The AI competition is shifting from a battle of models to a game of capital allocation – whoever miscalculates first gets eliminated.

According to 1M AI News's monitoring, Axios published an analysis article stating that the AI ​​competition is becoming less like a model competition and more like a capital allocation question. Purchasing computing power needs to be locked in one to two years in advance; buying too much leads to bankruptcy, while buying too little results in customers running out. Anthropic CEO Dario Amodei's exact words on the Dwarkesh Podcast were: if you purchase with a supposed 10x annual growth, in reality, it's only 5x or a year late, "there is no hedge in the world that can prevent bankruptcy." While the unit cost of computing power has indeed decreased, usage has increased faster, resulting in a continual rise in total spending, illustrating the classic Jevons Paradox.

The article points out that no one has answered this question correctly yet. Anthropic chooses restraint, preferring to throttle and lose customers rather than overbuy, steering training tasks away from user peaks; OpenAI chooses aggression, making significant investments in computing power. Each strategy has its costs: Anthropic's paying customers frequently encounter throttling and interruptions, with Dylan Patel from semiconductor analysis firm SemiAnalysis warning that they may be forced to shift to lower-quality computing power; OpenAI's spending discipline is already evident in the secondary market, as investors are transitioning from OpenAI to Anthropic. The AI ​​capital expenditures of ultra-large-scale cloud providers are expected to approach $700 billion this year. Even at this record level, the industry's computing power supply still can't keep up with demand. The closer it gets to an IPO, the harder it is to hide the answer to this question.