Meta said on 12 October 2026, at the Open Compute Project (OCP) Global Summit, that it is introducing IcePack, its first liquid-cooled network rack, and IceCube, a 102.4-terabit-per-second switch blade that delivers 128×800G of non-blocking switching capacity.
In an Engineering at Meta post, it also described a multiplane GPU fabric called Flattened Planar Fabric, an Ethernet scale-up switch called Tahan for its MTIA 300 accelerators, open-source implementations of its MetaRoCE transport protocol and a new OCP initiative it co-leads with Microsoft and Alibaba.
Business Pill · LATENCY AND THROUGHPUT
A one-minute explainer of latency and throughput: how long one request takes versus how much work a system handles over time. It teaches the general idea only and says nothing about any organisation in this story.
The key insight: As we read it, Meta’s OCP post treats the network as a core part of its AI build-out: it pairs a switch generation Meta says doubles bandwidth with a flatter GPU fabric, open-source software and a switch-platform specification submitted to OCP.
Meta says IcePack is based on OCP’s ORV3-HPRv1 rack standard, with IceCube switch blades inside each rack.
According to the post, IceCube comes in two variants, one built on Broadcom’s Tomahawk6C switch chip and one on NVIDIA’s Spectrum-6. Both run FBOSS, Meta’s open-source network operating system, which uses OCP’s Switch Abstraction Interface.
Meta says the new generation doubles bandwidth over its 51.2T platforms, and that the design is meant to support high power density, deliver higher thermal efficiency and enable gigawatt-scale AI clusters.
The post ties the work to Meta’s own AI workloads, naming its Muse Wonder and Muse Spark models and Muse, its personal AI agent.
Meta’s own figures for IceCube, the switch blade inside its first liquid-cooled network rack, IcePack, from its 12 October 2026 Engineering at Meta post: 102.4 terabits per second with 128x800G of non-blocking switching, double the bandwidth of Meta’s 51.2T platforms, in variants built on Broadcom’s Tomahawk6C and NVIDIA’s Spectrum-6.
A Flatter Fabric for GPUs
Meta calls Flattened Planar Fabric (FPF) its multiplane backend scale-out fabric. In it, each GPU connects through eight independent 100G links to eight different switches, one per network plane.
Meta says FPF moves from the three-tier Clos fabric of its NSF fabric to a two-tier Clos fabric, and that removing a tier of switches reduces network hops, latency and infrastructure cost.
In the post’s description, load balancing now starts at the network card, which sprays every packet across all eight planes in hardware, and the control plane includes both the switches and the hosts, so a host can steer traffic away from a plane that is down.
Meta says it worked closely with NVIDIA on FPF, using NVIDIA Spectrum-X Multiplane technology, and that the two companies are working with joint cloud partners to standardize multiplane fabrics based on what Meta learned.
Clusters That Span Buildings and Regions
Meta writes that it uses leased facilities and public cloud capacity alongside its own campuses, because “construction runs on multi-year lead times while demand moves in months.”
It says cloud partners now run a common suite of point-to-point and collective benchmarks, scaling from tens to over a thousand GPUs, before a cluster is handed over.
The post cites Prometheus at roughly 1 gigawatt in 2026 and Hyperion scaling toward 5 gigawatts, and says a single cluster now outgrows any single region. Its backend aggregation (BAG) layer, an Ethernet super-spine, links many fabrics into one cluster; Meta says inter-BAG capacity between a pair of regions reaches tens of petabits per second.
Meta also says it has deployed Segment Routing over IPv6 (SRv6) on FBOSS for its regional backbones, running on switch platforms powered by Cisco Silicon One.
Ethernet Inside the Rack
Meta introduced Tahan, an Ethernet-based scale-up switch for its MTIA 300 training accelerators. The post says it is built on Broadcom’s 51.2 Tbps Tomahawk 5 chip with 64×800G of capacity, half of it used for scale-up links to the accelerators, and that four Tahan switches interconnect 16 accelerators. Meta says Tahan is shipping today and delivers 6.4 Tbps per accelerator.
It also previewed a next-generation scale-up blade, a 216×400G switch on Broadcom’s 102.4 Tbps Tomahawk6C, with eight of these switches interconnecting 72 accelerators per rack. Meta says a future generation is planned to adopt Ethernet for Scale-Up Networking (ESUN), the OCP initiative Meta says it founded last year and now works on alongside AMD, Arista, ARM, Broadcom, Cisco, HPE, Marvell, Microsoft and NVIDIA.
Open Software and a New OCP Initiative
Meta says it is co-leading a new OCP initiative with Microsoft and Alibaba, the Open Switch Platform Architecture (OSPA), with specification v1.0 submitted to OCP. It is meant to standardize the switch platform layer: power sequencing, thermal and fan control, inventory, out-of-band interfaces, firmware updates and diagnostics.
The post says that layer has remained proprietary and fragmented, and that every new hardware platform requires months of duplicated, custom engineering to bring up.
Meta also says it is announcing open-source implementations of MetaRoCE, its RDMA transport protocol for AI clusters over Ethernet, and that switch vendors can now onboard their own platforms to FBOSS without Meta engineers in the loop.
As we read it, the IcePack and IceCube pitch is about fitting more bandwidth into tighter power and cooling envelopes: Meta says switches must deliver dramatically more bandwidth while fitting within increasingly dense power and cooling envelopes.
As we read it, FPF moves work out of the switches: Meta says load balancing now starts at the network card, and that removing a tier of switches reduces hops, latency and infrastructure cost.
As we read it, the cloud section explains why Meta wants standards: it says it uses leased facilities and public cloud alongside its own campuses, and that standardization is what makes heterogeneity affordable.
Engineering at Meta, Network Infrastructure team, 12 October 2026
“The principle we keep applying is that heterogeneity is manageable but standardization is what makes it affordable.”
Three Implications
SWITCH AND CHIP VENDORS Meta says vendors can now onboard their own switch platforms to FBOSS without Meta engineers in the loop, and that IceCube comes in Broadcom Tomahawk6C and NVIDIA Spectrum-6 variants.
CLOUD PROVIDERS Meta says cloud partners run a common benchmark suite, scaling from tens to over a thousand GPUs, before a cluster is handed over.
STANDARDS GROUPS Meta says OSPA specification v1.0 has been submitted to OCP, and that a future scale-up blade generation is planned to adopt ESUN as its data plane.
The framework puts the pattern this way: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, Meta’s post applies the open half of that pattern to networking: it is releasing FBOSS and MetaRoCE as open source and co-leading OSPA to standardize the switch platform layer, which it says has remained proprietary and fragmented.
Meta says each GPU in its Flattened Planar Fabric connects through eight independent 100G links to eight different switches, one per network plane, and that the fabric moves from a three-tier to a two-tier Clos design. This is Meta’s description from its 12 October 2026 post.
What Is Not Established
Meta’s post gives no cost figures and no count of IcePack racks or IceCube switches deployed, and it does not say whether other buyers can purchase the hardware. The bandwidth, latency and efficiency claims are Meta’s own. The post gives no date for the 72-accelerator scale-up design.
Business Engineer Framework
The Open vs Closed Meta-Framework
A Business Engineer framework on which layers of a technology stack to keep proprietary and which to open.
At OCP, Meta described a liquid-cooled 102.4T switch generation, a flatter eight-plane GPU fabric and Ethernet scale-up switching for its MTIA accelerators, and paired them with open-source software and a new switch-platform standard effort. The figures and claims are Meta’s own, from its 12 October post.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. On 13 October 2026 we read Meta’s Engineering at Meta post of 12 October in full. The specifications, figures and descriptions are Meta’s own; the post gives no cost figures or deployment counts. Nothing here is investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
