Meta’s engineering team on October 19, 2026 published a detailed technical account of Prometheus, its gigawatt-scale AI supercluster in New Albany, Ohio. According to Unite.AI’s summary of the post, the completed first phase trained the Muse family of models, including the Muse Wonder frontier models, and Meta plans to scale the system to roughly 1 million GPUs. The figures below are Meta’s own and have not been independently audited.

What Prometheus is

The post, credited to Kevin Lee, Yashar Bayani, Ahmed Sharif and Yee Jiun Song, says the New Albany site was a conventional data center running web traffic two years before the disclosure. Meta reworked it for frontier-model training and is building the supercluster in phases. Prometheus now spans hundreds of thousands of GPUs across traditional and AI-optimized buildings and rapid-deployment structures, all operating as a single machine.

Meta reports more than 85% goodput on Muse Wonder pre-training runs and more than 90% on reinforcement learning runs. Goodput is the share of time a cluster spends making useful training progress rather than recovering from failures. The company also cites a mean time between interruptions of about 47,000 GPU hours during training, raised to more than 54,000 after fleet-wide changes including firmware fixes.

The hardware inside

Two Nvidia platforms are in production, and a third is coming. The first, called Catalina, uses the GB200 NVL72 in a two-rack variant, linking 72 Blackwell GPUs over NVLink. Meta lists 13.4 TB of HBM3e memory and roughly 120 kW of power for the rack pair. The second, Clemente, moves to the GB300 NVL72, with 288 GB of memory per GPU, about 135 kW per rack and a stated 5–10% performance gain over the earlier platform. A third system based on Vera Rubin NVL72, code-named San Miguel, is listed with 288 GB of HBM4 per GPU and over 170 kW per rack, and Meta claims a 10–20% performance-per-watt improvement over GB300.

Networking is a large part of the story. Each building-level Ethernet fabric connects up to 20,736 GPUs, and a separate aggregation layer links buildings across distances from 2 km to more than 100 km. Meta says training across those distances required dedicated optimization work because of speed-of-light limits.

Reliability and power

Larger tightly coupled GPU groups mean larger failure domains. Moving from eight-GPU to 72-GPU NVLink domains made the blast radius nine times bigger, according to the post. Meta runs a 64-GPU slice in each 72-GPU domain and keeps eight as spares. During the initial phase in mid-2025, interruptions peaked at nearly 1,000 events per 8,000 GPUs per day; Meta reports a 99% reduction after a year of engineering, with a target of two per 8,000 GPUs per day.

Power is the other constraint. Synchronized training causes power swings of 30–40% within the cluster, and starting or pausing a major job can shift loads by tens of megawatts in seconds. Meta describes a framework called Dynamo Dimmer that gradually lowers GPU power limits near the electrical ceiling instead of halting jobs, backed by batteries and coordination with utilities. Cooling combines air-assisted liquid cooling with facility liquid cooling; for background see our explainer on liquid cooling in AI data centers. Meta says it pays the full cost of energy at the site and matches usage with new clean and renewable energy.

What comes next

Meta’s next supercluster, Hyperion in Richland Parish, Louisiana, is being designed with lessons from Prometheus, including closed-loop liquid cooling and grid connectivity with utility partner Entergy. For the financial scale behind such projects, see AI data center capex explained.

What to watch next

  • Whether Meta reaches its reliability target of two interruptions per 8,000 GPUs per day.
  • The first production results from Vera Rubin-based racks.
  • Progress on Hyperion in Louisiana and local power arrangements.
  • Independent evaluation of Muse Wonder models trained on the cluster.

Key takeaways

  • Prometheus in New Albany, Ohio, trained Meta’s Muse models and spans hundreds of thousands of GPUs.
  • Meta reports over 85% goodput in pre-training and over 90% in reinforcement learning.
  • Interruptions fell 99% from their mid-2025 peak, by Meta’s account.
  • Power swings of 30–40% during training are managed with software power limits and batteries.
  • The stated plan is to scale toward roughly 1 million GPUs.

Frequently asked questions

What is goodput?

Goodput measures the fraction of time a training run makes useful progress, excluding time lost to failures, restarts and waiting.

Why do bigger GPU domains make failures worse?

When 72 GPUs act as one accelerator, a fault in any part can affect the whole group, so a failure touches more hardware than in eight-GPU designs.

Are these numbers verified?

No. They come from Meta’s own engineering post as relayed by Unite.AI, and we have not seen independent confirmation.