AI's compute wars
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Allie Carl/Axios
Anthropic's runaway success is exposing AI's core problem: compute costs.
Why it matters: The closer AI labs get to IPOs, the harder it becomes to hide a structural margin problem: the more customers they win, the more they spend on the compute to serve them.
State of play: Anthropic's server capacity isn't keeping pace with demand, leaving paying customers stuck on usage limits and outages.
- Server capacity and compute power are finite resources that AI labs often have to purchase before they know how much demand they'll have from customers.
- Buy too much expensive capacity, erode your margins. Buy too little and you can't meet customer demand, and they'll run to your competitors as a result.
What they're saying: Anthropic CEO Dario Amodei said there's "no hedge on earth" against overbuying compute. Buying too much would bankrupt the company if demand falls short.
- Dylan Patel of SemiAnalysis warned Anthropic may be pushed toward lower-quality compute as OpenAI locks up premium supply.
- When Anthropic capped usage during peak hours, OpenAI said it would double limits.
- Amodei has signaled he'd rather lose customers in the short term than overbuy compute and torch his margins.
Between the lines: Compute isn't just fueling customer usage. Labs also need it to train upcoming models.
- Anthropic schedules training around peak hours to reduce costs, according to a source familiar with the matter.
- "Always watch the compute, other things matter, but any new capability breakthrough probably came from throwing more compute at it," Peter Gostev, AI capability lead at Arena AI wrote on X.
Yes, but: Compute costs are plummeting as efficiency in chips and software increases.
- But usage is skyrocketing faster so total spending keeps climbing: Classic Jevons Paradox.
Zoom out: AI capex from the hyperscalers is expected to hit nearly $700 billion this year as they race to build capacity.
- Some of that money goes toward future capacity — data center leases, power contracts — and maintaining existing systems, rather than raw compute.
- Translation: Even at record capex levels, the industry isn't buying enough compute to meet full demand.
Follow the money: While that could dissuade developers in the tech world, it has the opposite effect on Wall Street.
- Anthropic proved its spending discipline while OpenAI spent ferociously on compute, which is now resulting in less demand for shares in Altman's company, according to Bloomberg.
The bottom line: The AI race looks less like a model competition and more like a capital allocation problem — and the winners are still TBD.
