DeepSeek made AI cheap, and the American AI bubble was built on it being expensive

DeepSeek, a Chinese lab that grew out of a quantitative hedge fund, keeps doing the thing the American AI industry priced as impossible. It builds models that perform close to the frontier, trains them for a reported fraction of what its U.S. rivals spend, and then publishes the weights for anyone to download and run. Each release reopens the same argument: that the valuation of the entire American AI sector rests on an assumption DeepSeek is quietly dismantling, which is that intelligence has to stay expensive.

That assumption is not abstract. It is the wall holding up hundreds of billions of dollars in data-center construction, the share prices of the chipmakers and cloud providers sitting inside almost every index fund, and the monthly subscription many readers already pay for a chatbot. If a rival delivers comparable results for far less and gives the software away, the premium attached to scarce, costly compute starts to look less like a moat and more like a wager.

DeepSeek’s claim is about efficiency, not sorcery. Its engineers leaned on a mixture-of-experts design that wakes only a slice of the model for any single query, on aggressive use of lower-precision math, and on training pipelines tuned to run on fewer and partly export-restricted chips. The widely repeated figure for one of its main training runs sat under six million dollars. Comparable American runs are understood to cost many times that once the full bill is added up.

How the model is released matters as much as what it cost. DeepSeek puts out open weights, so a developer in Sao Paulo, a university lab in Warsaw, or a startup in Seoul can pull the model down and run it on their own machines, without paying a U.S. provider for every query and without shipping their data overseas. There is an irony in it: the export controls meant to slow Chinese AI by cutting off the most powerful chips appear to have pushed DeepSeek toward squeezing more out of less, and those efficient methods now travel everywhere the open weights do.

For the person who simply uses these tools, the immediate effect is choice. Cheaper models put downward pressure on subscription prices, push more capable assistants onto ordinary laptops and phones, and weaken the case for locking into a single provider. The thing that felt like a utility you rent starts to look like software you can own.

The bubble talk needs heavy caveats. That sub-six-million-dollar figure covers a single final training run, not the research, the dead ends, the salaries, or the hardware that made it possible, so setting it against a U.S. lab’s all-in spending compares two different things. Open weights are also not open source; the training data and the full method stay private. And the efficiency point cuts both ways. Asked about DeepSeek, Microsoft’s leadership reached for Jevons paradox, the old observation that when a resource gets cheaper to use, total demand for it tends to climb rather than fall. Cheaper intelligence may simply mean the world buys far more of it, which would be good news for the companies selling compute, not bad.

This is also not the first time the bubble has been pronounced dead. The same lab once triggered the largest single-day loss of market value in U.S. history, erasing close to six hundred billion dollars from one chipmaker in an afternoon, then watched the stock claw most of it back within weeks. The big American AI companies did not answer by spending less. They raised more and built bigger. Any claim that the bubble has finally burst has to survive the fact that the people with the most money on the table keep doubling their bet.

What DeepSeek has actually done is harder to dramatize than a burst bubble. It has stripped away the comfort of assuming the leading American labs are shielded by a wall of capital no one else can scale. If frontier-level capability can be approximated cheaply and handed out at no charge, the value stops living in owning the model. It moves to distribution, to the products built around the model, and to whoever controls the customer. The next test is already on the calendar, even without a date attached to it: every fresh DeepSeek release reopens the same question and lands in a market that has committed to spending more, not less, on the belief that scale still wins. Where it gets settled is in the earnings calls and capital-expenditure guidance of the coming quarters, not in a forum thread declaring the fight already over.