The remarkable features of the Chinese lab Moonshot’s Kimi K3, the largest open-weight large language model, have sparked a discussion that blends two aspects: the economic potential of American AI giants and the evolution of LLMs as a technology.
Dean W. Ball, OpenAI’s head of strategic futures, even suggested that the US government ought to fabricate a rationale to instill regulatory fear, uncertainty, and skepticism regarding the new models, as open-weight models should inherently discourage capital investment from the leading labs.
The tech community reacted strongly, with notable figures like Yann LeCun and Martin Casado contending that open software can drive innovation and exist alongside proprietary initiatives. Ball swiftly walked back his assertions that a regulatory clampdown was the White House’s “best strategy” and that open-weight models inevitably hinder technological progress.
Nonetheless, Axios reports that the Trump administration is contemplating a ban on K3 and other sophisticated Chinese models at the request of American frontier labs. Another report from Politico indicated that the Department of Commerce would not take that action in the near future.
The advantage for prominent AI firms is evident: Open-weight models, operating on independent infrastructure or within major enterprises, provide more affordable intelligence compared to Anthropic or OpenAI’s top-tier models. Should users increasingly allocate more resources outside the closed labs, it would result in diminished returns on their substantial investments in model training.
This perspective extends beyond OpenAI. “Robust, frontier-caliber open-source models will pressure the margins and lower the prices of the frontier companies,” remarked Braden Hancock, co-founder of Snorkel AI and a research partner at the Laude Institute, to TechCrunch. “It does not necessarily imply that the volume of AI usage will decrease at all. On the contrary, it’s likely the opposite.”
This isn’t an issue for those who do not own shares in Anthropic and OpenAI. AI will continue to expand. So what grounds does the government have to prevent Americans from buying something in our seemingly free markets?
Concerns surrounding Chinese models manifest in various ways. One pertains to safeguarding US data from the Chinese government; the US prohibited the importation of modern Chinese EVs due to fears about their data collection practices. However, experts generally believe that open-weight models hosted on US servers are unlikely to transfer data back to China, though such possibilities are not entirely out of the question.
Another concern is that the models might exhibit implicit bias towards the PRC — yet it remains uncertain what implications this could have for, for instance, coding assignments.
A third prevalent apprehension is that Chinese models may lack the safeguards mandated by the US government (through an unclear process), which aim to prevent leading US LLMs from being exploited for unauthorized access to closed computer systems or weapon creation. However, those same safeguards may expose US companies to greater risks: David Sacks, a venture capitalist and Trump advisor, has been sharing instances of US companies opting for Chinese LLMs to bridge security shortfalls when US frontier models decline to fulfill the tasks.
Yet, the principal impetus for limiting these models is the anxiety that China may surpass the US if the frontier labs decelerate.
Sam Bresnick, a research fellow at Georgetown’s Center for Security and Emerging Technology focusing on China, asserts that the increasing significance of AI for US military operations provides the US with a rationale to advocate for sustained investment in AI at frontier labs. However, he notes that the entire issue is complicated.
“Why should the U.S. government’s influence be directed at safeguarding these companies from rivals that are being excluded from the U.S. market based on their origins?” Bresnick questions.
Proponents of open AI argue that frontier companies are constructing a misleading binary between innovation and proprietary models.
“The more significant impact of having these open-source models emerge from China is less about potential backdoor intrusions, and more about them leading the innovation,” Hancock told TechCrunch. “You effectively end up with an expanded workforce on your model. PyTorch emerged as the industry standard due to its open-source nature, allowing the entire community to contribute to its development, as opposed to just one company, leading to its substantial growth while other deep learning libraries dwindled in comparison.”
Hancock and other supporters worry that Chinese LLMs may become the center of international research. Currently, US graduate programs largely rely on open-weight Chinese models, with Hancock noting that half of the papers students study originate from Chinese institutions, as American frontier labs increasingly hesitate to share their work broadly.
“Restricting open models wouldn’t enhance AI safety,” stated Clem Delangue, CEO of Hugging Face, an open AI collaboration platform. “It would merely conceal the risks, centralize power among a select few, and hinder the ability of the next generation of builders, researchers, academia, non-profits, and governments to contribute to making AI safer and more advantageous for everyone.”
Bresnick argues that effectively slowing down China would require focusing more on chip export controls. A better strategy for maintaining US AI superiority would involve ceasing sales of Nvidia H200 processors to China. “That,” he says, “could potentially keep us out of this complex debate surrounding the banning of open-source technologies that numerous US companies wish to utilize.”
Part of the dilemma lies in the ambiguity surrounding AI economics. “The open business model, the proprietary business model — neither is thoroughly established. AI firms are currently grappling with how to monetize their tools, particularly as training expenses continue to rise,” Bresnick remarks.
The same issues that unfold in the US are also manifesting in China, where AI companies are struggling to generate profits and access computational resources, with the government perceived as advocating for open releases for policy reasons despite the challenges in capitalizing on them.
Some US firms, such as Thinking Machines Lab and Nvidia, are attempting to create a business model around releasing open models. Hancock believes Nvidia would fare better “if there were dozens or hundreds of companies developing AI rather than just two or three that are sufficiently capitalized to produce their own chips,” which is part of the rationale behind its investment in Nemotron, a suite of open models.
“The essential point is the U.S. would greatly benefit from having its own highly capable, much more affordable open models,” Bresnick stated. “It simply conflicts with the approach the frontier labs have adopted.”
With additional reporting from Rebecca Bellan.
