Nvidia chief Jensen Huang admits affair as China's chip routes draw scrutiny
Nvidia is worth $4.5 trillion, on a par with Apple as one of the world's two most valuable companies, and the chips behind that valuation are banned from export to China. Yet Chinese AI labs keep releasing models that compete with US ones. How they get the computing power is the question the affair admitted by Nvidia chief executive Jensen Huang has pushed back into the open.
The GPUs Nvidia designs were built to render 3D graphics in video games, but their capacity to crunch vast volumes of data made them the standard engine for training and running artificial intelligence after ChatGPT launched in 2022. Washington blocks exports of Nvidia's top-end Blackwell and Rubin chips to China on security grounds, citing advanced AI's military applications, and to keep US labs ahead. This year it eased controls on the less powerful H200, with the US taking a 25 percent cut of sales, but conditions have made approvals difficult and Beijing is reportedly reluctant to buy.
Chips still reach Chinese users by more than one route. Lian Jye Su, chief analyst at Omdia, said Chinese vendors "are still using cutting-edge Nvidia chips to train their models" despite the controls. Stephen Wu, founder of the Carthage Capital investment fund, pointed to the gap in the rules: "US export controls restrict where a chip is shipped, not who uses it." A GPU can sit in Malaysia's Johor region "in full compliance while a customer in Shenzhen logs in", he said. White House adviser Michael Kratsios said this summer that Moonshot AI had accessed high-end Nvidia chips in Thailand, and according to Token Times, DeepSeek developed its new model using thousands of chips dismantled overseas and brought into China. Nvidia said it had not seen evidence of such smuggling, which it called "seems far-fetched".
Huang has lobbied Washington to allow sales to China, warning the restrictions are feeding homegrown rivals. Huawei is the leading one, and said in September its next-generation AI chip would be ready in the first quarter of 2027, three quarters earlier than planned; DeepSeek will reportedly train its next major model on Huawei silicon. Poe Zhao, founder of analysis publication Hello China Tech, said Nvidia chips "still make top-level AI work faster and easier, but they are no longer the only path", with cost the open question, since training on Chinese chips takes more time and engineering. Wu's arithmetic: more chips means more power and cost, "which they can afford, because Chinese electricity is cheap and the chips are state-subsidised".
