China’s artificial intelligence company, DeepSeek, has caught the attention of American lawmakers with its exceptional performance in AI development.
On Monday, January 27, DeepSeek, a Chinese artificial intelligence company, caused a stir on Wall Street, leading to a nearly $600 billion market value loss, marking the largest financial shock event in U.S. stock market history.
The release of DeepSeek’s R1 reasoning model has surpassed various American open-source language models, surprising and raising doubts among many AI researchers and U.S. officials.
American congressional members have urged action to slow down the development of DeepSeek, a Chinese tech startup, with some considering DeepSeek a “serious threat.”
John Moolenaar, Chairman of the House Committee on the Chinese Communist Party and Republican Federal Congressman, stated that the U.S. government should strive to implement stricter export controls on DeepSeek’s AI infrastructure and key technologies.
Chuck Schumer, Minority Leader of the Senate and Democratic Federal Senator, said, “Our competitors will use every opportunity they have to surpass the U.S. in future technological domains. We must lead the way.”
Schumer expressed that he will continue to prioritize advancing U.S. artificial intelligence development.
According to Bloomberg, the market’s response to DeepSeek causing the drop in U.S. stocks is seen as a strange delayed reaction. The sell-off of U.S. high-tech company stocks on Monday was actually the fifth trading day investors could have acted upon.
DeepSeek released the application on January 20, the day of the U.S. Presidential Inauguration. Initially, there was little attention given to it, but by the 24th, news about DeepSeek began to surge. There were suddenly 878 related news reports online, almost 600 more than the previous day.
Over the weekend, DeepSeek began to gain more presence on social media platforms. By the evening of the 26th, Google Trends showed a surge in search activity for DeepSeek.
While market reactions to news should be random, typically responding quickly when news breaks, reports regarding DeepSeek had been circulating for several months.
Reportage has emphasized that in the social media era, reaching a critical mass of information and rapidly spreading is common, but when the information on DeepSeek reached that critical threshold and caused a $580 billion loss on Wall Street in just one day on the 27th, it was unmatched in terms of impact.
There have been doubts raised about DeepSeek’s low-cost training being highly misleading, along with questions about their use of non-high-end chips for training.
DeepSeek released a report in December 2024 claiming that their soon-to-be-launched free open-source large language model V3 surpassed various domestic and foreign large model products in multiple data aspects. The company claimed the model was built in just two months, costing only $5.57 million, and trained using non-high-end H800 chips.
The $5.57 million cost overlooks the significant underlying infrastructure of DeepSeek, including chips that could potentially increase actual costs to hundreds of millions of dollars.
The report also mentioned that these costs did not include all other costs related to architecture, algorithms, or previous research and experiments in the data aspect.
On January 20, the company released the R1 reasoning model, claiming performance on par with the OpenAI O1 official version. They did not disclose the cost data of R1.
However, the $5.57 million figure became a headline in American media reports, also being misconstrued as the cost of R1. Stacy Rasgon, Senior Analyst at Bernstein Research, quipped, “DeepSeek really ‘founded OpenAI for over five million dollars’? Of course not.”
Alexandr Wang, CEO of Scale AI, a leading company in data labeling services, pointed out that DeepSeek may actually possess 50,000 H100 GPUs but might not be able to disclose this due to export control restrictions.
Daniel Newman, CEO of Futurum Group, also believes that if DeepSeek acquires equipment it should not have, further investigations will ensue. He anticipates that the U.S. government will investigate whether DeepSeek’s hardware violates export prohibitions.
Based in Hangzhou, the startup DeepSeek stated that it used Nvidia’s H800 chips to build the model, which was available until October 2023 when the U.S. government expanded its ban.
According to publicly available information and reports from DeepSeek founder Liang Wenfeng in interviews with mainland media, Liang had purchased a large number of Nvidia A100 chips well before anticipated U.S. sanctions, which are currently prohibited from being exported to China.
Chinese tech publication, 36Kr, estimated the company’s inventory of chips to exceed 10,000 pieces, while Dylan Patel, founder of the AI research consulting firm SemiAnalysis, believed the inventory to be at least 50,000 pieces. These reserves hold potential for training AI and can be combined with low-power chips like H800 to develop models.
“Our issue has never been money but high-end chips being embargoed,” Liang Wenfeng told 36Kr.
It is still unclear whether DeepSeek’s R1 model uses the banned high-end chips and in what quantities.
Michelle Giuda, CEO of the Krach Institute for Tech Diplomacy at Purdue University, told Bloomberg TV that the U.S. must maintain “truly robust defenses” by strengthening export controls. She added that, at the same time, if the U.S. wants to stay ahead in the tech race, it needs to catch up in cultivating excellent engineers compared to China.
“What we should do is redouble our efforts to make America act faster, smarter, and at lower costs in AI innovation,” she said.
Miles Brundage, former research lead at OpenAI, expressed concern on the ChinaTalk podcast that some might use the DeepSeek case as a final narrative against export controls.
“While I don’t think this argument holds, I understand why people might see it and conclude that export controls are counterproductive,” he said. “Export controls are effective – you can’t deny this situation is happening, and it’s more likely to happen because of export controls.”