Epoch AI's latest tracking shows frontier training facilities doubling peak compute every seven months, with the resulting buildout now approaching one percent of US economic output.
Peak compute at the world's largest AI training facilities has roughly doubled every seven months since August 2024, when SpaceXAI's Colossus 1 was completed, according to Epoch AI's latest tracking of frontier systems. That interval is the one to carry into every model release, funding round, and capacity announcement, because it lets a reader place each new project against a single ruler.
The chain of facilities that have held the top spot is short, and each one reads like a construction story more than a software story. SpaceXAI's Colossus 1, completed in 2024, was the first to set the new cadence. Anthropic and Amazon followed with New Carlisle, a hyperscale site in Indiana that has held the top spot, which Epoch AI describes as a facility built for AI training. Microsoft then opened Fairwater Atlanta, a facility in Georgia, and Meta's Prometheus cluster, announced earlier this year, has extended the cadence again, though its exact peak figure has not been disclosed publicly. Each new site has reset the bar, and each reset has landed inside the same seven-month window.
The physical buildout is now large enough to register in the national accounts. Epoch AI's analysis, drawn from public construction and capital-spending disclosures, puts AI infrastructure at roughly 0.8 percent of US gross domestic product, more than double its share in 2023. For a category that did not register as a line item five years ago, that is the threshold at which forecasters start treating it as a macro variable rather than a tech-sector curiosity.
The cadence is not, on its own, evidence that the resulting models are worth the spend. The same Epoch brief notes that Anthropic's Fable 5 currently tops the FrontierMath v2 leaderboard, a benchmark that Epoch itself describes as gameable. A research tool can flatter a model without measuring what the model does in the world. A companion piece on Mythos, an AI system whose cyber capabilities have been promoted ahead of any third-party evaluation, is the same caution in different packaging. The buildout is real, the compute is real, and the macro footprint is real; whether any specific model is buying commensurate capability is still an open question.
What to watch next is whether the next top-end facility lands inside the same seven-month window. If it does, the cadence has held through a full market cycle and the macro framing moves from suggestive to structural. If it slips, the buildout has hit a constraint worth identifying: power, chips, capital, or all three.