Smuggled Chips and Server Loopholes: How China Gets Nvidia GPUs

Chinese AI developers are still training their models on Nvidia's most advanced chips despite a U.S. export regime that bans the company's top-end Blackwell and Rubin GPUs from reaching China, according to Omdia chief analyst Lian Jye Su. "Chinese vendors are still using cutting-edge Nvidia chips to train their models" despite the controls, Su told AFP.

The admission from a senior semiconductor analyst sits at the center of a story that now spans three continents: Washington's export machinery, third-country data centers in Southeast Asia, alleged smuggling networks, and a fast-improving domestic Chinese chip industry led by Huawei. The stakes are straightforward. Whoever controls the supply of advanced AI compute controls the pace at which frontier models get built, and Washington has made that control explicit policy.

Why do Nvidia chips matter so much?

Nvidia designs Graphics Processing Units, components originally built to render 3D visuals in video games. Those same architectures crunch colossal volumes of digital data, which makes them the workhorse hardware for training and running artificial intelligence systems.

The launch of ChatGPT in 2022 demonstrated the commercial potential of large AI models and triggered an investment race among governments and technology giants. Chip orders poured in. Nvidia's market valuation has since climbed to $4.5 trillion, placing it on par with Apple as one of the world's two most valuable companies.

That valuation makes Nvidia the single most exposed company to U.S.-China technology policy. Its chief executive has lobbied the U.S. government directly, warning that the restrictions are fuelling the rise of competitive homegrown Chinese alternatives to Nvidia silicon.

Which chips does Washington restrict?

The U.S. government bans exports of Nvidia's top-end Blackwell and Rubin GPUs to China, citing security concerns, on the grounds that advanced AI has military applications. The policy also serves a commercial aim: keeping the United States' top AI laboratories ahead of their global rivals.

This year Washington eased controls on a less powerful chip, the previously banned H200, but attached an unusual condition. The United States takes a 25 percent cut of H200 sales under the eased regime described in the report.

The arrangement has not produced a flood of orders. U.S. conditions have made shipments difficult to approve, and Beijing is reportedly reluctant to buy H200s as its own chip industry gains momentum.

How does China actually get the hardware?

Two mechanisms dominate: third-country compute rentals and smuggling.

Recent model launches from Chinese startups including Moonshot AI, Z.ai and DeepSeek have turned up competitive pressure on U.S. rivals, and those launches were built on hardware that, in many cases, was never legally exported to China. Chips can be smuggled into the country. Alternatively, Chinese firms exploit a loophole that lets them rent computing power at data centers packed with Nvidia chips in a third nation, often in Southeast Asia.

"US export controls restrict where a chip is shipped, not who uses it," explained Stephen Wu, founder of the Carthage Capital investment fund. That means Nvidia GPUs can sit in Malaysia's Johor region, for example, "in full compliance while a customer in Shenzhen logs in," Wu told AFP.

The White House has named names. White House adviser Michael Kratsios said this summer that Moonshot AI had accessed high-end Nvidia chips in Thailand.

Washington has also accused Chinese AI laboratories of training systems to mimic the capabilities of U.S. models.

Then there is the smuggling question. Tech outlet The Information reported in December that DeepSeek developed its new model using thousands of chips dismantled overseas and brought into China. DeepSeek did not respond to a request for comment at the time. Nvidia told The Information it had not seen evidence of such smuggling, calling the claim seemingly far-fetched.

What about the personal scandal surrounding Nvidia's CEO?

Nvidia CEO Jensen Huang has admitted to an affair, reportedly with a Chinese executive at a Singapore-based company that the United States suspects of smuggling AI microchips into China. The admission has drawn fresh scrutiny to the exact question regulators have been asking for years: how advanced silicon reaches Chinese customers despite the export regime. Huang himself has been the most prominent corporate advocate for easing chip sales to China, arguing the restrictions hand the Chinese market to local competitors.

Can Chinese chips replace Nvidia?

The domestic alternative is improving, but it is not free. Huawei is China's leading Nvidia competitor, though it still lags the U.S. company. DeepSeek will reportedly train its next major AI model on Huawei chips.

"Nvidia chips still make top-level AI work faster and easier, but they are no longer the only path," said Poe Zhao, founder of the analysis publication Hello China Tech. "The open question is cost," because "training on Chinese chips usually takes more time and more engineering."

"Chinese-made chips are improving fast," Su said, adding that more collaborations between Chinese AI firms and local chipmakers are likely.

Huawei set a concrete timeline in September: its next-generation AI chip will be ready in the first quarter of 2027, three quarters earlier than planned.

To match the power of Nvidia hardware, AI companies compensate by deploying more Chinese chips. "They pay for it in power and cost — which they can afford, because Chinese electricity is cheap and the chips are state-subsidised," Wu said.

What happens next?

The squeeze works in both directions. Washington can tighten enforcement around third-country data centers and smuggling routes, but every restriction it imposes accelerates the shift toward Huawei and other domestic suppliers that Beijing subsidizes. If Huawei hits its first-quarter 2027 target, Chinese labs will face a genuine choice between smuggled Nvidia performance and state-backed domestic compute — and the smuggling pipelines that Omdia says are still in use may start to matter less.