Inside Prometheus: Meta trained Muse Wonder on phase one of its gigawatt Ohio supercluster, now scaling towards ~1M GPUs

What's new? Meta has published a detailed look at Prometheus, its gigawatt-scale AI supercluster in New Albany, Ohio. The first phase is complete and was used to train the Muse family of foundation models, including the Muse Wonder frontier models. Today it spans hundreds of thousands of GPUs across multiple interconnected buildings, from traditional and AI-optimised data centres to rapid-deployment structures, all working as a single machine. At full scale Meta wants it to reach roughly 1 million GPUs. Two years ago the site was a conventional data centre serving web traffic.

The hardware: Prometheus runs two NVIDIA platforms: Catalina, built on GB200 NVL72 with 800G of network bandwidth per GPU, and Clemente, built on GB300 NVL72 with 288 GB of memory per GPU in a single rack. A third, San Miguel, based on NVIDIA's Vera Rubin NVL72 with fully liquid-cooled racks of over 170 kW, is arriving next. Each 72-GPU domain trains on a 64-GPU slice and keeps eight GPUs as in-rack spares, so a failing host can be swapped out without halting the job.

Keeping it running: At this scale one underperforming host can slow an entire training run. Meta says it reached more than 85% goodput on Muse Wonder pre-training and more than 90% on reinforcement learning, with about 47K GPU hours of mean time between interruptions, since raised to more than 54K hours through firmware fixes. A power-management system called Dynamo Dimmer smooths the 30-40% load swings that occur when tens of thousands of GPUs compute and communicate in lockstep.

Why it matters: Frontier AI is increasingly a contest of infrastructure engineering as much as model research. Prometheus is the blueprint for Hyperion, Meta's next supercluster in Richland Parish, Louisiana, which is being designed with closed-loop liquid cooling and a full grid connection through utility Entergy, paid for by Meta. Meta says it is scaling Prometheus with the next frontier models from Meta Superintelligence LabsDealroom has a profile for this one. Try Dealroom → in mind.

Sources: Engineering at Meta, AI at Meta on X.