The impressive power of the Chinese lab Moonshot’s Kimi K3, a large open-weight language model, started a debate involving two things: the economic possibilities of American AI giants and the future of LLMs as a technology.
OpenAI’s head of future strategies, Dean W. Ball, even argued that the US government should find an excuse to create legal fear, uncertainty, and distrust in new models, as open-weight models should prevent the spending of large amounts of money by border labs.
People are shocked, with technology luminaries like Yann LeCun and Martin Casado arguing that open software can accelerate innovation and complement proprietary projects. Ball quickly retracted his claims that the breach was the White House’s “best strategy” and that open-weight models are slowing down advances in technology.
However, Axios reports that the Trump administration is considering banning the K3 and other Chinese-developed models at the behest of US border labs. Another report from Politico said that the Ministry of Commerce will not be quick to take that step.
The advantage of large AI companies is clear: Open-scale models, which work with independent infrastructure or within large enterprises, offer cheaper intelligence than the best-in-class Anthropic or OpenAI models. If users increasingly spend more money outside of closed labs, that means less return on their large investment in model training.
That vision extends beyond OpenAI. “Strong, high-quality open source models will limit margins and reduce costs for edge companies,” Braden Hancock, founder of Snorkel AI and a research fellow at the Laude Institute, told TechCrunch. “It’s not going to mean that the amount of AI usage is going down. You know, obviously, quite the opposite.”
That’s not a problem for people who don’t have shares in Anthropic and OpenAI. AI is still expanding. So what is the reason for the government to prevent Americans from buying something in our seemingly free markets?
Concerns about Chinese models come in many forms. One protects US data from the Chinese government; the US has banned the import of modern Chinese EVs due to concerns about data collection. But experts tend to think that open-source models running on US servers are less likely to leak data back to China, although it is unlikely that such a thing could be done.
Another is that the models may have an obvious bias towards PRC – but it’s not clear what that might mean for, say, coding functions.
A third common concern is that the Chinese models do not have the US government-mandated surveillance systems (through a vague process), which aim to prevent US-led LLMs from being used to exploit closed computer systems or create weapons. However, those precautions can make US companies more vulnerable: David Sacks, a financier and adviser to Trump, has been sharing cases of American companies turning to Chinese LLMs to cover security gaps where the models at the American border refuse to carry out operations.
But the most important motivation to limit the models is the fear that China will be able to overtake the US if the border labs come down.
Sam Bresnick, a China-focused researcher at Georgetown’s Center for Security and Emerging Technologies, says the growing importance of AI in US military operations gives the US reason to support continued investment in AI in border labs. But the whole question, he says, is full.
“Why should the weight of the US government be aimed at protecting these companies from competitors who are locked out of the American market based on their origins?” Bresnick asked.
Proponents of open AI say frontier companies create a false dichotomy between innovation and closed models.
“The biggest impact of these kinds of open sources coming from China is less that they’re coming in through back doors, and more that they’re proprietary,” Hancock told TechCrunch. “You end up with an extended workforce in your model. PyTorch became the industry standard because it’s open source, so the whole community was able to contribute to it instead of one company, and it grew and grew, and all the other deep learning libraries died in comparison.”
Hancock and other promoters fear that China’s LLMs will become a hotbed of international research. Already, US graduate programs are building largely on China’s open-minded models, and Hancock says that half of the papers students read come from Chinese institutions, while America’s frontier labs are increasingly excited about sharing their work more widely.
“Limiting open models will not make AI safe,” said Clem Delangue, CEO of Hugging Face, an open AI collaboration platform. “It would only hide the risks, concentrate power in the hands of the few and make it difficult for the next generation of developers, researchers, academics, non-profits, governments to play a role in making AI safe and beneficial for everyone.”
Bresnick says the real way to slow down China would be to focus more on controlling chip shipments. A better way to maintain US AI leadership would be to stop selling Nvidia H200 processors to China. He says, “That would get us out of this thorny debate about blocking open source technologies that a lot of American companies want to use.”
Part of the problem is uncertainty about the economics of AI. “An open business model, a proprietary business model – none of them are considered. AI companies are having trouble finding a way to monetize their tools, especially since training costs need to rise,” Bresnick points out.
The same challenges playing out in the US are playing out in China, where AI companies are also struggling to generate revenue and access computing power, and the government appears to be encouraging open source for policy reasons despite the challenge of implementing it.
Some American companies, including Think Machines Lab and Nvidia, are trying to do business by releasing open models. Hancock points out that Nvidia will do better “if there are dozens or hundreds of companies building AI rather than two or three and two or three have enough money to make their own chips,” which is one of the reasons for investing in Nemotron, a collection of open models.
“The bottom line is that the US would be best served to have its own more powerful, less expensive open models,” Bresnick said. “It contradicts the way the border labs have taken it.”
With additional reporting from Rebecca Bellan.
If you shop through links in our articles, we may earn a small commission. This does not affect our editorial independence.
