The impressive capabilities of Chinese research institute Moonshot’s Kim K3, the largest open-weight large-scale language model, have sparked a debate that confuses two things: the economic potential of the US AI giant and the future of LLM as a technology.
Dean W. Ball, head of strategic futures at OpenAI, went so far as to argue that because the open weight model would necessarily prevent Frontier Labs from making capital investments, the U.S. government should find an excuse to create regulatory fear, uncertainty, and mistrust about the new model.
People were outraged when tech luminaries like Yann LeCun and Martin Casado argued that open software could accelerate innovation and coexist with proprietary projects. Mr. Ball quickly walked back his claims that a regulatory crackdown was the White House’s “best strategy” and that an indiscriminate weight model would inevitably slow technological progress.
However, Axios reported that the Trump administration is considering banning the K3 and other advanced Chinese models at the request of the American Frontier Research Institute. Another Politico report said the Commerce Department would not take that action for now.
The benefits for large AI companies are clear. Openweight models, running on independent infrastructure or within large companies, provide intelligence at a lower cost than class-leading models from Anthropic and OpenAI. As users spend more outside of closed labs, it means a smaller return on their huge investment in model training.
Its views extend far beyond OpenAI. “A strong, frontier-level open source model will squeeze margins and drive down prices for frontier companies,” Braden Hancock, co-founder of Snorkel AI and research partner at Lord Institute, told TechCrunch. “It doesn’t necessarily mean we’re going to use a little less AI. Quite the opposite, obviously.”
For those who don’t own shares in Anthropic or OpenAI, that’s fine. AI will continue to proliferate. So what is the government’s justification for preventing Americans from purchasing something on the ostensibly free market?
There are several types of concerns about the Chinese model. One is to protect U.S. data from the Chinese government. The United States has banned imports of modern Chinese EVs, citing data collection concerns. But while experts tend to think it is unlikely that an open weight model running on US servers would leak data to China, it is not impossible that something like that could happen.
Another is that the model may have an implicit bias against China, but it’s not clear what that means for the coding task, for example.
A third common concern is that Chinese models lack guardrails mandated by the U.S. government (through an opaque process) aimed at preventing large U.S. LLMs from being used to exploit closed computer systems or create weapons. But those same guardrails could make U.S. companies more vulnerable. David Sachs, a venture capitalist and advisor to President Trump, has shared examples of US companies turning to Chinese LLMs to fill security gaps when the US frontier model refuses to do the job.
But the biggest motivation for restricting the model is the fear that China could overtake the United States if Frontier Labs slows down.
Sam Bresnick, a China expert at Georgetown’s Center for Security and Emerging Technologies, said the growing importance of AI in U.S. military operations provides reason for the U.S. to support continued investment in AI at frontier labs. But the whole issue is complex, he says.
“Why is the weight of the U.S. government focused on protecting these companies from competitors who are locked out of the U.S. market because of their origins?” asks Bresnick.
Proponents of open AI argue that frontier companies are creating a false dichotomy between innovation and closed models.
“The bigger impact of bringing these open source models from China is that rather than China having a backdoor in, it’s that China owns the innovation,” Hancock told TechCrunch. “What you end up with is, in effect, an expanded modeling workforce. PyTorch became an industry standard because it was open source, so the whole community could contribute, not just one company, and it just grew and grew. All the other deep learning libraries kind of died out by comparison.”
Mr Hancock and other advocates are concerned that LLMs in China will become hubs for international research. Already, U.S. graduate programs are largely built on a promiscuous Chinese model, with half of the papers students research coming from Chinese institutions, and U.S. frontier labs have become increasingly reluctant to share their research widely, Hancock said.
“Restricting open models does not make AI secure,” said Clem DeLang, CEO of Hugging Face, an open AI collaboration platform. “It simply hides risks, concentrates power in the hands of a few, and makes it harder for the next generation of builders, researchers, academia, nonprofits, and governments to participate in making AI safer and more beneficial for everyone.”
Bresnick says the real way to slow China down is to put more emphasis on chip export controls. A better way to maintain U.S. AI leadership is to stop selling Nvidia H200 processors to China. “That could take us away from this thorny debate about banning open source technology, which a huge number of American companies want to use,” he says.
Part of the problem is uncertainty about the economics of AI. “We don’t understand both open and proprietary business models. AI companies are having a hard time finding ways to make money with their tools, especially as training costs have to go up more and more,” Bresnick points out.
The same challenges occurring in the United States are occurring in China, where AI companies are struggling to make money and tap into computing power, and the government appears to be encouraging open releases for policy reasons, despite the difficulty in leveraging them.
Some US companies, such as Thinking Machines Lab and Nvidia, are trying to build their business around releasing open models. Hancock said NVIDIA “does better with dozens or hundreds of companies building AI than with two or three companies with enough capital to make their own chips,” which is one reason for investing in Nemotron, a collection of open models.
“The point is that it would be very beneficial for the United States to have its own very capable and much cheaper open model,” Bresnick said. “It just conflicts with the approach taken by Frontier Laboratories.”
With additional reporting from Rebecca Beran.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
