Nvidia Blackwell fingerprints have been detected in six Chinese AI models released since April 2026, according to a forensic analysis led by Will DePue, a former OpenAI researcher. DePue, who leads the nonprofit research group Unemployment Labs, said the findings provide the first technical evidence that leading Chinese AI labs have had access to Nvidia’s latest hardware despite United States export controls.
Nvidia Blackwell fingerprints: Evidence from published model weights
DePue analyzed the weights of large language models released openly by DeepSeek, Moonshot AI, MiniMax and Ant Group, according to The Decoder. The analysis found Blackwell signatures in Moonshot’s Kimi K3 releases and DeepSeek’s V4 models. Earlier models, including DeepSeek-V3 and Kimi K2.5, instead showed traces of Nvidia’s older Hopper generation.
The technique examines statistical patterns left in the numerical weights that define an AI model. Before training, PyTorch, the standard software framework, fills empty weight slots with random numbers generated from a starting seed. The GPU distributes the work across its compute cores, and different chip models have different core counts. That process leaves a characteristic pattern in the initial values.
Training adjusts the values millions of times. However, some slots associated with unused words or symbols retain their initial random values, allowing a faint statistical echo of the original distribution to survive, DePue said. He reported that he worked the pattern backward and reproduced the results for DeepSeek-V4-Flash and Kimi K3 exactly on real Blackwell chips. DePue also said senior researchers at OpenAI and Anthropic independently checked and confirmed his findings, and he published the verification code on GitHub.
Six models and two Blackwell patterns
Older Hopper chips, including the H100 and H800, use 132 compute cores. Blackwell chips use exactly 148 or 152 cores, depending on the model. The B200 and B300 have 148 cores, while the liquid-cooled GB200 and GB300 flagship rack systems have 152.
According to DePue’s analysis, DeepSeek-V4-Flash and MiniMax-M3 carry the 148-core pattern associated with B200 or B300 chips. DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ant Group’s Ling-3.1-flash carry the 152-core pattern associated with GB200 or GB300 systems. DePue said statistical testing ruled out chance and that his method correctly identified all 24 control models with known hardware, including Meta’s Llama 3.1 on H100.
Models from other Chinese laboratories, including Alibaba’s Qwen3.8 and systems from Zhipu, Tencent, Xiaomi and StepFun, showed only Hopper signatures. Meituan’s LongCat-2.0 pointed to Huawei hardware. The findings also align with separate analysis from Dylan Patel, chief analyst at SemiAnalysis, who told Sequoia Capital that DeepSeek-V3’s architecture remained tailored for Hopper chips, while DeepSeek-V4 was built for Blackwell GPUs and Huawei’s Ascend chips.
A major claimed training footprint
DePue’s report estimates that the five training runs examined required about 12,000 Blackwell GPUs combined. That represents roughly the computing power of 30,000 older H100 chips. Kimi K3 alone required an estimated 4,500 Blackwell processors over about 70 days, while DeepSeek-V4-Pro required about 3,000.
At a unit price of about $70,000, each of those clusters would cost $200 million to $300 million, according to DePue’s report. All five systems together would cost more than half a billion dollars. These are estimates based on the model analysis rather than disclosed procurement records.
The claimed scale matters because it suggests that restricted access was not limited to a small experimental cluster. The Decoder reported that DePue considers the analysis the first technical evidence of leading Chinese labs gaining access to Blackwell hardware. TechCrunch noted that US officials and news outlets had previously suspected DeepSeek and Moonshot AI were using Blackwell chips but had not published proof.
What the method can and cannot establish
Strictly speaking, the method identifies the hardware that initialized the model weights. The training itself very likely ran on the same systems, but the evidence does not reveal where the GPUs were located or who owned them. Chinese companies can rent computing capacity at data centres outside China, a route that TechCrunch reported has driven an industrial revolution in Malaysia.
Consequently, the findings indicate access to Nvidia’s top hardware but do not by themselves prove physical smuggling or a violation of United States export law. The only other technical explanation would be an AMD MI300X, which is also restricted and artificially throttled to produce the same pattern. DePue said he ruled out AMD for DeepSeek-V4-Flash by using tiny rounding differences between the chips.
Blackwell GPUs are banned from export to China or Chinese-owned firms. The Trump administration permits Nvidia to export the less powerful H200 only when shipments take a circuitous route that allows the United States to claim a 25 percent tariff. White House technology adviser Michael Kratsios said earlier in 2026 that Moonshot had used export-banned Nvidia chips, citing United States intelligence sources.
Nvidia’s wider China exposure
Epoch AI estimates that about two million advanced chips entered China in 2024 and 2025 through schemes designed to conceal their final destination. Epoch’s data also estimates that 275,000 Nvidia AI accelerators were imported to Malaysia in 2024 and 2025, illustrating the offshore route available to Chinese firms seeking computing capacity.
Reuters examined more than 150 procurement notices, investment plans and corporate records and identified Chinese AI data centre projects worth over 23 billion yuan, or $3.2 billion. Many specify Nvidia’s H100, H200 or H800 chips. Decrypt reported that the total project value includes buildings, electricity and other infrastructure. The documents do not establish that every processor was obtained or that any export law was broken.
An IDC report cited by Reuters estimated that Nvidia represented 34 percent of China’s AI accelerator shipments in the first half of 2026, compared with 28 percent for Huawei. Those figures cover the broader chip market rather than confirmed shipments of restricted models. The United States began restricting advanced AI chip sales to China in 2022. In January 2026, the Commerce Department moved H200 export applications to case-by-case review, although sales still require licences.
Corporate scrutiny and export enforcement
TechCrunch reported that Nvidia CEO Jensen Huang left the company this week after admitting to his affair with Alice Huang, an executive of Megaspeed. Megaspeed has been implicated in smuggling Nvidia chips into China and has also been alleged to be an agent of China’s intelligence service. The company has been central to the trade through its subsidiary Speedmatrix and is subject to investigations by both the United States and Singaporean governments.
TechCrunch said Nvidia’s potential corporate liability over export control violations may have factored into its decision to part ways with Huang. The report also noted that US companies have limited obligations to track what customers do with controlled AI chips. If investigators confirm chip smuggling and Huang’s knowledge of it, that relationship could create a more serious problem for Nvidia.
The analysis therefore adds a technical dimension to the dispute over Nvidia Blackwell fingerprints. It supports the claim that Chinese laboratories accessed the newest generation of AI accelerators, but it does not settle the location, ownership or legality of that access. Its importance lies in showing how export restrictions designed to constrain hardware ownership may also need to address where computing capacity is rented and which systems initialize frontier model weights.