Meta Completes First Phase of Prometheus Gigawatt-Scale AI Supercluster

Prometheus Supercluster Development

Meta has completed the first phase of Prometheus, its gigawatt-scale AI supercluster located in New Albany, Ohio. This project represents Meta's most significant AI infrastructure undertaking to date, transforming a conventional data center site into a facility for training frontier AI models.

The supercluster is being built in phases, with the initial phase already operational. This completed phase was utilized for training Meta's Muse foundational models, including the Muse Wonder models, which are designed for agentic tasks.

Infrastructure and Scale

Prometheus currently consists of hundreds of thousands of GPUs spread across multiple interconnected buildings, including traditional and AI-optimized data centers. These components are linked by extensive fiber optic networks, functioning as a single, unified machine.

Meta plans to scale Prometheus to include approximately one million GPUs. The infrastructure incorporates technologies such as closed-loop liquid cooling to enhance operational efficiency.

Performance and Reliability

The supercluster is engineered for high reliability and stability to sustain performance during large training jobs. Meta achieved over 85% 'goodput' for pre-training runs and more than 90% 'goodput' for reinforcement learning runs of Muse Wonder.

During these training runs, the overall wall-clock mean time between interruptions (MTBI) was approximately 47,000 GPU hours, which has since been increased to over 54,000 GPU hours through firmware fixes and ongoing improvements.

Future Capabilities

The capacity provided by Prometheus enables Meta to pursue the development of larger AI models and longer context windows. The company is continuing to scale the supercluster with future frontier models from Meta Superintelligence Labs in mind.