Meta said in an engineering post on 19 October 2026 that it is building Prometheus, an AI supercluster in New Albany, Ohio. The first phase is done and trained Meta's Muse models, including Muse Wonder. Meta said Prometheus now has hundreds of thousands of GPUs, the chips that train AI models, and will hold roughly 1 million at full size.
Meta said interruptions peaked near 1,000 events per 8K GPUs a day in mid-2025, then fell 99% in a year. It credits tracing every interruption source and fixing each failure type, starting with weak NVLink links among over 600,000. Newer racks link 72 GPUs, not 8, so one fault reaches nine times as many chips. Eight GPUs per rack are spares, and one faulty model copy can restart alone; otherwise the job rolls back to a saved checkpoint. These limit the cost of faults that remain. The post gives no date for full size.
The post suggests that at this size, how a cluster limits and recovers from faults matters as much as GPU count. Meta's figures rest on its own definition, hardware and spares, so compare like with like. GPU cloud buyers can ask providers how they define an interruption and how they use spares.