OpenAI is scared of open-weight models. Should the US be?

China's Moonshot company's Kimi K3, the largest open-source large language model, has sparked a US policy debate. OpenAI's head of strategic futures, Dean W. Ball, once publicly argued that the US government should create regulatory fear, uncertainty, and distrust because open-source models inevitably hinder capital expenditure by frontier labs. However, he later retracted this statement, acknowledging that a regulatory crackdown is not the White House's "best strategy" and that open-source models may not slow technological progress.

Axios reported that the Trump administration is considering banning K3 and other advanced Chinese models at the request of US frontier labs, but Politico says the Commerce Department will not take this step in the short term.

The impact of open-source models on major AI companies is clear: open-source models running on independent infrastructure or within enterprises provide cheaper intelligence than top models from Anthropic or OpenAI. If users increasingly turn away from closed labs, it means the massive investments these labs have made in model training will see diminished returns.

Snorkel AI co-founder Braden Hancock notes: "Powerful frontier-level open-source models will squeeze profit margins and lower prices for frontier companies. But this won't reduce AI usage—quite the opposite."

Concerns about Chinese models include data leakage risks, implicit pro-China biases, and a lack of safety guardrails mandated in the US. However, experts believe that open-source models running on US servers are unlikely to leak data to China.

Interestingly, safety restrictions on US companies may actually make them more vulnerable: venture capitalist and Trump advisor David Sacks shared a case where a US company turned to a Chinese LLM to fill security gaps because frontier models refused to perform certain tasks.

Open-source advocates argue that frontier companies are creating a false dichotomy between innovation and closed models. Hancock warns that Chinese LLMs could become the center of international research. US graduate programs are primarily built on Chinese open-source models; Hancock says half of student research papers come from Chinese institutions, while US frontier labs are increasingly reluctant to share their work widely.

Hugging Face CEO Clem Delangue says: "Restricting open-source models won't make AI safer—it will only hide risks, concentrate power in a few hands, and make it harder for the next generation of builders, researchers, nonprofits, and governments to participate."

Georgetown University CSET researcher Sam Bresnick believes that the real way to slow China is to strengthen chip export controls, such as stopping the sale of Nvidia H200 processors to China, which could avoid the thorny debate over banning open-source technology.

Bresnick notes that the AI economy itself is full of uncertainty: "Neither the open-source nor the proprietary business model has been finalized. AI companies are all struggling to find profitability, especially as training costs keep rising." The same challenges exist in China, where AI companies also struggle to generate revenue and access computing power.

Some US companies, such as Thinking Machines Lab and Nvidia, are trying to build businesses around open-source models. Hancock points out that Nvidia would benefit more if there were "dozens or hundreds of companies building AI, rather than just two or three with enough capital to make their own chips."