastructure cost. Its MetaRoCE AI transport protocol now has open-source implementations, and Meta is co-leading the Open Switch Platform Architecture initiative with Microsoft and Alibaba. Specification version 1.0 has been submitted to OCP.

Our top picks

• Liquid-cooled 102.4T switching
  IcePack and IceCube pair higher stated switching capacity with liquid cooling for dense AI infrastructure.
• A flatter GPU network
  Flattened Planar Fabric uses eight independent network planes per GPU to support larger clusters with fewer network hops.
• Open AI transport software
  Meta says MetaRoCE now has open-source implementations, giving infrastructure teams more room to inspect and adapt the networking stack.
• A shared switch-platform specification
  Meta, Microsoft and Alibaba are co-leading an OCP initiative, with version 1.0 of its specification submitted.
• Ethernet for accelerator scale-up
  Meta describes Tahan, an Ethernet-based switch for its MTIA 300 accelerators, extending Ethernet’s role inside AI systems as well as between them.

Why it matters

AI capacity is not just a matter of buying more accelerators. Moving data between them, managing heat and keeping training jobs running across sprawling sites are also major infrastructure challenges. These designs show Meta pursuing more bandwidth and larger clusters while trying to make networks more open and easier to operate across different settings.

The practical test is whether these designs, and the standards work around them, travel beyond Meta’s own infrastructure. Meta says it is sharing some of its network learnings with cloud partners, but the announcement does not establish how widely the systems will be adopted or provide independent performance comparisons.

Our read

This is a substantial infrastructure announcement, not a new AI model, but it gets at a basic constraint on building bigger AI systems: the network has to keep pace with the chips. The welcome part is the push to share protocols and develop a common switch-platform specification. The less glamorous question, as ever, is whether open designs prove easier for others to deploy than the bespoke alternatives.

What to watch

• Whether other operators adopt the Open Switch Platform Architecture specification.
• How MetaRoCE and Flattened Planar Fabric perform in deployments beyond Meta.
• Whether Meta publishes comparative results on bandwidth, latency, power use and cost.
• How the next-generation Ethernet scale-up design develops.

Discussion spark: Should AI infrastructure move towards shared, open networking standards, or do operators gain more from tightly integrated proprietary systems?

not affiliated with, endorsed by, or operated by Meta