Chinese AI Models Are Still Finding Their Way Onto Banned Nvidia Chips, New Reporting Says

Chinese frontier AI labs have still been training leading models on Nvidia chips that Washington has banned from export to China, according to new reporting and research published by TechCrunch and The Straits Times. The findings matter because they suggest US chip controls remain porous even as Beijing races to narrow the gap with American AI leaders.

What the new reporting says

TechCrunch reported that research found training fingerprints consistent with Nvidia’s export-banned Blackwell GPUs in several top Chinese open-weight models, including Moonshot AI’s Kimi K3 releases and DeepSeek V4, along with other open models. The Straits Times, in a separate report, said the broader picture shows how Chinese labs are still obtaining high-end Nvidia hardware through routes that sidestep formal export rules.

Those claims sit against a tightening and loosening policy mix in Washington. The Straits Times said the US has banned Blackwell and Rubin GPUs from sale to China on security grounds. TechCrunch added that the Trump administration later allowed H200 exports only if they are shipped through a circuitous route that triggers a 25% tariff, a policy that illustrates how controls have become more conditional rather than absolute.

Why it matters

This is not just a trade compliance story. If Chinese AI developers can still access top Nvidia hardware, then the US may be slowing, but not stopping, the flow of compute that powers frontier model training. That has consequences for national security, for the commercial AI race, and for companies trying to forecast where the next generation of powerful models will emerge.

The reporting also points to a second-order effect: restrictions on direct shipments appear to be pushing demand into offshore data centers and third-country arrangements. That means the chip may never cross the border in the usual way, but the compute still reaches Chinese customers through rented infrastructure, intermediaries, or concealment schemes.

How could this still be happening?

According to the reports, there are several possible channels. The Straits Times said Chinese labs may be reaching Nvidia chips through smuggling or through data centers in third countries, especially in Southeast Asia. TechCrunch said offshore data centers, notably in Malaysia, have become a major workaround for compute access.

Epoch AI estimates cited by TechCrunch suggest the scale can be large. The research group estimates about 2 million advanced chips entered China in 2024 and 2025 through concealment schemes. It also estimated that 275,000 Nvidia AI accelerators were imported to Malaysia in the same period, highlighting how regional compute hubs can become part of a broader supply-chain detour.

Who is being singled out?

Moonshot AI and DeepSeek are central to the new scrutiny, with The Straits Times saying recent launches by Moonshot AI, Z.ai and DeepSeek intensified attention on where Chinese model developers are sourcing compute. TechCrunch said White House adviser Michael Kratsios stated that Moonshot used export-banned Nvidia chips, citing US intelligence. The Straits Times reported that Kratsios said Moonshot accessed high-end Nvidia chips in Thailand.

Those claims do not by themselves settle whether the models were trained entirely on banned chips or only partly. That remains one of the most important open questions, because even partial use of restricted hardware could complicate both enforcement and corporate liability.

“US export controls restrict where a chip is shipped, not who uses it”

The quote above, attributed by The Straits Times to Stephen Wu, founder of Carthage Capital, captures the enforcement problem at the center of the story. A controlled shipment can still end up powering work for a buyer in China if the chip is rerouted, rented, resold or used through a foreign data center.

Background: how the chip squeeze developed

The US export-control regime grew out of fears that advanced AI hardware could help China’s military modernization and its domestic AI industry at the same time. Nvidia’s GPUs have become the essential tool for training large language models, which is why Washington has treated access to that hardware as strategically sensitive.

According to the reporting, the policy response has changed over time, but the underlying contest has not. TechCrunch said Blackwell GPUs are banned from export to China or Chinese-owned firms. The Straits Times said Washington also banned Blackwell and Rubin GPUs to China on security grounds. Yet even with those restrictions in place, the reports describe a market that adapts quickly around them.

The adaptation includes offshore capacity. Rather than importing chips directly into China, customers can rent compute from data centers in another country. That model is especially relevant in Southeast Asia, where regional infrastructure can serve Chinese customers without making the final user immediately visible in customs records.

What the company and market reaction looks like

TechCrunch said Nvidia and Jensen Huang have vouched for Megaspeed data centers as legitimate, even as the company and its subsidiary Speedmatrix have been investigated by US and Singaporean authorities. The same report said Bloomberg reported that a company linked to Alice Huang helped smuggle banned Nvidia chips to Shanghai.

That mix of denials, investigations and reported links shows how quickly the issue moves from abstract policy to corporate scrutiny. It also places Nvidia in a difficult position: its chips remain the industry benchmark, but the company is also navigating a market where the same hardware can become part of a sanctions story.

The Straits Times noted Nvidia’s scale in stark terms, citing a market value of US$4.5 trillion, or S$5.75 trillion. That valuation helps explain why the company remains central to every debate over AI compute, even when the debate is about shipping routes rather than model architecture.

Why it matters for China’s AI race

The direct implication is that Chinese AI developers may not be as cut off from the best available training hardware as US policymakers hoped. The indirect implication is that domestic chipmakers have a longer runway before they are forced to replace Nvidia entirely.

The Straits Times said Chinese-made chips are improving fast, with Huawei leading the domestic race. But it also said Chinese chips still need more time, engineering, power and state support to match Nvidia. That gap matters because training frontier models is not just about having a chip; it is about having enough of the right chips, enough power, and enough software compatibility to make the whole stack work.

The same report quoted Omdia chief analyst Lian Jye Su as saying, “Chinese vendors are still using cutting-edge Nvidia chips to train their models.” That underlines the core tension: China is building its own supply base, but it has not yet escaped dependence on the outside one.

By the numbers: what the reports say

Several figures in the reporting help show the scale of the problem and the policy response.

• 2 million advanced chips are estimated by Epoch AI to have entered China in 2024-2025 through concealment schemes.
• 275,000 Nvidia AI accelerators were estimated to have been imported to Malaysia in 2024-2025.
• 25% is the sales cut or tariff tied to the eased H200 export policy described in the reporting.
• Q1 2027 is Huawei’s target for its next-generation AI chip readiness, according to The Straits Times.
• September was the month Huawei said its next-gen AI chip would be ready earlier than planned, according to The Straits Times.

What investigators and policymakers are watching

According to the fact sheet, investigators are looking at several companies and several routes. Megaspeed and its subsidiary Speedmatrix have been investigated by US and Singaporean authorities. A company linked to Alice Huang was reported by Bloomberg to have helped smuggle banned Nvidia chips to Shanghai. Meanwhile, the White House adviser Michael Kratsios has publicly linked Moonshot AI to export-banned Nvidia chips in statements cited by the reports.

Policy makers are also facing a timing problem. The US eased H200 sales in 2026 but kept a 25% cut of sales, while the Trump administration’s route-based H200 policy still leaves room for workarounds. In practice, that means some less powerful Nvidia hardware can still reach the market, even as the top-end chips remain restricted.

That creates a partial-containment regime rather than a clean cutoff. If the goal is to prevent China from getting the hardware most useful for frontier training, the reporting suggests the controls have not yet achieved that objective.

What happens next

The next developments will likely come from investigations rather than product launches. The key questions are whether the training described by TechCrunch relied entirely on banned Nvidia chips or on a mix of hardware sources, whether any laws were broken, and how the chips actually moved across borders or into offshore data centers.

Another issue is whether Singapore, Thailand and Malaysia become more active enforcement frontiers. The reporting points to those places not as end markets but as transit points and compute hubs, which makes them central to any attempt to tighten controls.

Finally, the long-term competitive picture will depend on China’s domestic chip progress. Huawei’s timeline suggests a faster pace of development, but the Straits Times said domestic hardware still needs more time and support before it can match Nvidia at the frontier. Until then, the market for banned or indirectly accessed Nvidia chips is likely to keep shaping China’s AI output.

Timeline: how the story developed

Bottom line

The reporting points to a simple but uncomfortable conclusion for US policymakers: banning the sale of the most advanced Nvidia chips has not prevented Chinese AI labs from using them. The difference between a forbidden shipment and a usable compute supply is now often a matter of geography, intermediaries and enforcement capacity.