🔥Meta Finishes Phase One of 1-Gigawatt Prometheus AI Supercluster in Ohio

Meta's training cluster just got a lot more interesting

Meta just finished phase one of Prometheus, a gigawatt-scale supercluster in New Albany, Ohio. It's already running, and the target is roughly 1 million GPUs working in lockstep. Why should you care? Because the design choices are the interesting part. Meta built the network to match the hierarchy of large training jobs and to minimize exposure of any communication, so GPUs spend less time waiting. If you're planning multi-building training runs, this is a preview of how the hyperscalers are thinking about interconnect and fault tolerance. The hardware is two NVIDIA platforms now, with a third coming. Catalina uses GB200 NVL72, where the Ariel variant pairs each Blackwell GPU with a Grace CPU and 800G of I/O per GPU. Clemente moves to GB300 NVL72, lifting HBM per GPU from 192 GB to 288 GB and doubling softmax attention acceleration. Meta targets over 85% goodput for pre-training and over 90% for RL runs, with MTBI climbing from about 47K to more than 54K GPU hours after firmware fixes. Watch the goodput numbers. Those are targets, not measurements, and the full scale is still ahead.