Meta reveals details on Prometheus, its Ohio AI supercluster

Meta has shared fresh details about its Prometheus AI supercluster in New Albany, Ohio, which currently runs hundreds of thousands of Nvidia GPUs across several buildings with plans to expand to about 1 million, according to The Next Web. The company outlined the scale and engineering challenges of the gigawatt-scale facility in a blog post on Monday, showing how a former web traffic site was transformed.

Key takeaways

• The first phase of the Prometheus supercluster is finished and was used to train Meta's Muse model family.
• Prometheus currently runs Nvidia Catalina and Clemente platforms, with a third platform named San Miguel coming.
• Meta cut training interruptions by 99% after engineering work that addressed signal quality and memory faults.
• The facility relies on battery backups and software called Dynamo Dimmer to handle rapid power swings.

What is the Prometheus AI supercluster?

The facility in Ohio sits across traditional data centres, AI-optimised buildings, and rapid-deployment structures linked by miles of fibre. Two years ago, the location was a conventional data centre handling web traffic. The first phase of construction is already complete, and Meta used it to train its Muse family of models, which includes the Muse Wonder models. According to The Next Web, the company stated that at full scale the setup will comprise roughly 1 million distributed GPUs functioning in lockstep.

Hardware inside the site spans three generations of Nvidia racks, each housing 72 GPUs that act as a single accelerator. Catalina uses GB200 NVL72 racks, Clemente uses newer GB300 NVL72 racks, and the upcoming San Miguel platform will bring Nvidia Vera Rubin NVL72 racks. Each building's network connects up to 20,736 GPUs, while a separate aggregation network joins buildings over distances ranging from 2km to more than 100km, supplemented by leased partner capacity.

How does Meta manage Nvidia GPUs and power swings?

Running thousands of chips together created massive reliability hurdles. In mid-2025, interruptions peaked at nearly 1,000 events per 8,000 GPUs every day. Engineers traced early problems to signal quality on more than 600,000 NVLink connections, while memory faults grew as GB300 racks scaled up. A year of fixes cut those disruptions by 99%. Training runs now average over 54,000 GPU hours between interruptions, up from 47,000, while the company eyes a target of 100,000.

Hardware redundancy is part of the solution. Every 72-GPU rack runs a 64-GPU slice and keeps eight chips as spares, with Meta holding about 2% of racks in reserve. When computers fire up together, the cluster experiences power draw swings of 30% to 40%, shifting tens of megawatts in seconds. Meta controls this with software named Dynamo Dimmer, alongside batteries deployed at the rack, facility, and grid levels.

What happens next?

The energy demands of the Ohio site are substantial. Meta pays the full cost of the power it uses in New Albany and matches it with new clean and renewable energy, though it left the RE100 pact in July as gas construction for AI outpaced renewables. Lessons learned from this project are already guiding the next facility, named Hyperion, which is designed in Richland Parish, Louisiana. Meta has also pledged $1bn to communities near its data centres.