Inside Prometheus: Meta’s Journey to a Gigawatt-Scale Supercluster

Why it matters

Meta has completed the first phase of Prometheus, its gigawatt-scale supercluster in New Albany, Ohio, and is now scaling the site toward roughly a million GPUs. The cluster, which began as a conventional web data center, was reworked over two years into a training facility for frontier models, and its first phase produced the Muse foundational models, including the frontier Muse Wonder models designed for personal and professional agentic tasks. Meta emphasizes reliability as the central engineering challenge: a fault on a single host can stall an entire training job, degrading goodput. The company reports strong goodput for both pre-training and reinforcement learning runs, along with a mean time between interruptions it has since improved through fleet-wide firmware fixes. The site currently runs two NVIDIA GPU platforms, with a third generation arriving soon, and uses closed-loop liquid cooling. Meta says the scale enables larger models and longer context windows than previously practical, and it is building with future frontier models from Meta Superintelligence Labs in mind.