The US bans the supply of Nvidia's flagship chips, Blackwell and Rubin, to China, citing security threats: advanced artificial intelligence (AI) could have military applications. The restrictions are also meant to preserve the leadership of American laboratories. However, Chinese companies continue to train their models on advanced Nvidia accelerators.
Nvidia develops graphics processing units (GPUs), devices originally created to process three-dimensional graphics in video games. They are capable of processing enormous volumes of data, so they are widely used to train and run AI systems. After the launch of ChatGPT in 2022, interest in the technology surged, and Nvidia's market value reached 4.5 trillion dollars.
Nvidia chief Jensen Huang lobbied for permission to sell to China, warning that bans would spur the development of local competitors. This year Washington eased controls on the less powerful H200 chip, but on one condition: the US government takes 25% of sales revenue. Approval of supplies remains complicated, and Beijing, according to reports, is in no hurry to buy the H200, as China's own chip industry is gaining momentum.
Recent launches of powerful models from Chinese startups Moonshot AI, Z.ai and DeepSeek have increased pressure on American competitors. "Chinese manufacturers still use advanced Nvidia chips to train their models," Lian Jye Su, chief analyst at Omdia, told AFP.
Chips can enter the country through smuggling or through a loophole: Chinese companies rent computing capacity in data centers in third countries, often in Southeast Asia, where Nvidia accelerators are installed. "US export controls restrict where a chip is supplied, not who uses it," explained Steven Wu, founder of the investment fund Carthage Capital. According to him, Nvidia GPUs could be located, for example, in the Malaysian state of Johor "in full compliance with the rules, while a client from Shenzhen connects to them."
White House adviser Michael Kratsios said in the summer that Moonshot AI had gained access to high-end Nvidia chips in Thailand. Washington has also accused Chinese AI laboratories of training systems to imitate the capabilities of American models. In December, The Information reported that DeepSeek had developed a new model using thousands of chips disassembled abroad and brought into China. DeepSeek did not respond to a request for comment, and Nvidia said it had seen no evidence of such smuggling and considers it "unlikely."
"Nvidia chips still accelerate and simplify top-level AI work, but they are no longer the only path," noted Po Zhao, founder of the analytical publication Hello China Tech. "The open question is cost," since training on Chinese chips usually requires more time and engineering effort. Nvidia's main competitor in China is Huawei, although it lags behind the American giant. DeepSeek is reported to be training its next major model on Huawei chips. In September, Huawei said its next-generation AI chip would be ready in the first quarter of 2027, three quarters earlier than planned.
To increase the power of Chinese chips, AI companies need to use more of them, Wu explained. "They pay for this with energy and money, and they can afford it because electricity in China is cheap and chips are subsidized by the state."