China’s technological advances have “narrowed” the U.S. headstart in AI and semiconductors and it’s now “unrealistic” for an advantage to last more than a few years, warns a U.S. think tank, as the Chinese startup DeepSeek rattles U.S. markets with its efficient and powerful AI models.
In late January, DeepSeek’s R1 model sent U.S. tech shares into freefall as investors questioned the billions of dollars in capital being poured into AI development. At one point, DeepSeek’s chatbot overtook OpenAI’s ChatGPT as the most downloaded free app on the U.S. app store.
Since DeepSeek’s release, U.S.-based commentators have tried to dismiss the startup’s claims about how it trained the model. But “DeepSeek’s technological innovations are real” and not just propaganda, argues a Friday research paper from the Center for Strategic and International Studies, a Washington D.C.-based think-tank.
“All DeepSeek’s technological innovations relate to algorithmic and architectural improvements,” Gregory Allen, director of CSIS’s Wadhwani AI Center and author of the report, wrote. While some of DeepSeek’s innovations were independently discovered by the U.S., others were “genuinely new,” he added.
Taken broadly, the gap between the U.S. and China in AI has “narrowed significantly, and it is unrealistic to expect a lead of more than a year or two, even with extremely aggressive export controls,” Allen wrote.
Still, the paper argues that export controls may still be the best way that the U.S. can maintain its narrow lead over China when it comes to AI applications and the semiconductor industry.
Despite extensive investment, China’s chip industry still lacks a domestic alternative for tools made by the likes of ASML and Applied Materials. That restricts how quickly China’s own chipmakers, like Semiconductor Manufacturing International Corporation (SMIC) can make the advanced chips needed to run AI applications.
Citing unnamed sources in the chip industry, the CSIS paper alleges that SMIC is still struggling with low yields of about 20% for advanced chips due to a lack of equipment. That pushes back against a Financial Times report in late February that suggested that Huawei’s SMIC-manufactured AI chips now have a 40% yield, enough to be profitable.
Yet the paper argues export controls alone won’t be enough to cement a U.S. lead. Washington will need to find a way to prevent controlled chips from getting into the Chinese market.
Countries like Singapore may have inadvertently become part of networks that transport advanced Nvidia chips to China, in violation of U.S. sanctions. About 18% of Nvidia’s revenue is generated by customers using Singapore as a billing location, but both the company and the Singaporean government note that only about 1-2% of revenue is generated from shipments to the country.
The complex chip supply chain means Washington has to count on multiple companies and governments to be responsible for enforcing U.S. law.
CSIS alleges that Huawei used shell companies to get access to TSMC’s manufacturing capacity, allowing the Chinese tech company to acquire more than 2 million AI chip dies, a building block for processors. That, in turn, could ensure a regular supply of Huawei’s Ascend AI chips, used for inference in DeepSeek’s R1 model.
In October, TSMC halted shipments to one of its Chinese customers as part of a probe into how one of its chips made it into a Huawei product.
Huawei did not immediately respond to a request for comment.
Beijing is also pushing the sector to produce domestic alternatives for advanced chipmaking equipment.
Together, chip smuggling and state-driven efforts to build a local alternative are two facets of what Allen calls a “formidable combination,” alongside poaching skilled talent and “substantial government investment.” CSIS suggests that Huawei might also be helped by “DeepSeek’s open-source community enthusiasm,” who may improve the Ascend chip’s AI software, making it a bigger challenge to Nvidia.
“Huawei and DeepSeek have a credible path to a million-Ascend-chip AI supercluster should they seek to build one,” Allen wrote.
