Meta Details Prometheus Supercluster With Path to 1 Million GPUs

Meta says the first phase of its Prometheus AI supercluster in New Albany, Ohio, is complete, with plans to expand the system to about 1 million GPUs. The company detailed the system in an engineering post, saying it currently links hundreds of thousands of GPUs across multiple buildings.

The completed phase trained the Muse Wonder frontier models. Meta recorded more than 85% useful compute time during pretraining and more than 90% during reinforcement learning. Mean time between interruptions reached about 47,000 GPU hours during those runs and now exceeds 54,000 GPU hours after firmware fixes.

Prometheus currently uses NVIDIA GB200 NVL72 and GB300 NVL72 systems. The GB300 configuration places 72 GPUs in one rack, provides 20.7 TB of HBM3e memory, and delivers a 5% to 10% performance increase over the GB200 configuration. Meta plans to add NVIDIA Vera Rubin NVL72 systems, which it expects to improve performance per watt by 10% to 20% over GB300 systems.

The cluster uses dedicated Ethernet networks for communication between GPUs and closed loop liquid cooling for high power racks. Meta says it reduced daily interruptions by 99% from a peak of nearly 1,000 events per 8,000 GPUs in mid 2025 through firmware updates, hardware screening, spare GPUs, checkpointing, and automated fault recovery.