DeepSeek sends a shockwave through markets
It did not take much for the euphoria over artificial intelligence (AI) to turn into alarm. On January 27th stockmarkets in America and Europe convulsed. The share price of Nvidia, America’s AI-chip champion, fell by 17%, erasing $600bn of market value, the biggest one-day loss in the history of America’s stockmarket. Other businesses in the AI supply chain, from data-centre landlords to makers of networking gear, suffered a similar fate.
The cause of investors’ panic was DeepSeek, an obscure Chinese hedge fund turned AI startup that has blown analysts away with its latest large language model, R1, released on January 20th. Consumers have flocked to DeepSeek’s chatbot, which last weekend became the most downloaded app on iPhones. Innovative techniques have allowed the firm to train AI models that perform about as well as the most sophisticated Western ones with only a fraction of the computing power—and therefore a fraction of the cost.
In the following days, calm returned to markets. Share prices halted their descent, and in some cases partially regained their losses. Yet the episode is set to leave a lasting impression on investors, who have been forced to rethink who will profit most from AI, and consider what would happen were the bubble to burst.
The sprawling AI supply chain consists of hundreds of firms. Some of them, such as Nvidia, produce the hardware that sits inside data centres. Others rent out that gear as cloud-service providers (Amazon, Microsoft and Google). Model-makers (OpenAI and Anthropic) train AI systems in the cloud, and software firms (Salesforce and SAP) build applications on top of those models to sell to customers.
The clearest losers from DeepSeek’s breakthroughs are suppliers of AI hardware. If training models requires less computing power, then fewer chips and related equipment will be sold. Nvidia, which before the carnage on January 27th was the world’s most valuable firm, looks particularly exposed. Its most advanced chips, which are widely used in developing cutting-edge AI models, are said to generate gross margins in excess of 90%. If demand for them falls, the company’s monopoly-like margins will be squeezed.
Nvidia’s rivals, such as AMD, will also feel the pressure, though their valuations have not been as frothy. Share prices of AI-chip firms suffered a median decline of 6% on January 27th (see chart 1). The value of TSMC, a Taiwanese firm that makes most of the world’s cutting-edge semiconductors, tumbled by 13%.
Other suppliers of AI hardware are also being reassessed by investors. HPE and Dell, two American electronics firms that make server racks that sit inside data centres, saw their share prices sink by 6% and 9%, respectively, on January 27th. Chips need to talk to each other to train leading-edge models, which is why the value of Arista, a maker of networking gear, plunged by more than a fifth during the rout. Vertiv and Modine Manufacturing, two firms that make cooling equipment to stop AI chips from overheating, both lost over a quarter of their value.
Energy firms, too, have been caught up in the bloodbath. Many investors had assumed that training cutting-edge AI models would require ever greater amounts of electricity. But DeepSeek has thrown this into doubt. The share prices of Siemens Energy, a maker of electrical equipment, and Cameco, a producer of uranium used for nuclear power, fell by 20% and 15%, respectively, during the rout.
Another group of losers are the model-makers, such as the privately held OpenAI and Anthropic, whose businesses risk being undercut. They have been burning through cash, and could find it harder to raise capital now that DeepSeek has shown it is possible to do more with less.
Yet cheaper AI models will also create winners. Firms that build software applications on top of them, such as Salesforce and SAP, will benefit from falling costs. Many of these companies saw their share prices jump this week as others plunged. Apple, maker of the iPhone, may be another winner. It has not invested as heavily in AI infrastructure or model-making. Cheaper AI may also lead to a wave of new consumer apps, which could help perk up sluggish sales of its iPhones.
What all this means for the cloud giants is harder to predict. Alphabet, Amazon and Microsoft operate across the AI supply chain. Their software applications, such as Microsoft’s Copilot, may become more profitable as cheap models become more prevalent. But they have also invested in model-making, both directly (Alphabet has a large in-house team) and indirectly (through their stakes in startups such as OpenAI and Anthropic).
They have poured vast sums into AI infrastructure, too. Last year the combined spending on data centres by the cloud trio and Meta (which has also been developing AI models) reached about $180bn, an increase of 57% on 2023 (see chart 2). Although shareholders might welcome a reprieve from further capital spending, they may now be wondering what will become of the investments made to date.
The cloud giants have also been venturing into chip design, in an effort to reduce their reliance on Nvidia. Mario Morales of IDC, a research firm, notes that their semiconductor-engineering teams are each now roughly as big as that of a large chipmaker. Investors seem to think that these ambitions will be cut back. During the rout the share prices of Broadcom and Marvell, two firms that help the cloud giants design their own chips, fell by 17% and 19%, respectively. That contrasts with the share prices of the cloud giants, which fell slightly (in the case of Alphabet and Microsoft) or not at all (for Amazon).
Plenty in the industry remain bullish. SoftBank, a spendthrift Japanese investor, is reportedly in talks to plough $15bn-25bn into OpenAI. On January 29th Mark Zuckerberg, Meta’s boss, said his firm would invest “hundreds of billions” of dollars in AI over the long term.
One reason is that so-called reasoning models, including OpenAI’s o3 and DeepSeek’s own R1, are deploying much more computing power at the inference stage, where the model responds to questions, to generate better answers. That could help offset decreases in the use of computing power for training.
Another source of optimism relates to demand. On January 29th Microsoft said that the growth of its AI cloud revenue— 157%, year on year, for the quarter—was constrained by supply, not demand. Some have argued that as AI models become cheaper to train, usage will rise further.
Yet other hurdles remain to adoption. America’s Census Bureau surveys firms about their use of AI. Of those that have no plans to use it in the next six month, only 4% cite cost as the reason. The vast majority think that AI simply does not apply to their business.
If DeepSeek’s innovations lower the cost of AI by orders of magnitude, companies may well discover new applications for the technology. Reasoning models will help, too, by improving the performance of AI models. Any effect on demand, however, will take time to materialise. And as the market turbulence demonstrates, investors are getting jittery. Their patience may not last for ever. ■
To stay on top of the biggest stories in business and technology, sign up to the Bottom Line, our weekly subscriber-only newsletter.