Until about two years ago, Meta’s campus in New Albany, Ohio, served ordinary website traffic. It now houses Prometheus, the cluster on which Meta trained its Muse models, according to a 19 October post on Engineering at Meta, the company’s technical blog.
The first phase of Prometheus is finished and produced the Muse family, including Muse Wonder. Meta calls these its new frontier models and says they deliver “state-of-the-art performance on personal and professional agentic tasks.” That is Meta’s own characterization of its own models. The post includes no independent benchmark results.
Meta calls the project a “gigawatt-scale” supercluster and says it is working toward roughly one million GPUs. Neither phrase is a measurement of today. The post says the site currently holds hundreds of thousands of GPUs spread over several linked buildings, and gives no exact count, no actual power draw, no cost, and no completion date for the larger target.
The most useful number in the post is about breakage. Training a large model means thousands of machines working in lockstep, so one faulty machine can slow or halt the whole job. Meta tracks “goodput,” the share of the cluster’s time that produces useful training. It reports goodput above 85 percent on the pre-training runs of Muse Wonder and above 90 percent on its reinforcement learning runs. The post does not say over what period those averages were taken.
The interruption data is more concrete. In mid-2025, Meta says, failures peaked at nearly 1,000 events per 8,000 GPUs per day. A year of fixes cut that by 99 percent. During the Muse Wonder runs, the cluster averaged about 47,000 GPU hours between interruptions, and firmware fixes since then have pushed that above 54,000. Meta’s target is two interruptions per 8,000 GPUs per day, which still works out to a stop roughly every 12 hours for a cluster that size.
The hardware explains why failures are so costly. Prometheus runs Nvidia’s GB200 and GB300 systems, which tie 72 GPUs into one unit, and Nvidia’s Vera Rubin generation is due next. Under the previous design that unit held eight GPUs, so a single fault now disrupts nine times as many chips. Meta runs 64 GPUs of each 72 and keeps eight as spares. For a 32,768-GPU job it provisions 523 racks, or 37,656 GPUs, about 13 percent more than the job uses.
Scale creates other problems. Each building’s network fabric links up to 20,736 GPUs, and the buildings sit between 2 km and more than 100 km apart. Power is its own hazard: when thousands of GPUs start or pause together, load swings of 30 to 40 percent can move tens of megawatts in seconds. Meta’s power software lowers GPU limits gradually near the electrical ceiling instead of stopping jobs.
On cooling, the post says Meta’s newest AI data centres use closed-loop liquid cooling. Prometheus itself mixes air-assisted liquid cooling in some buildings with facility liquid cooling in its purpose-built AI building. The post says liquid, not water specifically, and it states that Meta pays the full cost of its energy in New Albany. Those are the company’s claims, with no outside check attached.
Stability may matter more than raw chip speed here. Every hour a stalled job sits idle is an hour of an enormous machine producing nothing, and Meta says the cluster now suffers 99 percent fewer interruptions than at its mid-2025 peak. Meta says its next site, Hyperion in Richland Parish, Louisiana, is being designed with these lessons in mind. Anyone budgeting for large training runs should treat the 54,000-hour figure as the number to watch, since Meta says it is still moving.
Engineering at Meta, the Meta engineering blog, post on the Prometheus supercluster, published 19 October 2026.
