Meta details Prometheus, its gigawatt-scale Ohio training supercluster: phase one built and trained Muse Wonder at 85%+ pre-training and 90%+ RL goodput, interruptions down 99%, scaling toward ~1 million GPUs with Vera Rubin next

Why it matters

【Key points】

Engineering at Meta published "Inside Prometheus" at 21:00 CST on Oct 19, the first detailed look at its gigawatt-scale supercluster in New Albany, Ohio. Phase one is complete and trained the Muse models, including Muse Wonder (e/833). It now spans hundreds of thousands of GPUs across several buildings, with a target of about 1 million.

Self-reported numbers: Muse Wonder goodput above 85% for pre-training and 90% for RL; mean time between interruptions about 47K GPU-hours during those runs, now above 54K after firmware fixes, with a ~100K goal. Interruptions peaked near 1,000 per 8K GPUs per day in mid-2025 and are down 99%. Platforms: GB200 NVL72 (Catalina) and GB300 NVL72 (Clemente), with Vera Rubin NVL72 (San Miguel, 2.52 EFLOPS FP4 per rack, 170kW+) coming.

Details: each 72-GPU rack runs 64 GPUs with 8 in-rack spares, so a 32K-GPU job needs ~13% headroom. Building fabrics reach 20,736 GPUs; aggregation spans 2 to 100+ km. Training causes 30–40% power swings, handled by Dynamo Dimmer power capping, rack batteries and storage. Hyperion in Louisiana will build on this design.

【Verifiable】

Meta engineering blog: https://engineering.fb.com/2026/10/19/data-center-engineering/inside-prometheus-meta-gigawatt-scale-supercluster/

Muse Wonder (e/833): https://agipulse.org/e/833

【Our read】

1. Frontier labs rarely publish cluster reliability data; interruption rates, goodput and spare ratios say more about usable compute than GPU counts.

2. NVL72 widens the failure blast radius 9x; Meta's fix is reserving ~11% of GPUs, a hidden cost of big racks.

3. It builds the case for MSL's next models but gives no model timeline.