Nvidia GPUs have become central to training and running AI systems because they can process enormous volumes of data, and demand surged after ChatGPT launched in 2022.

The AI investment boom has helped push Nvidia’s market capitalization to $4.5 trillion, underscoring how valuable access to its chips has become as companies and governments race to build advanced AI.

US export restrictions bar Nvidia’s top-end Blackwell and Rubin GPUs from China, but Chinese AI firms can still access advanced Nvidia computing through overseas data centres or, according to allegations, smuggled hardware.

The Information reported that DeepSeek trained a new model using thousands of chips dismantled overseas and brought into China; DeepSeek did not comment, while Nvidia said it had seen no evidence and called the account far-fetched.

Remote access offers another route because US controls regulate where hardware is shipped, not necessarily who uses it; examples cited include data centres in Malaysia’s Johor region and a White House adviser’s claim that Moonshot AI accessed high-end Nvidia chips in Thailand.

Chinese-made chips are improving, with Huawei the leading domestic competitor, but they still trail Nvidia, and using them typically takes more engineering time and may require more chips to match performance.

Those disadvantages can be manageable in China, where electricity is cheap and chips are subsidised by the state, while Nvidia hardware can help Chinese AI companies such as Moonshot AI, Z. ai and DeepSeek train models faster.

DeepSeek reportedly plans to train its next major model on Huawei chips, and Huawei has said its next-generation AI chip is expected in the first quarter of 2027.