Key points
- Meta released the IcePack liquid-cooled rack and IceCube 102.4T switches to increase bandwidth and thermal efficiency for AI workloads.
- The company launched the Open Switch Platform Architecture initiative with Microsoft and Alibaba to standardize switch hardware layers.
- New networking fabrics include the Flattened Planar Fabric for scale-out and MetaRoCE, an open-source transport protocol for AI networks.
Meta shared new open networking infrastructure at the 2026 Open Compute Project Global Summit, according to Engineering at Meta. The release includes the IcePack liquid-cooled network rack and the IceCube 102.4T switch platform. These components are designed to support high-power-density AI clusters while improving thermal efficiency. The company also announced the Open Switch Platform Architecture initiative and updates to its scale-out and scale-up fabrics.
How it unfolded
- Meta presented the IcePack rack and IceCube switches at the 2026 Open Compute Project Global Summit, marking its first liquid-cooled network rack design.
- The company introduced the Flattened Planar Fabric, a multiplane backend scale-out fabric that changes how GPUs connect to the network for larger clusters.
- Meta, Microsoft, and Alibaba co-launched the Open Switch Platform Architecture initiative, submitting specification version 1.0 to the Open Compute Project community.
Who is affected
System administrators and engineers building large-scale AI infrastructure are the primary audience for these updates. Meta states that the IcePack rack and IceCube switches are designed for gigawatt-scale AI clusters, such as those running the Muse Wonder and Muse Spark models. The 102.4T switches deliver 128 times 800G non-blocking switching capacity, representing a two-fold bandwidth increase over previous 51.2T platforms. This hardware targets operators needing higher thermal density and bandwidth for next-generation data centers.
The new networking fabrics affect how GPU clusters are architected. The Flattened Planar Fabric enables larger, highly efficient GPU clusters with lower latency. Meta shared learnings from this multiplane fabric with cloud partners. Additionally, the scale-across networking updates address multi-site clusters, using petabit-scale backend aggregation super-spines to handle failures and long distances. These systems use SRv6 regional backbones built with FBOSS and SAI.
The Open Switch Platform Architecture initiative impacts silicon vendors, system integrators, and operators. By standardizing the switch platform layer, the initiative aims to create more open and sustainable AI infrastructure. Both variants of the IceCube switch, one using Broadcom’s Tomahawk6C ASIC and the other using NVIDIA Spectrum-6 ASIC, run Meta’s open-source FBOSS network operating system. This leverages the OCP Switch Abstraction Interface for broader compatibility.
What happens next
Meta is inviting engineers, silicon vendors, system integrators, and operators to join the Open Compute Project community. The goal is to meet the challenges of building open, efficient, and sustainable AI infrastructure. The company highlighted its ongoing work in Ethernet for Scale-Up Networking, noting that the MTIA 300 system uses the Tahan 51.2T scale-up switch. Future designs will move toward ESUN-based implementations with next-generation chips.
The open-source MetaRoCE transport protocol is now available for AI networks. This allows developers to integrate RDMA capabilities into their AI networking stacks. Meta continues to push boundaries in data center networking, building on initiatives like Ethernet for Scale-Up Networking introduced at OCP 2025. The focus remains on scaling AI workloads across buildings, campuses, and multiple regions with resilient fabric architectures.
What to do and how to stay safe: Meta
- Review current data center cooling capacities to determine if existing infrastructure can support high-density liquid-cooled network racks like IcePack.
- Evaluate switch ASIC options, comparing Broadcom Tomahawk6C and NVIDIA Spectrum-6 performance metrics against your specific AI workload requirements.
- Test the open-source MetaRoCE transport protocol in non-production environments to assess compatibility with existing RDMA implementations.
- Monitor the Open Compute Project repository for updates to the Open Switch Platform Architecture specification version 1.0 and related FBOSS releases.
Step-by-step guide: Injection attacks examples: how prompt leaks break AI systems
General security guidance from the Circuit Observer newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is the bandwidth capacity of the new IceCube switches?
The IceCube switches provide 102.4 Tbps of bandwidth with 128x800G non-blocking switching capacity, doubling the bandwidth of previous 51.2T platforms.
Which companies are leading the Open Switch Platform Architecture initiative?
Meta, Microsoft, and Alibaba are co-leading the Open Switch Platform Architecture initiative to standardize the switch platform layer within OCP.
Does the IceCube switch support different ASICs?
Yes, IceCube comes in two variants: one built on Broadcom’s Tomahawk6C switch ASIC and another on NVIDIA Spectrum-6 Switch ASIC.
