AI competitions are shifting from model battles to capital allocation games; whoever makes the first miscalculation is the first to be eliminated.

According to 1M AI News monitoring, Axios published an analysis article saying the AI race is becoming less like a model competition and more like a capital allocation puzzle. For computing power procurement, you need to lock in purchases one to two years in advance—buy too much and you can lose to the point of bankruptcy, buy too little and customers will run off. Anthropic CEO Dario Amodei’s exact words on the Dwarkesh Podcast were: “If you procure at a 10x annual growth rate, the reality is only 5x, or it’s a year late—there is no hedging instrument in the world that can stop you from going bankrupt.” And while the unit cost of computing power does indeed fall, usage grows even faster, total spending keeps climbing—this is the classic Jevons paradox.

The article notes that so far, nobody has solved this question. Anthropic chose restraint: it would rather throttle and lose customers than overbuy, and its training tasks avoid user peak hours. OpenAI chose aggression, making major investments in computing power. Both strategies come with trade-offs: Anthropic’s paying users frequently hit throttling limits and interruptions. Dylan Patel of the semiconductor analysis firm SemiAnalysis warned it might be forced to shift to lower-quality compute. OpenAI’s spending discipline, meanwhile, is already reflected in the secondary market, with investors moving from OpenAI to Anthropic. This year, the AI capital expenditures of hyperscale cloud providers are expected to be close to $700 billion. Even at this record level, compute supply across the entire industry still can’t catch up with demand. The closer you get to the IPO, the harder it is to hide the answer to this question.