The development

What happened

Meta says the completed first phase of Prometheus in New Albany, Ohio, has trained Muse models including Muse Wonder. Its 19 October engineering report describes a system spanning hundreds of thousands of GPUs across connected buildings. Roughly one million GPUs remains the full-scale target.

Key facts

  • Current platforms include NVIDIA GB200 and GB300 NVL72; a Vera Rubin platform is described as forthcoming.

  • Meta reports over 85% productive training time, or goodput, for Muse Wonder pretraining and over 90% for reinforcement learning. These are company-reported measures.

  • Each 72-GPU domain uses 64 active GPUs and eight local spares, with additional racks reserved for wider failures.

  • Isolated data-parallel replicas can recover while other training continues; broader failures use checkpoint restarts. Ethernet fabrics and Backend Aggregation connect the cluster.

The context

Why it matters

The report shows why useful training capacity depends on fault recovery and networking alongside GPU count. Completed first-phase training is operational evidence, while the million-GPU expansion and comparisons with other clusters remain separate questions.