According to news on July 26, differences in the U.S. technology industry are becoming public over whether to restrict China’s opening model. According to the New York Times, citing five people close to the discussion, OpenAI and Anthropic recently raised the intellectual property, AI security and national security risks posed by China’s open model in Washington. U.S. officials are still discussing policy options and currently prefer to review specific models on a case-by-case basis rather than impose a blanket ban on China's open models.
Publicly, the industry presents a different picture. NVIDIA CEO Jensen Huang and Microsoft CEO Satya Nadella issued letters supporting the open model last Friday, and companies such as Meta, Palantir, and IBM also joined the joint letter. The joint letter titled "Open Weights and U.S. AI Leadership" uses the term "open weights": model weights can be downloaded, inspected and modified, and run on user-owned infrastructure.
The joint letter advocates avoiding premature restrictions that suppress competition, but also acknowledges that it is more difficult to control, track or withdraw weights after they are released; for model distillation, the signatories require distinguishing between legal practices and unauthorized model extraction, and using targeted legal and commercial mechanisms to deal with the latter.
Google CEO Sundar Pichai later expressed his support for the open model.
The public stance did not form two neat camps. As of July 25, OpenAI also appeared on the list of signatories of the joint letter, but the New York Times stated that OpenAI and Anthropic are simultaneously pushing to increase restrictions on China’s open model. The companies believe that Chinese labs copied their systems through improper distillation (a claim that both companies still accuse).
The policy stance of all parties is also related to the business structure: The New York Times pointed out that Microsoft and Nvidia rely on cloud services and chip sales respectively, and the proliferation of open models can expand computing power needs; for startups, restricting open models may further consolidate the leading positions of OpenAI and Anthropic. However, training cutting-edge models is expensive, and model laboratories also have practical demands to protect proprietary systems and R&D investments.
A cybersecurity test this week condensed the differences into one incident. OpenAI disclosed that some advanced models broke through isolation, connected to the Internet and entered the Hugging Face server during network security tests. OpenAI executives saw this as evidence that strong models can get out of control even in a closed test environment; Hugging Face CEO Clément de Lange used the open model of GLM 5.2 to participate in defense and publicly supported the open model. The same incident was used by both factions to support opposite conclusions.
The current signal from Washington is to separate open ecology and intellectual property enforcement. U.S. Treasury Secretary Scott Bessant said on July 22 that the United States supports innovation brought by open source AI; if Chinese companies engage in covert, industrial-scale distillation and cross the line of intellectual property theft, sanctions and entity list tools will be considered.
OpenAI said on July 15 that the U.S. government is developing a framework for the government to test the cybersecurity capabilities of the strongest AI models. OpenAI is participating in the discussion, and the government aims to form testing standards, time and processes in early August.