Analysis Finds Blackwell Signatures in Leading Chinese AI Models

Open-weight AI models released by Chinese companies including DeepSeek, Moonshot AI, MiniMax and Ant Group show evidence consistent with training on Nvidia Blackwell GPUs, which are restricted from export to China and Chinese-owned firms. The finding comes from Will DePue, former OpenAI researcher and head of the nonprofit Unemployment Labs, who examined the published weights of large language models.
DePue reported compelling indications that Moonshot’s Kimi K3 releases and DeepSeek’s V4 models retain hardware-related signatures associated with Blackwell configurations. The analysis adds technical detail to an earlier assertion by White House technology adviser Michael Kratsios that Moonshot had used export-banned Nvidia chips, citing US intelligence sources.
Weights can reveal training infrastructure
Model weights are the numerical parameters that encode a trained model’s structure. DePue’s approach rests on the premise that those weights can still reflect aspects of the chip configurations used during training. The reported evidence does not establish that the models were trained entirely on Nvidia hardware, nor does it by itself prove an export-control violation.
One potential route is offshore computing. Chinese firms may access advanced compute through data centres outside mainland China, an arrangement that has helped fuel rapid data-centre development in Malaysia. Epoch AI estimates that roughly two million advanced chips entered China in 2024 and 2025 through schemes intended to conceal their final destination, while its figures put imports of Nvidia AI accelerators to Malaysia at 275,000 during the same period.
Access, rather than ownership, is the policy challenge
The apparent availability of Blackwell-class compute to Chinese frontier labs raises questions about whether export rules focused on ownership, destination and installation location can achieve their intended effect. Blackwell GPUs are barred from export to China or Chinese-owned companies, while the Trump administration permits exports of the less powerful H200 through a shipping route subject to a 25% tariff.
Sam Bresnick of Georgetown’s Center for Security and Emerging Technology has argued that restrictions lose coherence if Chinese companies cannot use chips in mainland China, Hong Kong or Macau but can access equivalent capacity elsewhere. The policy issue is not limited to physical shipments: it also concerns who can operate the compute and for what purpose.
Scrutiny has also centred on Megaspeed and its subsidiary Speedmatrix, which have been investigated by US and Singaporean authorities. Nvidia and Jensen Huang have vouched for Megaspeed’s data centres, saying Nvidia officials visited them to confirm their legitimacy. The concern around those relationships is sharpened because allegations involving Megaspeed leadership outlines allegations involving Megaspeed’s leadership and possible surveillance of Nvidia’s chief executive.
What organisations should examine
For businesses that buy, host or provide advanced AI infrastructure, the episode underlines the need to examine more than the immediate buyer and shipping address. A practical compliance review should map the end user, the operating location of data-centre capacity, the entities controlling workloads and any intermediaries involved. That broader view can help organisations identify where access to controlled compute may diverge from the formal destination of the hardware.

