The Mystery of 445 Rows: Nvidia Blackwell Chip Traces Found in Six Chinese AI Models, Including DeepSeek and Ant Group

A new technical analysis has identified numerical patterns consistent with Nvidia‘s advanced Blackwell processors in six Artificial Intelligence models developed by four Chinese companies, including DeepSeek, Moonshot AI, MiniMax and Ant Group. The findings have raised questions about how Chinese AI developers may have accessed computing hardware restricted under US export controls.

The research, published on November 12, was conducted by Will DePue, a researcher at Unemployment Labs. His analysis focused on numerical patterns embedded in publicly available AI model weights. Although the results provide technical evidence linking certain models to Blackwell hardware, they do not establish where the processors were used, who owned them or whether any export regulations were violated.

Key Highlights

• Researcher Will DePue identified Blackwell-compatible numerical patterns in six Chinese AI models.
• Models from DeepSeek, Moonshot AI, MiniMax and Ant Group were included in the analysis.
• In DeepSeek-V4-Flash, 445 out of 469 selected rows matched exactly in a hardware reproduction test.
• The findings do not establish whether restricted Nvidia processors were used inside mainland China or whether export rules were breached.
• The research emerged amid scrutiny of Nvidia’s leadership transition and ongoing questions about access to advanced AI computing hardware.

What Did the 445 Rows Reveal?

The central finding involves DeepSeek-V4-Flash, one of the AI models examined by DePue. He recovered the starting number, known as a seed, used to generate the model’s initial weights. He then ran the same software configuration on an Nvidia Blackwell B300 processor to compare the resulting numerical patterns.

Out of 469 selected rows, 461 showed correlations above 0.99. More notably, 445 rows matched exactly. When the same test was performed using an older Nvidia H100 processor, the hardware did not reproduce those rows.

The unusually close match formed the basis of the researcher’s claim that the model contained numerical fingerprints consistent with Blackwell hardware.

However, the finding should not be interpreted as direct proof that a particular Blackwell processor was physically used by the Chinese developers. The analysis identifies similarities in computational outputs, not the location, ownership or supply route of the hardware.

How Can AI Model Weights Reveal the Chips Used?

AI models rely on numerical weights that determine how they process information and generate responses. Before training begins, software creates an initial set of numbers using a starting value called a seed.

The way these numbers are generated can vary depending on the hardware and software configuration. Different processor architectures or computing environments may produce distinct numerical arrangements, creating patterns that researchers can potentially identify later.

Some of these initial values can survive the training process. By examining publicly released model weights and comparing their numerical patterns with outputs generated on known hardware, researchers can look for clues about the computing systems involved in model development.

DePue said his technique was consistent with publicly disclosed hardware information for 24 models. However, the research also acknowledged that specially configured AMD processors could produce similar patterns.

According to DePue, additional tests ruled out that alternative explanation for DeepSeek-V4-Flash, but the possibility was not conclusively eliminated for the other five models. This distinction is important when assessing how broadly the findings can be applied.

Why DeepSeek’s AI Models Attracted Attention

The investigation builds on earlier observations involving DeepSeek’s model development. According to DePue, the company’s V3 model, released in December 2024, displayed a numerical fingerprint compatible with Nvidia’s older Hopper generation.

However, its V4 models, released in April 2026, showed patterns associated with the Blackwell generation in the researcher’s analysis.

The difference has attracted attention because Blackwell represents a newer generation of Nvidia AI computing hardware, access to which is subject to US export restrictions for China.

In February 2026, Reuters reported that a senior US official believed DeepSeek had trained an AI model using Blackwell processors in mainland China. The new technical analysis offers another line of evidence concerning possible hardware access, but it does not independently establish where the processors were located.

US Export Restrictions on Advanced Nvidia Chips

Washington has imposed restrictions on exports of advanced artificial intelligence processors to China since 2022. These measures are intended to limit access to computing technology considered strategically important for advanced AI development.

In January 2026, the US Commerce Department introduced a case-by-case licensing framework for Nvidia’s H200 processors and comparable chips, subject to specified conditions. Blackwell processors, however, continue to require export licences for shipments to China.

These restrictions have made access to high-performance AI computing an important issue for Chinese technology companies. Developers need substantial computing capacity to train and operate advanced models, creating demand for access to processors through permitted channels and overseas infrastructure.

The central question raised by DePue’s research is therefore not simply whether the models contain Blackwell-compatible patterns, but how the developers obtained access to the computing resources that may have produced them.

Chinese AI Companies Look Towards Overseas Data Centres

Reuters reported on November 12 that Chinese AI companies were seeking Nvidia computing capacity at overseas data centres amid concerns about potential restrictions on remote access.

This development highlights the growing importance of international data-centre infrastructure in the AI industry. Companies can use computing resources located outside their home countries, but access to restricted technology and the rules governing remote use remain subject to regulatory requirements.

A bill addressing access to advanced computing resources through overseas facilities had passed the US House of Representatives but was still awaiting action in the Senate, according to the supplied report.

DePue’s findings do not determine whether the processors connected to the numerical patterns were located in China or overseas. They also do not establish whether any company violated US export-control requirements.

Why the Research Does Not Prove Export Violations

Although the 445 exact matches in DeepSeek-V4-Flash represent a significant technical finding, identifying a hardware fingerprint is different from proving an export-control violation.

The numerical patterns can provide clues about the hardware environment used to generate initial model weights. They cannot independently establish the physical location of the processors, the identity of their owners or the route through which computing access was obtained.

The research also has limitations beyond the question of location. The AMD alternative was not conclusively ruled out for all six models, and the reproduction scripts were available to reviewers on request rather than being publicly released.

These limitations mean the findings should be treated as technical evidence requiring further verification, not as a final determination of unlawful chip transfers.

Nvidia Leadership Changes Add to Industry Scrutiny

The research appeared during a period of significant developments at Nvidia. According to a US securities filing, the company’s board accepted Jensen Huang’s departure effective November 11 and appointed Chief Financial Officer Colette Kress as interim chief executive.

The leadership change followed scrutiny surrounding allegations involving Huang and former Megaspeed executive Alice Huang. Nvidia had previously stated that it had no reason to believe national-security information or company secrets had been compromised.

However, DePue’s research does not identify any connection between the numerical fingerprints found in Chinese AI models, Jensen Huang or Megaspeed. It also does not establish that the leadership developments were related to the chip-access questions raised by the analysis.

What the Findings Could Mean for the AI Industry

The research highlights how technical analysis of publicly available AI models may help investigators and researchers examine the computing infrastructure behind advanced systems.

As AI development becomes increasingly dependent on powerful processors, the ability to identify hardware-related patterns could become relevant to discussions about technology supply chains, export controls and international computing access.

For policymakers, the findings underline the difficulty of monitoring advanced computing resources when companies can rely on overseas data centres and complex technology arrangements. For AI developers, they highlight the growing importance of transparency around computing infrastructure and compliance with applicable regulations.

Nevertheless, the findings do not establish wrongdoing by the companies whose models were examined. Further independent testing and additional evidence would be needed to determine the precise hardware arrangements involved.

Conclusion

The discovery of 445 exact numerical matches in DeepSeek-V4-Flash has brought renewed attention to the possible use of Nvidia Blackwell technology in Chinese AI development. Researcher Will DePue also identified compatible patterns in models from Moonshot AI, MiniMax and Ant Group, although some alternative explanations remain unresolved.

The analysis provides technical clues about the computing hardware associated with these models, but it does not prove that restricted processors were used inside China or that any export rules were violated. As US restrictions and demand for advanced AI computing continue to shape the industry, questions about hardware access, overseas data centres and regulatory oversight are likely to remain important.