Meta says its Prometheus AI supercluster has begun training Muse models

Meta says the first phase of Prometheus, its gigawatt-scale AI supercluster in New Albany, Ohio, is complete and has trained models in the Muse family. The planned full build is roughly one million GPUs, a scale that makes the infrastructure behind frontier AI part of the story, not just the models it produces.

Meta AI Watch analysis

What happened

In a 19 October engineering post, Meta describes Prometheus as hundreds of thousands of GPUs across interconnected buildings, operating as one system. It says the site was converted from a conventional data centre and that the completed first phase was used to train Muse models, including Muse Wonder.

The project is still being built out. Meta says the finished cluster is intended to reach roughly one million distributed GPUs. Its current hardware includes NVIDIA GB200 and GB300 NVL72 platforms, with a further platform planned.

Why it matters

Training frontier models at this scale depends on more than adding accelerators. The network, cooling, power, hardware reliability and software all have to work together; a failure in one part can interrupt a large training job. Meta’s account shows the engineering choices it says are helping Prometheus keep those jobs moving, including spare GPUs and dedicated networking between buildings.

It also gives a concrete view of the infrastructure being built to train Meta’s models. The reported reliability figures are Meta’s own, and do not by themselves establish how Prometheus compares with other clusters.

Our read

The striking detail is not simply the target of a million GPUs. It is the effort to make a sprawling installation behave like one dependable machine. Meta is offering an unusually detailed look at that engineering, while the scale and the performance figures remain the company’s account. The next test is whether the completed cluster delivers useful results as it grows.

What to watch

• Whether Meta reaches its planned Prometheus scale and on what timetable.
• Whether it publishes further reliability or training-performance results.
• How the next hardware platform changes the cluster’s capacity and operation.

Discussion spark: As AI clusters grow to this scale, should their success be judged mainly by the models they enable, or by the infrastructure’s efficiency and reliability?

not affiliated with, endorsed by, or operated by Meta