The footnote that cost $589bn

What is DeepSeek

DeepSeek lives in Hangzhou, a city better known for tea and a lake, inside the fortune of a quantitative fund called High-Flyer. Its founder, Liang Wenfeng, had the foresight, or the luck, to buy ten thousand Nvidia chips before Washington forbade it. With them he built a laboratory of young, unfamous graduates and a habit of publishing everything. Its models arrive like a neighbour’s fireworks: late at night, unannounced, and annoyingly good.

What happened

On 26 December 2024 DeepSeek published a report on its new model, V3. One sentence gave the cost of training it: “only $5.576M”. The next excluded “the costs associated with prior research” and every experiment before. It was the price of the last run, not of the model. Almost nobody quoted the second sentence.

Why we found this curious

The sentence crossed the Pacific in an afternoon. On 27 December an investor on Xueqiu, China’s stock forum, wrote: “If DeepSeek’s numbers are real, Nvidia halves.” Chinese computing stocks fell 4.25% that week. In New York the same number was praised by the most famous engineers in the world, and on 6 January Nvidia closed at a record. In Shanghai most shares are owned by individuals. In New York, by institutions.

Tracing the tracks

We followed the number through 130 voices. Each retelling dropped something. Andrej Karpathy, a founder of OpenAI, called it “a joke of a budget”. An engineer at Meta wrote that every leader of its AI group earned more than DeepSeek’s model had cost. A post saying it “only cost them $6million to make vs the Americans spending billions” collected 146,000 likes.

It caught fire in one week of January. On the 20th DeepSeek released R1, free to anyone with a phone. On the 21st Donald Trump announced Stargate, $500bn of data centres. Then came a weekend. By Sunday DeepSeek was the most downloaded free app in America.

A daunting comparison

$500bn against $5.6m is ninety thousand to one. Pedro Domingos, a computer scientist, wrote the next day that Stargate “would look a lot better” if DeepSeek had not just matched OpenAI’s best model “at a tiny fraction of the cost.” The sum was wrong. SemiAnalysis, a research firm, later put DeepSeek’s spending on servers above $1bn. That figure collected fewer likes.

What if we sort them by taste

Most research sorts voices by fame. We sorted them by taste: how early, and how exactly, each one saw. Taste: outsiders who saw first. Fame: the great names of technology and politics. Money: executives and fund managers. Then the crowd, and the press.

Taste heard it first. Fame made it loud. In November Dylan Patel, an analyst, wrote that DeepSeek had “over 50k Hopper GPUs”, Nvidia’s best chips; in January the whole market was asking the same question. On 27 December Arnaud Bertrand, who writes about China, called DeepSeek a “Sputnik moment”. On 26 January Marc Andreessen said the same, and it became the headline of the world.

Insights

By the evening news, $5.576M for one training run had become $1trn of America’s stock market. The crowd, which said sell, bought. The funds, which said nothing, sold. Nine months later Nvidia was worth $5trn, and the footnote had been right all along. Next time we will read the small print first, and in Chinese.