A nonprofit research group Unemployment Labs and independent analysts say several leading Chinese companies’ open-weight language models show signs consistent with training on Nvidia Blackwell GPUs — a family of accelerators that U.S. export rules ban from direct shipment to China or Chinese-owned entities.
Will DePue, who leads Unemployment Labs and is a former OpenAI researcher, analyzed the weight structures of models published openly by DeepSeek, Moonshot AI, MiniMax, and Ant Group. DePue reports that the most closely watched releases — including Moonshot’s Kimi K3 series and DeepSeek’s V4 models — carry configuration signatures that match training on Blackwell-class hardware.
Why this matters for export policy and AI competition
U.S. export controls aim to protect American leadership in AI by restricting access to the most powerful computing hardware. Blackwell GPUs are prohibited for export to China, while a less powerful Nvidia part, the H200, has been allowed under constrained conditions.
Epoch AI, a nonprofit research organization, estimates roughly two million advanced AI accelerators reached China in 2024 and 2025 through schemes designed to obscure their final destination. Epoch also estimates about 275,000 Nvidia AI accelerators were imported into Malaysia in 2024–2025, where offshore data centers have become a common route for Chinese firms to obtain compute capacity.
Those patterns complicate the enforcement objective: shipping channeled through third countries or offshore facilities can allow Chinese labs to train on hardware U.S. policy intended to block.
Offshore training, legal ambiguity and differing interpretations
The technical evidence that model weights reflect Blackwell training does not, on its own, prove that U.S. export laws were violated or that the models were trained exclusively on Blackwell chips. Hardware could have been used in offshore data centers or routed through intermediary jurisdictions. Still, such circumvention runs contrary to the export bans’ stated purpose of denying the most capable accelerators to Chinese AI developers.
Experts debate the effectiveness of existing controls. Sam Bresnick at the Georgetown Center for Security and Emerging Technology argued that restricting only ownership or installation location is insufficient if the same hardware can be legally or illicitly accessed in nearby jurisdictions.
Corporate and personal entanglements under scrutiny
Megaspeed, led by Alice Huang, and its Speedmatrix subsidiary have been implicated in the contested trade for several years and have drawn investigations by U.S. and Singaporean authorities. Bloomberg reported that an Alice Huang–linked company was allegedly involved in smuggling banned Nvidia chips to a Shanghai data center.
This controversy intersected with a personnel development at Nvidia: Jensen Huang, Nvidia’s chief executive, left the company this week after acknowledging a personal relationship with Alice Huang. Alice Huang has been alleged in some reports to be an agent of China’s intelligence services. While media attention has focused on Jensen Huang’s personal ties, Nvidia’s potential corporate exposure to export control violations likely influenced the company’s decision-making.
Nvidia and Jensen Huang have said that Nvidia representatives visited Megaspeed’s data centers and deemed them legitimate.
Market consequences and unanswered questions
A May report from short-selling research firm Culper argued that Nvidia faces a “China problem,” noting that a significant portion of its revenue is driven by Chinese demand — whether through remote data centers or, as Culper alleged, illegal channels. The report outlined Megaspeed’s ties to Nvidia and suggested that Alice Huang’s businesses are backed by Chinese entities barred from accessing U.S. chips.
Analysts note several open questions. Some 600,000 H200 chips have reportedly been stockpiled in Hong Kong awaiting shipment into China; observers disagree on why those shipments have not been authorized, offering hypotheses that range from preparatory bargaining for trade deals to concerns about potential surveillance capabilities in the hardware. Graham Webster, a Stanford research scholar focused on U.S.–China tech policy, has argued that holding chips back alone is unlikely to spur domestic production given the scale of demand.
Oversight, enforcement and potential fallout
Currently, U.S. firms face limited obligations to track exactly how buyers use controlled goods such as AI accelerators. That regulatory gap becomes more consequential if investigations substantiate allegations of systematic smuggling and show awareness by corporate leaders. In that case Nvidia could face legal and reputational consequences beyond the immediate personnel changes.
The agencies that approve export licenses are also under scrutiny. Reports suggest Howard Lutnick, the Secretary of Commerce overseeing those processes, could lose his position after allegations that investigators failed to detect suspicious buyers and were too close to Nvidia officials.
Conclusion
The weight-analysis findings and related reporting point to a potentially significant issue: access to banned Nvidia Blackwell chips may have contributed to the training of prominent open-weight Chinese models. Ongoing government investigations and corporate disclosures will be required to determine whether export controls were violated, whether smuggling occurred on a scale alleged by some organizations, and what legal or policy changes should follow.
