Meta has launched the first phase of its Prometheus AI supercluster in New Albany, Ohio. The facility, which became operational in October 2026, is designed to reach a total capacity of 1 GW. Currently, the cluster utilizes hundreds of thousands of GPUs, with plans to scale toward one million units.
The primary function of Prometheus is training the Muse family of foundational models, including Muse Wonder. Engineering details were presented at the OCP Global Summit on October 12, 2026. The infrastructure relies on the Backend Aggregation network using Ethernet and a closed-loop liquid cooling system to manage heat.
Performance metrics indicate over 85% goodput during pre-training and over 90% during reinforcement learning. The mean time between interruptions is approximately 47,000 GPU hours. Meta has invested over $7.8 million in local grants and sponsorships in Franklin and Licking counties since 2017.
Lessons from the Prometheus project will inform the development of Hyperion, a future supercluster in Louisiana. Meta aims for Hyperion to reach a capacity of up to 5 GW. The company shared its engineering approach at an open hardware summit to encourage industry study of its networking standards.
