Meta's Prometheus supercluster hits gigawatt scale on the road to 1 million GPUs

Meta's GB{2,3}00 Cluster

Meta has completed the first phase of Prometheus, a gigawatt-scale supercluster in New Albany, Ohio, that trained its Muse Wonder frontier models. The cluster runs hundreds of thousands of GPUs across multiple buildings as a single machine, with GB200 and GB300 NVL72 racks and a dedicated Ethernet back-end network spanning up to 100 km. Training runs hit over 85% goodput for pre-training and 90% for reinforcement learning, with mean time between interruptions above 54,000 GPU hours. At full scale it will reach roughly 1 million GPUs.