DeepSeek changes the narrative in the middle of the gold rush

Financial markets are the largest and most complex incentive system ever devised by humankind. Sometimes brutally efficient, sometimes brutally dysfunctional, and almost always a bit of both. Ever since OpenAI launched the first version of ChatGPT on November 30, 2022, and the whole world watched as a machine appeared to think (though it’s more like searching and drafting), the financial music has been played in one key: artificial intelligence, the next promised land of finance, the technology that will (supposedly) change the world. Like the steam engine, the internet or computing. Who would want to stay on the sidelines?

The magic of the markets caught fire: investors have been willing to pay anything not to be left out of the party, so the big techs and their deep pockets went on a feverish arms race: they need to train (and show the market, of course) better language models, and for that they need more data, and much more processing and storage capacity. AI was a question of scale and, in this view, the best position goes to the one who invests the most. It is all a matter of scale and brute force; performance would come later and, based on recent experience, the winner takes it, as happened with search engines (Google), operating systems (Microsoft), social networks (Meta), online commerce (Amazon) or high-end technological devices (Apple) or, momentarily, high-end electric vehicles (Tesla). The Magnificent Seven are seven magnificent monopolies or quasi-monopolies.

The first winner of this euphoria has been Nvidia, the technological version of the pick-and-shovel sellers in the gold rush. It went from earning $16 billion in 2021 to expecting revenues of $200 billion in 2026 thanks to the chips (based on graphics processors) with which Google, Meta, Amazon and company were seeking the marginal advantage of their training models. And with a profit margin typical of a monopoly, around 50%. There were already some voices warning of the diminishing returns of the best-known language models. But the market wheel is unstoppable. Last week Meta boasted about its $65 billion investment in AI.

DeepSeek’s model has supposedly been developed with less than $6 billion (a figure that is not verified). Its model is open source and accessible to the public. And it explains its reasoning. It is an amendment to the narrative embraced by Silicon Valley and Wall Street. And in the markets, numbers count, but sometimes the narrative counts even more. Today it is questionable whether such large investments in equipment are necessary. And the profitability of these investments is in question (given that results can be obtained more cheaply). And it is questionable whether the market is willing to pay dizzying sums for a competitive advantage that may not be so unique after all.

Nvidia’s vertical and backward-feeding rise, its rise to three trillion and change in stock market value, has been a breeding ground for a whole financial ecosystem around this seemingly easy money: stock options (including those with overnight expiration, rather like a sports bet), leveraged exchange-traded funds (also bearish funds), massive entry of hedge funds and, to add a touch of folklore, parties at results presentations. The echo of past bubbles is evident, even though AI (like the internet, like so many other technologies) is here to stay. It is simply that it is no longer easy to play this card in the markets, and they will take it into account when defining the premium they pay for shares. That is the first and clearest implication. There are other possible ones, such as the impact of losses in a financial world, the potential impact of the fact that the US is no longer the exclusive leader in the world of AI, or whether less expensive models can really democratize this technology. But those are even harder to gauge.

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition