Meta says the first phase of Prometheus, its gigawatt-scale AI supercluster in New Albany, Ohio, is complete and has been used to train its latest foundation models, including the Muse family. The site was a conventional data centre running web traffic about two years ago and was reworked for frontier-model training. Every figure here is Meta's own and has not been independently checked.
Scale and hardware
Prometheus currently spans multiple interconnected buildings holding hundreds of thousands of GPUs, and Meta says it will reach roughly one million GPUs working in lockstep at full scale. It lists three Nvidia rack platforms: a Blackwell design with 72 GPUs per rack pair drawing about 120 kilowatts, a Blackwell Ultra design with 72 GPUs in a single rack drawing about 135 kilowatts and offering a 5 to 10 per cent performance gain, and an upcoming Vera Rubin design drawing over 170 kilowatts per rack with a 10 to 20 per cent performance-per-watt gain.
Power, cooling and networking
Synchronised training can swing the cluster's electrical load by 30 to 40 per cent, and starting or pausing large jobs can shift tens of megawatts within seconds. Meta uses a cluster-wide power management system that gradually lowers GPU power limits near the electrical ceiling instead of stopping jobs, backed by rack-level batteries, capacitors, facility-level battery storage and coordination with utilities. Cooling combines air-assisted liquid cooling in free-air and rapid-deployment buildings with facility liquid cooling in a purpose-built AI building. A dedicated Ethernet network links GPUs, with each building fabric connecting up to 20,736 GPUs and a hub-and-spoke network linking buildings across distances from 2 kilometres to more than 100.
Reliability
Meta reports goodput above 85 per cent for pre-training runs and above 90 per cent for reinforcement-learning runs, a mean time between interruptions of about 47,000 GPU hours that later improved to more than 54,000 with firmware fixes, and a 99 per cent reduction in interruptions since a mid-2025 peak of nearly 1,000 per 8,000 GPUs a day. Its stated target is two interruptions per 8,000 GPUs a day, roughly 100,000 GPU hours between interruptions.
