Only rarely does a single company’s new product provoke a major market sell-off. But that’s exactly what happened on Monday, when a large language model from a Chinese company named DeepSeek drove the entire Nasdaq index of tech companies down more than 3 percent and shaved more than 17 percent off the market capitalization of the chipmaker Nvidia—which, until that moment, had been the most valuable company in the world.
The panicked selling of Nvidia had a surface logic. The company provides almost all of the computer chips (called GPUs) that companies such as Alphabet, OpenAI, Microsoft, and Meta rely on to train their LLMs. (The Atlantic entered into a corporate partnership with OpenAI in 2024.) Consequently, it has been the biggest beneficiary of the huge boom in corporate spending on AI that we’ve seen over the past few years. (Nvidia’s annual revenue has quadrupled since 2022.) Although DeepSeek also used Nvidia chips to train its model, the company said that they were an older type of GPU—U.S. export controls imposed by the Biden administration have prevented Chinese companies from buying cutting-edge chips. DeepSeek’s disclosure raises the possibility that future progress in training LLMs could be made with fewer, simpler chips, and at a lower cost than previously anticipated. That would obviously put a big dent in Nvidia’s profits. So investors dumped its stock.
[Read: The DeepSeek wake-up call]
If investors are very concerned about how DeepSeek might hurt chipmakers, they seem surprisingly unconcerned about how it might affect big AI software companies. Meta’s stock price, for instance, actually rose on Monday, and although the stocks of Alphabet and Microsoft did take a hit, they bounced back over the next couple of days. Some of that is because the underlying business of these companies, independent of AI, remains enormously profitable. But it also suggests that investors aren’t paying enough attention to the way DeepSeek’s success could disrupt the AI market, and in doing so threaten the future profits of the tech companies that are currently spending many billions of dollars every year on their LLMs.
Tech investors have historically profited by spotting the new new thing. But at the moment, they seem implicitly to assume that all of the fundamental change in the LLM business has already happened and that its future will look much like its present, with the companies that currently dominate the space—many of which are not simply competitors but also financial partners—continuing to do so indefinitely. What happened over the past week is a reminder that these assumptions may not be so solid.
The large language model that caused such a stir on Monday, DeepSeek-R1, is clearly comparable with LLMs such as ChatGPT o1-mini and Claude 3.5. Measured by industry benchmarks that rate subject knowledge, reasoning, and accuracy, the DeepSeek model seems to deliver similar performance while costing much less to develop—though just how much less remains a matter of debate. Beyond dispute is that it’s cheaper to use: Consumers can get access to DeepSeek’s core functions for free, and third-party developers are being charged a fraction of the cost of a product such as ChatGPT. DeepSeek also uses open-source technology, meaning that, in theory, you could download the program and run your own AI on your desktop if you had a powerful-enough computer. The fact that the LLM offers reasonable performance—results that, even a year ago, would have seemed startlingly good—at a significantly lower cost means that it has to be taken seriously as a competitor.
From one angle, in fact, DeepSeek looks like what the business-school professor Clayton Christensen, in his book The Innovator’s Dilemma, dubbed a “disruptive technology”: a product that’s less powerful than the products at the top of the market but also much cheaper, and that has the possibility of improving in quality over time to the point where it offers a superior combination of price and performance for most customers. In this regard, the rapid uptake of DeepSeek by users around the world has been striking. The LLM still has miles to go in market share to catch ChatGPT, which has more than 300 million weekly users, but since its release on January 20, its mobile-app version has been downloaded more than 3 million times from Google Play and Apple, making it the most popular app on both stores. That suggests that the cost of switching from one AI tool to another is very low, and that the moats big AI companies are building around their business may be much shallower than they’d hoped.
[Read: China’s DeepSeek surprise]
The underlying wager that these companies have made is that the big money they’re investing will result in radically better performance, which in turn will enable them to charge hefty sums to businesses and, to a lesser extent, consumers. (OpenAI, for instance, is reportedly targeting $100 billion in revenue by 2029.) And these companies remain committed to that bet. This week, the CEOs of both Microsoft and Meta said that enormous spending is essential to staying competitive in the market. Dario Amodei, a co-founder and the CEO of Anthropic (in which both Amazon and Google have invested heavily), wrote in a blog post that companies are going to continue to “spend more and more on training powerful AI models, even as … the cost of training a given level of model intelligence declines rapidly,” because “the economic value of training more and more intelligent models is so great.” In the long run, such investment may well result in the kind of performance improvement that a company like DeepSeek (which can’t even get access to the most powerful GPUs)—or the many other low-cost LLM developers that are sure to try to emulate it—cannot keep up with.
When you look at ordinary users’ embrace of DeepSeek, though, you can also see an alternative future. In this one, AI performance improves so much that most customers are happy with cheap, good-enough LLMs, and AI models end up as essentially interchangeable, commoditized products, with the small profits that always follow that type of commercial diffusion. We’re going to find out whether the great authors of the disruptive technology that’s transforming the business world might themselves get disrupted.