Meta has switched on the first phase of Prometheus, its AI supercluster at the company’s data center campus in New Albany, Ohio. The facility became operational in October 2026 and is built to run at gigawatt scale.
What Meta actually built
Prometheus is currently packed with hundreds of thousands of GPUs. Meta’s stated goal is to scale toward approximately one million GPUs and a total capacity of 1 GW.
The cluster’s main job is training Meta’s Muse family of foundational models. That includes the Muse Wonder models, which Meta says currently deliver industry-leading performance on both personal and professional tasks.
Meta walked through the engineering behind the project at the OCP Global Summit on October 12, 2026. The company’s official engineering blog post on Prometheus was updated on October 19, 2026.
Two pieces of infrastructure carry most of the weight. The first is the Backend Aggregation network, or BAG, which links GPUs spread across multiple buildings using high-capacity Ethernet connections.
The second piece is closed-loop liquid cooling. Rather than relying purely on air, these systems circulate coolant through a sealed loop to pull heat away from hardware, with the aim of improving efficiency and reliability.
The numbers Meta is bragging about
Meta reports more than 85% goodput during pre-training runs and more than 90% during reinforcement learning runs.
Goodput is the share of compute time that actually advances training, as opposed to time lost to crashes, restarts, or idle hardware.
Meta also reports a mean time between interruptions of about 47,000 GPU hours. That figure tracks how much cumulative chip time passes, on average, before something breaks a training job.
A long relationship with central Ohio
Prometheus did not appear out of nowhere. Meta began its New Albany project in 2017, and the AI cluster is the latest chapter on a campus the company has been building for years.
Since then, Meta says it has put more than $7.8 million into local grants and sponsorships across Franklin and Licking counties. The project also supports hundreds of jobs in the area.
What this means for Meta and the AI race
Meta expects the lessons from Prometheus to shape Hyperion, a follow-on supercluster in Louisiana that is expected to target up to 5 GW of capacity.
Meta’s choice of Ethernet for the BAG network is also worth noting. Building multi-building connectivity on a widely used networking standard, and presenting the work at an open hardware summit, signals that Meta wants its approach to be studied and potentially adopted beyond its own walls.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.