Evidence that leading Chinese artificial intelligence developers may have trained some of their most advanced models on Nvidia’s restricted Blackwell processors is raising fresh questions about the effectiveness of US semiconductor export controls and the extent to which Chinese companies can access American computing technology through overseas infrastructure.
A research effort led by Will DePue, a former OpenAI researcher who heads nonprofit research group Unemployment Labs, found what he described as compelling evidence of Blackwell hardware fingerprints in the model weights released by several Chinese AI developers. The analysis examined models from DeepSeek, Moonshot AI, MiniMax and Ant Group, including Moonshot’s Kimi K3 releases and DeepSeek’s V4 models.
The findings indicate that Chinese AI laboratories may have obtained access to computing capacity built around processors that Washington has prohibited from being exported to China or Chinese-owned companies. They do not establish that every model was trained entirely on Blackwell GPUs, identify the physical location of the processors, or prove that export laws were violated.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Chinese companies have found ways to access advanced computing infrastructure outside mainland China. Overseas data centers, particularly in Southeast Asia, have become an alternative route for obtaining computing power without importing restricted processors directly into China.
The research nevertheless presents a difficult question for Washington: if Chinese developers can train advanced models using restricted American chips located outside China, export controls based primarily on the destination of hardware may not prevent the transfer of computing capabilities the restrictions are intended to limit.
The findings also arrive amid a wider controversy surrounding Nvidia and its former chief executive, Jensen Huang, whose departure has followed reports about his personal and business relationships with Alice Huang, an executive at Singapore-based cloud computing company Megaspeed. The company and its subsidiary, Speedmatrix, have faced scrutiny over alleged shipments of restricted Nvidia processors into China.
Offshore Data Centers Expose Gaps in US Chip Restrictions
The US export control framework is designed to preserve America’s advantage in advanced computing by restricting China’s access to the most powerful AI processors. Nvidia’s Blackwell GPUs are among the products barred from export to China and Chinese-owned companies. The Trump administration has allowed sales of the less powerful H200 under conditions that include a circuitous shipping arrangement and a 25% tariff.
However, restrictions on direct shipments do not necessarily prevent Chinese developers from renting computing capacity in countries where the processors can legally be installed. A company can obtain access to advanced chips through an overseas data center without necessarily owning or importing the hardware itself.
Epoch AI, a nonprofit research organization, estimates that approximately two million advanced chips entered China in 2024 and 2025 through schemes designed to conceal their final destination. Its data also estimates that 275,000 Nvidia AI accelerators were imported into Malaysia during the same period, revealing the scale of Southeast Asia’s role in the regional computing supply chain.
The distinction between chip ownership and access to computing power has been key to the policy debate. Washington can restrict the sale of a processor to a particular customer or destination, but monitoring how computing capacity is subsequently rented, accessed, and used presents a more complicated enforcement challenge.
Sam Bresnick, a researcher at Georgetown University’s Center for Security and Emerging Technology, questioned the logic of allowing Chinese companies to access restricted computing capacity simply by operating outside mainland China.
“If we care about the U.S. tech edge, it doesn’t make any sense to me that you would not allow Chinese companies to access chips in mainland China or Hong Kong and Macau, but then say, actually, if you just go over to Malaysia, you can do whatever you want,” he told TechCrunch earlier this summer.
The debate is moving beyond the physical location of processors towards the identity of the end users and the purposes for which computing resources are deployed. China-focused policymakers have argued that restrictions should cover access to controlled chips, not only their ownership or installation in particular countries.
White House technology adviser Michael Kratsios said earlier this year that Moonshot had used export-banned Nvidia chips, citing US intelligence sources. The new research offers a separate line of evidence based on the technical characteristics of released AI models, although it does not independently establish the origin or legality of the hardware used.
The findings come with additional weight because of the scrutiny surrounding Megaspeed, a company led by Alice Huang, whose relationship with Jensen Huang has become part of a broader investigation into Nvidia’s exposure to China-related chip restrictions.
Megaspeed has been at the forefront of concerns about the movement of advanced Nvidia processors through its subsidiary Speedmatrix. Both US and Singaporean authorities have investigated the company, according to the supplied reports. Bloomberg also reported on Friday that a company linked to Alice Huang was involved in smuggling restricted Nvidia chips to a data center in Shanghai.
These remain allegations and reported investigative findings, not established proof of wrongdoing by Nvidia or its former chief executive.
In May, short-selling research firm Culper Research published a report arguing that Nvidia had a significant China exposure because Chinese demand was reaching the company through both overseas data centers and alleged smuggling networks. The report examined Megaspeed’s relationship with Nvidia and alleged that companies associated with Alice Huang had backing from Chinese entities barred from accessing US chips.
Nvidia and Huang had vouched for Megaspeed’s data centers, saying company officials had visited the facilities to verify that they were legitimate. If investigators establish that restricted processors were diverted to prohibited users and that Nvidia executives knew about the activity, the company could face serious questions about its compliance procedures and potential corporate liability.
Huang’s personal and professional connections with Alice Huang could intensify that scrutiny, but the existence of a relationship does not mean that he knew about or facilitated illegal shipments. The major issue for investigators is what Nvidia knew about the customers, the end users, and the movement of the processors.
The controversy also puts pressure on US export enforcement. Commerce Secretary Howard Lutnick, who oversees agencies responsible for export licensing, has faced scrutiny over reports that investigators failed to identify suspicious buyers and maintained overly close relationships with Nvidia officials. Reports have suggested his position could be at risk, although no formal outcome has been established.
China’s Domestic Chip Ambitions Complicate Nvidia’s Outlook
Culper’s investment thesis also focused on the possibility that China’s response to US restrictions could ultimately hurt Nvidia more than any direct enforcement action.
The Chinese government has encouraged domestic semiconductor development as part of a broader effort to reduce reliance on American technology. If local manufacturers can produce competitive AI processors at scale, Chinese demand for Nvidia products could decline, removing a potentially significant market for the US company.
The question matters to reports that approximately 600,000 Nvidia H200 chips are stockpiled in Hong Kong awaiting possible shipment into China. Although Chinese companies have an appetite for advanced computing capacity, observers have offered different explanations for why the processors have not been cleared for import.
Stanford research scholar Graham Webster, who focuses on US-China technology policy, questioned the argument that withholding the chips would necessarily accelerate domestic production, given the scale of Chinese demand for computing power. Other explanations include the possibility that the stockpile is connected to preparations for a broader trade arrangement involving US goods, or concerns in Beijing about potential surveillance technology embedded in the processors.
The competing explanations finger the uncertainty surrounding China’s semiconductor policy. Beijing must balance its need for advanced computing capacity against its objective of developing domestic alternatives and reducing exposure to US export restrictions.
The risks are therefore coming from two directions for Nvidia. Analysts have warned that US investigations could tighten controls on customers, intermediaries and overseas data centers, while China’s industrial policy could gradually reduce the market available to American chipmakers.