AI Competition Shifts to Capital Allocation Game: Computing Power Procurement Is a Life-or-Death Issue, Anthropic and OpenAI Adopt Different Strategies

The AI industry competition has shifted from model rivalry to a capital allocation game. Computing power procurement needs to be locked in 1-2 years in advance—buying too much leads to losses and bankruptcy, while buying too little causes customer churn. Anthropic's CEO says no hedging tool can avoid bankruptcy from wrong decisions.

Currently, no one in the industry can perfectly solve this problem. Anthropic chooses a restrained strategy, preferring to limit traffic and lose customers rather than overbuy; OpenAI invests aggressively in computing power, but its spending discipline has led investors to shift away. The industry's computing power supply still lags behind demand, even with hyperscale cloud vendors' AI capital expenditure expected to reach nearly $700 billion this year.

Although the unit cost of computing power is falling, usage is growing faster, leading to rising total支出, which fits the Jevons paradox. The closer to IPO, the harder it is to hide companies' capital allocation decisions.