US companies running Qwen, Kimi or DeepSeek could face disclosure duties in SEC filings and product documentation if Congress adopts the commission's proposals.
The U.S.-China Economic and Security Review Commission (USCC) wants Congress to stop new deployments of Chinese open-weight AI models in critical infrastructure, national security work and systems that handle bulk personal data unless their safeguards pass independent review. The proposal is one of three AI recommendations in the 2026 Annual Report to Congress, delivered on November 18, and all three are among the 10 of its 30 recommendations that the commission singles out as most important. The other two would require public disclosure of Chinese model use and tie AI chip exports to China to audited proof that US manufacturing capacity is not diverted.
The recommendations rest on Chapter 8, "Silicon to Sovereignty: China’s Bid to Control the AI Stack from Wafer to Weights, Part One." Its central finding is that China still trails at the bottom of the stack, in chipmaking, memory and Nvidia’s CUDA software platform, and likely will for some time, yet is gaining ground in models through cheap open-weight releases. The commission, which Congress created in 2000, approved the report unanimously with 10 members; two seats are vacant. Its recommendations bind no one until lawmakers act on them.
The timing puts the report against a calmer diplomatic backdrop. Nikkei Asia led with the commission’s warning that frequent Trump-Xi meetings are masking Beijing’s efforts to compete with the United States economically, technologically and militarily. The executive summary notes that the two leaders agreed at their September summit in Washington to start an AI dialog and an incident channel, and says leader-level stability must not delay competitive policy.
The commission finds China behind on chips but ahead in open-model adoption
On compute, the chapter cites EpochAI estimates that the United States hosted about three-quarters of observed global AI compute capacity in May 2025, against roughly 15% in China. It cites the Financial Times for Huawei’s Ascend chips, made by SMIC, reaching yields of nearly 40% by February 2025, up from about 20%, against an estimated 60% to 80% for Nvidia’s Blackwell. Public estimates put Huawei’s 2025 AI chip output at 200,000 to 400,000 units, compared with four million to five million for Nvidia. A source the commission did not name assumed Huawei would make about one million Ascend chips in 2026 and hold two-thirds of China’s AI chip market.
The chapter credits some progress. It says SMIC’s N+3 process reached higher transistor density than TSMC’s N6 for Huawei’s Kirin 9030 phone chip, and that CXMT began small-volume production of HBM3E, the high-bandwidth memory stacked beside AI processors, according to August reports by Reuters and The Information.
Models are where the balance shifts. Chinese open-weight models accounted for more than 60% of all model downloads on Hugging Face as of June 2026, the commission said, citing the platform’s data. Developers released more than 150,000 derivatives of Alibaba’s Qwen models between January and July. The report lists Airbnb, Cursor developer Anysphere, DoorDash, Lindy AI and Coinbase among US companies using or testing Qwen, Kimi, DeepSeek or GLM models, and notes that Amazon Bedrock and Azure AI Foundry both offer Chinese models.
Three proposals target model deployment, disclosure and chip capacity
The first recommendation would have the Commerce Department’s Office of Information and Communications Technology and Services, working with the Center for Advancing Innovation and Standards for Super Intelligence, define sensitive uses within 180 days. New deployments of covered models there would need a time-limited authorization based on independently evaluated safeguards. Agencies and operators would also inventory existing deployments and keep plans to replace the models.
The second would make it US policy that Americans know when Chinese-developed models sit behind public services or products, including fine-tuned or derived versions wherever they are hosted. Businesses that failed to disclose would face Federal Trade Commission action, and listed companies would report the extent of their use as a material risk in SEC filings. Reuters, in a report carried by AsiaOne, said the Trump administration is separately considering a disclosure requirement for US companies.
The third would amend the Export Control Reform Act so that exporters of advanced AI chips to China certify that the sales do not divert US capacity in advanced wafer fabrication, high-bandwidth memory or advanced packaging. Exporters would submit auditable data to the Bureau of Industry and Security, with criminal penalties for false certification.
Distillation and remote access are the gaps the commission flags next
The chapter treats distillation, training a model on another model’s answers, as a way around compute limits. It cites Anthropic’s attribution of more than 16 million exchanges across about 24,000 fraudulent accounts to DeepSeek, Moonshot AI and MiniMax, and a September advisory from the NSA, CISA and FBI naming those three plus Alibaba, StepFun and Z.ai.
It also says existing rules have not fully stopped Chinese firms from renting advanced chips in third-country data centers. The Commerce Department moved on that two days before the report, when it required licenses for overseas data centers serving Chinese companies. Reuters also reported the commission’s finding that China’s military is seeking advanced Nvidia chips for AI-enabled weapons, quoting the report as saying "China is seeking to evade US export controls."
The commission says Part Two will examine software, servers, cloud providers and energy. The nearer test is whether the incoming 120th Congress writes the disclosure or certification proposals into bills.