Meta’s infrastructure chief Santosh Janardhan described an Ethernet “super spine” and giant tent-like structures used for Prometheus at the Open Compute Project Global Summit in San Jose on October 12, 2026. The network connects a sprawling AI campus, while the structures shorten construction from years to months, he said, according to Linda Hardesty’s reporting for Fierce Network.
Prometheus is Meta’s AI cluster in New Albany, Ohio. In its February 9 engineering account, Meta said the cluster would deliver 1 gigawatt of capacity once complete. Janardhan’s presentation put its physical scale at more than 30 structures, tens of thousands of GPUs, and roughly 3–4 million kilometers of network fiber, Hardesty reported.
The interesting engineering problem is how those separate buildings work together. Packing a campus with processors is one challenge; keeping them communicating efficiently during AI training is another.
The network behind Prometheus’s Ethernet super spine
Janardhan explained that AI training advances at the pace of the slowest communication between GPUs. His description of the super spine was straightforward: connect the network spines—the high-capacity connections within the buildings—and aggregate them locally.
Meta’s engineering account gives that idea more detail through backend aggregation, or BAG. BAG is a network layer that links regional network fabrics, the interconnected switches carrying traffic inside the cluster, and connects them with Meta’s backbone network. It lets the architecture extend beyond an individual building or fabric.
Meta describes inter-BAG capacity of 16–48 petabits per second per pair of regions. That is a capacity range for connections between aggregation layers, rather than a per-GPU transfer rate. The distinction matters: a huge aggregate network still has to distribute traffic across the paths connecting individual workloads.
Simpler management or more paths around failures
Meta describes two approaches to connecting BAG layers: planar and spread connections.
Planar connections simplify management, but concentrate potential failure domains—the portions of a network that could be affected by the same fault. Spread connections distribute links to provide greater path diversity and resilience.
For an operator, those are different priorities. A simpler arrangement is easier to manage; a more distributed arrangement offers additional routes through the network. Meta’s explanation makes the super spine more than a story about raw bandwidth: how the links are arranged also shapes how the network handles failures.
Oversubscription, routing, and link security
Meta gives a typical L2-to-BAG oversubscription ratio of about 4.5:1. Oversubscription means the combined capacity feeding a network layer exceeds the capacity available onward. Here, L2 refers to a level in Meta’s fabric hierarchy.
That ratio describes a capacity tradeoff: all downstream connections cannot send at their full combined rate through the aggregation layer simultaneously. Meta’s separately illustrated non-scheduled fabric, or NSF, arrangement has an effective ratio of 4.98:1.
The BAG hardware description includes modular chassis with Jericho3 ASIC line cards—cards containing specialized networking chips—with up to 432 × 800G ports per line card. For routing, Meta uses external Border Gateway Protocol, or eBGP, with bandwidth attributes and unequal-cost multipath routing, or UCMP. These mechanisms let routing account for paths with different available bandwidths.
Meta also specifies MACsec, a link-layer security technology, on BAG-to-BAG connections. The architecture therefore covers aggregation, routing, and protected links as well as the headline capacity figures.
The hardware Meta described in its earlier OCP presentation
An earlier technical presentation, published by Open Compute Project on October 22, 2025, offers a closer look at the platforms Meta described for Prometheus. Meta technical program managers Maggie Han and Abhinav Triguna covered accelerator racks, storage, general-purpose computing, and networking.
Han described Clemente as Meta’s platform using NVIDIA GB300 accelerators, with 72 GPUs in a GPU rack. Catalina used NVIDIA GB200 accelerators, with two GPU racks containing 36 GPUs each and connected in a 36×2 configuration. These were the platform configurations presented in that earlier talk.
The same presentation showed how the supporting systems fit around the accelerators. Han identified Grand Canyon for bulk storage on hard drives, and Olympic and Yosemite V3.5 Crater Lake for flash-based storage. General-purpose computing platforms included Yosemite V3.5 Halfdome and Yosemite V4. The design encompassed more than GPU racks: it also separated storage and compute into named platform families.
Triguna described two network-fabric approaches: disaggregated scheduled fabrics, or DSF, based on Broadcom Jericho 3 AI and Ramone 3 switches, and non-scheduled fabrics using 400 GbE Minipack 3 switches. He also discussed the need for spare-hardware strategies at this scale and flexibility to support multiple accelerator generations. Those operational concerns help explain why a large cluster needs an adaptable network rather than a layout built around one accelerator platform.
Giant tents and AI-assisted concrete
The construction disclosures addressed a different bottleneck: getting the physical infrastructure built. According to Hardesty’s summit account, Janardhan said Prometheus’s tent-like structures are at least two football fields long, take months rather than years to build, and are rated for a Category 1 hurricane.
He also described Meta’s AI-assisted concrete work, saying a resulting formulation used 19% less carbon, cured 43% faster, and added no cost. Those were his summit claims about the formulation.
How BOxCrete proposes a mix
Meta’s March 30 account of its concrete work explains the process behind Bayesian Optimization for Concrete, or BOxCrete. The system uses historical data and constraints to propose concrete mixes, then incorporates laboratory results into subsequent iterations.
Think of it as a proposal-and-feedback loop: the model suggests a mix, laboratory testing supplies results, and those results inform the next proposal. Meta says laboratory validation, field trials, engineering sign-off, and compliance with building codes remain necessary. The software participates in formulation; engineers and physical testing remain part of the construction process.
Meta also reported a site-specific result at a data center in Rosemount, Minnesota. A mix used for a building slab reached full structural strength 43% faster than the original formula and reduced cracking risk by nearly 10%, according to the company. That Rosemount result concerns time to structural strength and cracking risk, while Janardhan’s summit figures concerned curing and carbon in a separate formulation claim.
Hyperion’s separate Louisiana plan
Janardhan also previewed Hyperion, a separate planned Meta cluster in Richland Parish, Louisiana. Hardesty reported his description of a 5 GW project covering about 4,000 acres.
The two projects have different locations and capacity plans:
| Project | Location | Stated capacity |
| Prometheus | New Albany, Ohio | 1 GW once complete, according to Meta |
| Hyperion | Richland Parish, Louisiana | 5 GW planned, as described by Santosh Janardhan at the summit |
Janardhan characterized Hyperion as the biggest data center on the planet to the best of his knowledge. His preview placed Meta’s Louisiana project on a roughly 4,000-acre site.
