Meta Reveals Prometheus: 1M-GPU AI Supercluster
Meta detailed Prometheus, its gigawatt AI supercluster in Ohio with hundreds of thousands of GPUs, scaling toward 1 million to train its Muse models.
Meta has opened up about Prometheus, the giant AI supercomputer that powers its newest AI models. In a detailed engineering post published on October 19, 2026, the company described the gigawatt-scale supercluster in New Albany, Ohio, which currently runs hundreds of thousands of GPUs and is built to scale to about 1 million. It is the machine that trained Meta's Muse family of AI models, including the Muse Wonder models launched last week.
What Prometheus is
Prometheus is not a single building. It is several buildings working as one machine. Just two years ago, the New Albany site was a normal data center handling regular web traffic. Meta rebuilt it step by step: web servers were swapped for GPU racks, new halls were added as hardware arrived, and the company also leased extra capacity from partners to bring more chips online sooner. Today the campus mixes older buildings, a purpose-built AI building, and rapid-deployment structures, all joined by miles of fiber-optic cable. The network links buildings from 2 km to more than 100 km apart, so Meta had to tune its training software around the speed of light itself.
The GPU racks inside
The cluster runs on Nvidia's NVL72 rack design, in which 72 GPUs share one high-speed NVLink network and act as a single giant accelerator. Meta has built three generations of it: Catalina (based on the Nvidia GB200, already running), Clemente (based on the Nvidia GB300, already running), and San Miguel (based on the upcoming Vera Rubin platform, arriving soon). At full scale, the cluster will hold roughly 1 million distributed GPUs, all working in lockstep.
How reliable it is
Training a huge AI model can be stopped by a single broken chip, so reliability is the real engineering story. Meta reports that more than 85% of compute time actually moved training forward during Muse Wonder's pre-training, and more than 90% during reinforcement learning. The mean time between interruptions started at about 47,000 GPU hours and was pushed past 54,000 GPU hours through firmware fixes and other fleet-wide changes. Interruptions peaked near 1,000 a day per 8,000 GPUs in mid-2025; Meta says it has cut them by 99%.
What comes next
The first phase of Prometheus is finished, and Meta Superintelligence Labs is now using the growing cluster for its next frontier models, which need longer context windows and larger parameter counts than earlier hardware allowed. The company counts millions of miles of fiber strand, tens of thousands of switches, and millions of optical parts in the cluster. Prometheus is Meta's answer to a simple fact of the AI era: frontier models are won or lost in the data center.