China’s Bargain-Basement DeepSeek Deep Sixes Artificial Intelligence Stocks on Wall Street
Investors question whether billions of dollars of investment in language models was too much too soon.
Wall Street is rattled today by the possibility that billions of dollars planned and already invested in artificial intelligence hardware may have been misdirected.
Nvidia, the leading supplier of computer chips for AI gear, slid 16.9%. That dragged down other AI hardware makers — and perhaps unexpectedly utility companies — but there were some winners too as investors rushed for safe havens amid the confusion.
The angst came courtesy of a Chinese company called DeepSeek, whose R1 language model competes with OpenAI’s ChatGPT, the market leader, at least until last week.
What seems to have caught everybody’s attention was the apparent disclosure that DeepSeek has produced a serious competitor to ChatGPT for $6 million, a piddling sum compared with the $500 billion that OpenAI coincidentally announced on Monday would be invested in building new American infrastructure over the coming four years in a project called Stargate. Other investors include Softbank and Oracle, and Nvidia, Microsoft, and Arm will provide technology backing.
The $6 million figure is misleading. It comes from what DeepSeek termed a technical report about the third version of its language model. That report was released on December 26, and if one reads far enough one will find that it only includes the cost of training the third version of the model, leaving open the possibility that the total investment in “prior research and ablation experiments on architectures, algorithms, or data” was significantly higher.
Still, it seems clear that DeepSeek is capable of producing a credible AI system for a fraction of the cost being committed in America, and unlike its increasingly proprietary rivals the Chinese program is open source. That means anybody can use it and make tweaks to improve or customize it, which would aid in its competitiveness versus ChatGPT, Meta’s Llama, and Google’s Gemini.
DeepSeek used an approach called “mixture of experts” to cut the computing power needed to perform AI. This seems to be a case of necessity leading to invention because American export rules prevent the Chinese company from acquiring the most powerful Nvidia chips.
Nvidia, which last week was the most valuable company in the world, ended up in third place after Monday’s decline, according to CompaniesMarketCap, trailing Apple and Microsoft. Apple, which has had a rough start to the year as investors questioned its outlook for iPhone growth in China and lackluster AI offering, was up 3 percent but is still about 8 percent lower on the year.
For the day, the blue-chip Dow Jones industrial average eked out a 0.7% gain, while the technology-laden Nasdaq composite was off 3%. The broader Standard & Poor’s 500 split the difference with a 1.5% decline. Among the S&P’s sectors, tech was down 4.9 percent and utilities fell 2.3 percent.
High-performance computing, which includes support for power-hungry AI, has been supporting utility prices and capital spending. Meanwhile investors rotated into consumer staples, up 2.7 percent, and health care, which gained 2.3 percent.