Meta Turned an Ohio Web Data Center Into a Gigawatt AI Supercluster, and Its First Phase Already Trained Muse
Synopsis
Meta converted a conventional data center in New Albany, Ohio, into Prometheus, a gigawatt-scale supercluster whose first phase is complete and was used to train the Muse family of foundation models. Because one failing host can slow an entire large training job, Meta focused on system stability and reached more than 85% goodput on Muse Wonder pre-training runs and more than 90% on reinforcement learning runs. The cluster now spans hundreds of thousands of GPUs, with a target of roughly one million at full scale.
Interpretation
Prometheus's first phase is built and training, and it supported Meta's Muse foundation models, including Muse Wonder, described as new frontier models delivering state-of-the-art performance on personal and professional agentic tasks. Two years ago New Albany was a conventional data center site running web traffic; hardware, networking and building-design advances turned it into a home for frontier-model training. The completed first phase was used to train the Muse foundation models, which is the comparison that makes the buildout concrete.
A large training job's output depends on the whole system working together, since a problem on a single host can slow or stop the entire job and lower goodput, the share of compute that produces useful training progress. Meta concentrated on system stability and reliability rather than raw compute alone. Muse Wonder pre-training runs exceeded 85% goodput and reinforcement learning runs exceeded 90%, with an overall wall-clock mean time between interruptions of about 47,000 GPU hours.
The cluster is organized around NVL72 racks, where 72 GPUs act as a single accelerator inside one network fabric, and a dedicated back-end Ethernet network carries direct GPU-to-GPU traffic across buildings. Traditional web-service networks move requests and responses between instances that do not need high bandwidth or tight lock-step coordination, while large-scale AI training needs both. Two NVIDIA platforms, GB200 and GB300, are deployed with a third based on Vera Rubin NVL72 arriving soon, and full scale targets roughly one million GPUs running in lockstep.
Perspective
The report covers Meta's public account of building and operating Prometheus: the first phase is complete and trained the Muse foundation models, the cluster currently comprises hundreds of thousands of GPUs, full scale targets roughly one million, and future capacity will mix GB300 and Vera Rubin NVL72. The intended reader is a researcher or engineer concerned with large-scale training infrastructure. Meta says it will keep expanding capacity with NVIDIA and that its next supercluster, Hyperion, is being designed with Prometheus's lessons in mind.
The goodput, MTBI and interruption-reduction figures are Meta's own operational data and are not independently verified. The 99% reduction in interruptions refers to the first phase from mid-2025, and later firmware fixes raised MTBI above 54,000 GPU hours, but the stated reliability target of two interruptions per 8K GPUs per day has not yet been reached.