Meta opens up about Prometheus, its gigawatt-scale AI cluster
Meta says the first phase of Prometheus, its Ohio supercluster, is finished and it is working to grow it to about 1 million GPUs. The Muse model family was trained there.
What happened
On 19 October Meta published an engineering post on Prometheus, its cluster in New Albany, Ohio, by engineers Kevin Lee, Yashar Bayani, Ahmed Sharif and Yee Jiun Song. The first phase is finished and trained the Muse model family, including Muse Wonder. Meta is now working to scale it to about 1 million GPUs running in lockstep. The post came two days before Meta's third-quarter results on 21 October.
The numbers
Prometheus runs hundreds of thousands of GPUs across several interconnected buildings, some of them rapid-deployment structures. Reported goodput is above 85% on pre-training runs and above 90% on reinforcement learning. Mean time between interruptions was about 47,000 GPU hours, rising above 54,000 after firmware fixes. Two Nvidia NVL72 generations are in production, starting with Catalina, built from GB200 NVL72 racks on Meta's Open Rack V3 design; Vera Rubin NVL72 is coming soon.
Speed
About two years ago the campus was a conventional web-traffic data centre. Meta added capacity in months using tent-like rapid structures rather than waiting years for traditional construction.
Why reliability matters
When hundreds of thousands of GPUs run one job, a single failure can stall it. Every point of goodput saved is worth millions of dollars in compute. Highlighting efficiency before earnings looks like an effort to show investors the spending is paying off.
The energy angle
A gigawatt cluster draws power close to a mid-sized city. Meta is relying on nuclear agreements running to 2035. Energy has become the scarcest input; grid connection delays can hold projects back for years, as Nscale's UK site shows.
Who else is building
Meta is not alone. OpenAI, Microsoft, Google, Amazon and xAI are all building or leasing gigawatt-class sites, and SemiAnalysis forecast the same week that roughly 50 GW of new data centre capacity will come online worldwide in 2027. What Meta's post adds is unusual operational detail: few companies publish reliability figures for clusters of this size.
What it means for your business
- Expect model prices to fall, but don't build plans on it; energy scarcity can make them volatile.
- Cut waste in your own AI use: repeated queries, oversized models.
- Treat energy access as strategy if your work is compute-heavy.