Meta details Prometheus supercluster, scaling toward 1 million GPUs

Meta on October 19, 2026 published technical details of Prometheus, its gigawatt-scale AI supercluster in New Albany, Ohio, which currently links hundreds of thousands of graphics processors across multiple buildings to operate as a single large computer, according to aitimes.com. The company said the site is designed to reach roughly 1 million GPUs at final scale.

Meta said its recently released Muse Wonder model was trained at Prometheus. The cluster runs on Nvidia's GB200 and GB300 systems, with 72 GPUs per GB300 rack, and tens of thousands of network switches and millions of optical components tie the processors together.

To limit downtime, Meta said it worked with Nvidia for about a year to cut training interruptions by 99%. The site keeps spare GPUs available as failover capacity and uses liquid cooling.

Meta said it has completed the first phase of the Prometheus build-out and plans to expand the cluster with Nvidia's next-generation Vera Rubin systems. The company intends to apply what it learns at Prometheus to Hyperion, another supercluster planned in Louisiana.

From web services to AI training

Two years before the disclosure, the New Albany data center handled ordinary web services before being converted into a large-scale AI training site. On October 12, 2026, Meta announced Muse Wonder and Muse Wonder Full, saying the models reached state-of-the-art results on key personal, work-agent, memory and coding benchmarks.