What is DeepSeek, the Chinese AI company upending the stock market?

People reflected in a window with a slogan about AI at a representation of a company ahead of the World Economy Forum in Davos, Switzerland, Sunday (AP Photo/Markus Schreiber)

A frenzy over an artificial intelligence chatbot made by Chinese tech startup DeepSeek was upending stock markets Monday and fueling debates over the economic and geopolitical competition between the U.S. and China in developing AI technology.
DeepSeek’s AI assistant became the No. 1 downloaded free app on Apple’s iPhone store Monday, propelled by curiosity about the ChatGPT competitor. Part of what’s worrying some U.S. tech industry observers is the idea that the Chinese startup has caught up with the American companies at the forefront of generative AI at a fraction of the cost.
That, if true, calls into question the huge amounts of money U.S. tech companies say they plan to spend on the data centers and computer chips needed to power further AI advancements.
But hype and misconceptions about DeepSeek’s technological advancements also sowed confusion.
“The models they built are fantastic, but they aren’t miracles either,” said Bernstein analyst Stacy Rasgon, who follows the semiconductor industry and was one of several stock analysts describing Wall Street’s reaction as overblown.
“They’re not using any innovations that are unknown or secret or anything like that,” Rasgon said. “These are things that everybody’s experimenting with.”

What is DeepSeek?
The startup DeepSeek was founded in 2023 in Hangzhou, China and released its first AI large language model later that year. Its CEO Liang Wenfeng previously co-founded one of China’s top hedge funds, High-Flyer, which focuses on AI-driven quantitative trading. The fund, by 2022, had amassed a cluster of 10,000 of California-based Nvidia’s high-performance A100 graphics processor chips that are used to build and run AI systems, according to a post that summer on Chinese social media platform WeChat. The U.S. soon after restricted sales of those chips to China.
DeepSeek has said its recent models were built with Nvidia’s lower-performing H800 chips, which are not banned in China, sending a message that the fanciest hardware might not be needed for cutting-edge AI research.
DeepSeek began attracting more attention in the AI industry last month when it released a new AI model that it boasted was on par with similar models from U.S. companies such as ChatGPT maker OpenAI, and was more cost-effective in its use of expensive Nvidia chips to train the system on troves of data. The chatbot became more widely accessible when it appeared on Apple and Google app stores early this year.
But it was a follow-up research paper published last week — on the same day as President Donald Trump’s inauguration — that set in motion the panic that followed. That paper was about another DeepSeek AI model called R1 that showed advanced “reasoning” skills — such as the ability to rethink its approach to a math problem — and was significantly cheaper than a similar model sold by OpenAI called o1.
“What their economics look like, I have no idea,” Rasgon said. “But I think the price points freaked people out.”

The ‘Sputnik’ backdrop
Behind the drama over DeepSeek’s technical capabilities is a debate within the U.S. over how best to compete with China on AI.
“Deepseek R1 is AI’s Sputnik moment,” said venture capitalist Marc Andreessen in a Sunday post on social platform X, referencing the 1957 satellite launch that set off a Cold War space exploration race between the Soviet Union and the U.S.
Andreessen, who has advised Trump on tech policy, has warned that overregulation of the AI industry by the U.S. government will hinder American companies and enable China to get ahead.
But the attention on DeepSeek also threatens to undermine a key strategy of U.S. foreign policy in recent years to restrict the sale of American-designed AI semiconductors to China. Some experts on U.S.-China relations don’t think that is an accident.
“The technology innovation is real, but the timing of the release is political in nature,” said Gregory Allen, director of the Wadhwani AI Center at the Center for Strategic and International Studies. Allen compared DeepSeek’s announcement last week to U.S.-sanctioned Chinese company Huawei’s release of a new phone during diplomatic discussions over Biden administration export controls in 2023.
“Trying to show that the export controls are futile or counterproductive is a really important goal of Chinese foreign policy right now,” Allen said.
On Monday, Trump said DeepSeek’s breakthrough was “good because you don’t have to spend this much money.”
Speaking Monday to House Republicans in Miami, Trump called the DeepSeek news “positive” if it is accurate because “you won’t be spending as much and you’ll get the same result.” He called the development a “wakeup call for our industries that we need to be laser focused on competing to win.”
Trump signed an order on his first day in office last week that said his administration would “identify and eliminate loopholes in existing export controls,” signaling that he is likely to continue and harden Biden’s approach.
DeepSeek’s progress on AI without the same amount of spending could possibly undermine the potentially $500 billion AI investment by OpenAI, Oracle and SoftBank that Trump touted at the White House.
Nvidia’s stock dropped 17% Monday, but the company in a statement commended DeepSeek’s work as “an excellent AI advancement” that leveraged “widely-available models and compute that is fully export control compliant.”

What makes DeepSeek different?
One thing that distinguishes DeepSeek from competitors such as OpenAI is that its models are “open source” — meaning key components are free for anyone to access and modify, though the company hasn’t disclosed the data it used for training.
But what’s attracted the most admiration about DeepSeek’s R1 model is what Nvidia calls a “perfect example of Test Time Scaling” — or when AI models effectively show their train of thought, and then use that for further training without having to feed them new sources of data.
“It’s just thinking out loud, basically,” said Lennart Heim, a researcher at Rand Corp.
OpenAI’s reasoning models, starting with o1, do the same, and it’s likely that other U.S.-based competitors such as Anthropic and Google have similar capabilities that haven’t been released, Heim said.
But “it’s the first time that we see a Chinese company being that close within a relatively short time period. I think that’s why a lot of people pay attention to it,” Heim said. “I used to believe OpenAI was the leader, the king of the hill, and that nobody could catch up. Turns out this is not completely the case.”
MDT/AP