DeepSeek sends a shockwave through markets

In a flashAIAIAI RAI AIAIAIAIAIAI AIAI AI AIAIR1AI AIAIAIAI, euphoria over artificial intelligence () has turned to panic. Since early trading began on January 27th the market value of Nvidia, an chipmaking champion, has slumped by 17% at the time of writing. The share prices of Alphabet, Amazon and Microsoft—America’s cloud-computing triumvirate—have fallen by 3%, 1% and 3%, respectively. All told, American tech companies have shed around $1trn in value.The immediate cause of investors’ panic is DeepSeek, an that last week blew analysts away with its latest large language model, 1. Consumers are flocking to DeepSeek’s chatbot, which was the most downloaded app on iPhones over the weekend. Innovative techniques have allowed the company to train models that perform about as well as the most sophisticated Western models with only a fraction of the computer power—and so a fraction of the cost.DeepSeek’s entrance on the scene comes at a time when ever more on infrastructure. Last year the combined spending on data centres by the three cloud-computing giants and Meta (which has also been developing models) reached about $180bn, up by 57% from the year before. At the start of this month Microsoft said it would spend a further $80bn in 2025 on infrastructure. Last week Meta said it was planning to pour $65bn into this year.Yet if high-performing models can be trained with less computing power, all that investment may prove excessive. Although shareholders in the cloud-computing giants might welcome a reprieve from further capital spending, they may now be wondering what will become of the investments made to date. Even more worrying is what all this could mean for Nvidia and other companies supplying data-centre equipment, as well as those involved in providing the energy that fuels them. Shares in Siemens Energy, a maker of electrical equipment, and Cameco, a producer of uranium used for nuclear power, have tumbled by 20% and 13%, respectively. A rout in public markets will also ripple through to private companies. In 2024 venture capitalists invested $132bn in startups, an increase of more than half from the previous year, reckons PitchBook, a research firm. Model-makers that have been burning through cash, including Open and Anthropic, may find it harder to raise capital now that DeepSeek has shown it is possible to do more with less. Other venture-backed firms such as Groq and Cerebras, two chipmakers, and CoreWeave, an cloud-computing firm, could also face trouble.How bad will the rout be? Three uncertainties hang over the market. The first relates to the economics of . DeepSeek’s innovations suggest that the upfront cost of training a model may plunge. Yet that is happening just as a wave of so-called reasoning models, including Open’s o3 and DeepSeek’s own , are deploying much more computing power at the inference stage, where the model responds to questions. These models are able to, in effect, think harder about queries in order to generate better answers. What exactly the opposing forces of cheaper training and pricier inference mean for spending on computer power may not become clear for some time.A second source of uncertainty is geopolitics. America is trying to stymie China’s efforts by restricting the export of cutting-edge chips and the equipment used to make them. DeepSeek’s success shows that this strategy has so far failed. President Donald Trump may conclude that the answer is to further tighten export restrictions, worsening the blow for companies in the supply chain.The final area of uncertainty is demand. Some Wall Street bulls are already seeing a positive side to DeepSeek’s breakthroughs. As models become cheaper to train, companies could use the technology more. Already shares in businesses that are building services on top of others’ models, including Salesforce, an enterprise-software firm, and Apple, maker of the iPhone, have risen as investors bet on falling costs.Yet there is a worse scenario, too. For many enterprises, the cost of is less of a problem than difficulties implementing it on a large scale. If this persists, underlying demand for the technology could start to sag. And that would be enough to turn today’s market slump into a collapse.