• Sources: primary
  • Summary: Prometheus in New Albany, Ohio, spans hundreds of thousands of GB200 and GB300 NVL72 GPUs and targets about 1 million. Meta reports over 85% goodput for Muse Wonder pretraining, over 90% for RL, and mean time between interruptions raised from about 47K to over 54K GPU hours after firmware fixes. A 32,768-GPU run needs 523 racks, about 13% above the active footprint, with 11% coming from the 64-of-72 slice geometry.
  • Why it matters: The spare-slice arithmetic is a concrete capacity-planning model for anyone sizing jobs on 72-GPU failure domains.