EL7.AI
GeopoliticsMedium•13 November 2026•
4 min read

Blackwell fingerprints emerge in weights of 6 Chinese models from 4 labs

Key Facts

1The report identified fingerprints consistent with Blackwell in the weights of 6 Chinese models, but said they do not prove full training on those chips.
2The strongest evidence for each model ranged from 49 to 1,385 sigma, and the weakest individual test came in at 20 sigma.
3The report linked two models to a configuration with 148 processing units and the other four to one with 152 units.
4Axios said Trump authorized action for the third time during 2026 against companies suspected of diverting advanced chips to China.

A report published by Unemployment Labs on November 12, 2026, identified configuration fingerprints consistent with Blackwell chips in the weights of 6 Chinese AI models from four labs. The report said the fingerprints support the possibility that the chips were used for training, but do not prove that Blackwell hardware carried out the models’ full training. The findings put the use of advanced hardware at the center of debate over U.S. export controls, without establishing where the chips were located or how they reached their users. The report’s technical significance therefore remains separate from proof of a legal violation.

The report linked DeepSeek-V4-Flash and MiniMax-M3 to a configuration with 148 processing units, while linking the other four models to GB200 or GB300 configurations, each with 152 units. It said the first two models match a 148-unit B200 or B300 configuration, while the others are consistent with a 152-unit GB200 or GB300 configuration. The comparison was based on configuration characteristics visible in the model weights, according to the report. This match forms the basis for the reference to Blackwell, but does not identify the chips’ location or owner. It also does not by itself show whether the matching hardware was used in every stage of training.

The report said the strongest evidence for each of the 6 models ranged from 49 to 1,385 sigma, while the weakest individual test reported came in at 20 sigma. It described the findings as strong indications that the configurations matched the Blackwell layouts it examined. However, the strength of the statistical match does not answer questions about the chips’ location, ownership or route of access. The report explicitly said the weights do not reveal who owned the chips or where they were located. The fingerprints also do not prove that the chips themselves carried out the full training, or that export controls were breached.

The findings come amid repeated U.S. efforts to counter the diversion of advanced chips to China. Axios reported on November 11 that President Donald Trump had authorized action for the third time during 2026 against companies suspected of diverting advanced NVIDIA chips to China. But the report said the Commerce Department had not implemented the measures at the time. Axios added that Trump had approved action on previous occasions in May, August and November 5, and that Commerce Secretary Howard Lutnick and another official had resisted pressure to proceed. A White House official said Trump still trusted Lutnick, while the secretary faces scrutiny from members of Congress over his handling of concerns about chip diversion.

U.S. Bureau of Industry and Security policy sets conditions for reviewing semiconductor export licenses, including proof that exports will not reduce production capacity available to U.S. customers. The policy also requires the Chinese buyer to adopt compliance and customer-screening procedures, and the chips to undergo independent third-party testing inside the United States. Jeffrey Kessler, under secretary of commerce for industry and security, said export controls should evolve with technology while protecting national security. He said allowing H200 sales to China under monitoring requirements would support the U.S. technology ecosystem. These conditions outline the regulatory framework for the report’s findings, without determining whether any of the chips cited were covered by such licenses or violated them.

This comes alongside a shift in the choices of Chinese AI companies as they seek to reduce reliance on U.S. suppliers. The Associated Press reported that the open Pro and Flash versions of DeepSeek V4 partially support Huawei chips. The agency said this support reduces DeepSeek’s reliance on U.S. companies such as NVIDIA. DeepSeek drew market attention with the launch of its R1 reasoning model in January 2025, according to the agency. Separately, in a case involving the risk of server diversion, the Associated Press said a U.S. indictment accused officials at Super Micro and others of a scheme to divert high-performance servers assembled in the United States to China during 2024 and 2025. The agency reported that authorities cited orders for servers worth $2.5 billion and the diversion of company servers worth at least $510 million after they were assembled in the United States.

Open on chart