Meta Details Prometheus Gigawatt AI Supercluster Scaling Toward 1M GPUs in Ohio

Meta has completed the first phase of Prometheus, a gigawatt-scale AI supercluster in New Albany, Ohio, and is working toward scaling the system to roughly 1M distributed GPUs operating in lockstep. The cluster currently runs hundreds of thousands of GPUs across multiple buildings connected by a Backend Aggregation network spanning 2 km to over 100 km, and was used to train the company's new Muse foundational models. Prometheus deploys two Nvidia platforms—Catalina (two-rack GB200 NVL72 with 13.4 TB of HBM3e) and Clemente (single-rack GB300 NVL72 with 20.7 TB of HBM3e and over 1 exaFLOP of peak FP4 compute)—with a third platform based on Nvidia's Vera Rubin NVL72 arriving soon. Because moving from NVL8 to NVL72 increased the fault blast radius by 9x, Meta runs a 64-GPU slice per 72-GPU rack and holds the remaining 8 GPUs as immediate in-rack spares alongside roughly 2% reserve racks. This requires 37,656 provisioned GPUs (about 13% over-provisioning) to sustain a 32,768 active-GPU run. During Muse Wonder training, Meta achieved over 85% goodput for pre-training, over 90% goodput for reinforcement learning, and an initial wall-clock mean time between interruptions of roughly 47k GPU hours, which subsequent firmware and fleet fixes raised past 54k GPU hours.