Meta says its Prometheus AI supercluster in New Albany, Ohio, now connects hundreds of thousands of GPUs across multiple buildings. The first completed phase was used to train its Muse foundation models, including Muse Wonder. At full scale, Meta plans to run roughly 1 million distributed GPUs together.
Two years ago, the site was a conventional data center handling web traffic. Meta says it has rebuilt the site in phases for large AI training jobs. Prometheus currently spans traditional data center buildings, buildings designed for AI workloads and rapid deployment structures.
Meta is adding newer GPU platforms
Prometheus currently uses two NVIDIA GPU platforms. Meta’s Catalina platform uses GB200 NVL72 hardware, with 72 GPUs connected across two racks. Its newer Clemente platform uses GB300 NVL72 hardware and fits a 72-GPU group into one rack. According to Meta, GB300 increases memory per GPU from 192 GB to 288 GB.
To connect GPUs across the site, Meta uses a dedicated Ethernet network alongside its standard data center network. Each building-level network can connect up to 20,736 GPUs. A separate network links the buildings over distances ranging from 2 km to more than 100 km, the company says.
Meta reports fewer training interruptions
A failure in one part of a large training job can slow or stop the whole run. Meta says each 72-GPU group uses 64 GPUs for training and keeps eight as spares. It also reserves roughly 2% additional racks for rarer failures that take a whole rack offline.
Meta reports more than 85% goodput during Muse Wonder pre-training and more than 90% during its reinforcement learning runs. Goodput measures productive training work. The company says the runs averaged roughly 47,000 GPU hours between interruptions. Later changes across its fleet, including firmware fixes, raised that figure to more than 54,000 GPU hours.
Meta says it has cut interruptions by 99% from a peak of nearly 1,000 events per day for every 8,000 GPUs during Prometheus’s initial phase in mid-2025. It is continuing to expand the cluster for future models from Meta Superintelligence Labs. A third GPU platform based on NVIDIA’s Vera Rubin NVL72 is due soon, according to the company.
