Meta has published details of Prometheus, a gigawatt-scale compute site in New Albany, Ohio, which the company says will eventually target roughly one million GPUs. The stated purpose is training its Muse model family.
A gigawatt at a single site is the scale at which a data centre stops being a building and becomes a grid planning problem. For comparison, the three US nuclear plants the Department of Energy is lending $4.42 billion to restart - Palisades, Crane and Duane Arnold - total 2.25 GW between them. Prometheus alone is on the order of half that, for one company, for one model family.
The Muse context matters. Meta launched Muse publicly in September, and the model family now spans Muse Spark Open at 816B parameters with open weights, Muse Wonder, and the agent product that this week shared a seller's home address on Facebook Marketplace. Prometheus is the training capacity behind all of it, and the timeline Meta has given runs well beyond the current generation.