Open Models Scare OpenAI. Should the US Be Afraid of Them?

The arrival of Kimi K3 — the largest open-weight language model to date, from the Chinese lab Moonshot — has kicked off a discussion that, as TechCrunch notes, blends two very different questions: the commercial prospects of America's AI giants and the future of large language models as a technology.

OpenAI's head of strategic future, Dean W. Ball, went so far as to suggest that the US government find a pretext and artificially cultivate an atmosphere of regulatory uncertainty and distrust around the new models — since open weights inevitably undercut frontier labs' incentives to invest in training. The reaction was fierce: Yann LeCun and Martin Casado pointed out that open software accelerates innovation and coexists just fine with proprietary software. Ball soon walked back his words.

Still, according to Axios, the Trump administration is nonetheless weighing a ban on K3 and other advanced Chinese models at the request of American labs; Politico, meanwhile, reports that the Commerce Department is in no hurry to take such a step.

Where the Real Interest Lies

The upside for the big AI companies is obvious: open models running on independent infrastructure or inside corporations deliver cheaper intelligence than the flagships of OpenAI and Anthropic. The more users spend money outside the closed labs, the lower the return on those labs' enormous investments in training. “Strong open frontier-class models will compress margins and drive down the prices of the closed companies,” says Snorkel AI co-founder Braden Hancock, adding that overall AI usage will only grow as a result.

Various arguments are marshaled for a ban: leakage of data to China (which experts consider unlikely for models running on American servers), a possible hidden pro-PRC bias, and the absence of mandatory safeguards. Paradoxically, it is precisely these safeguards, according to investor and Trump adviser David Sacks, that push American companies toward Chinese models — when their own refuse to carry out a task, such as patching vulnerabilities.

The Price of Isolation

The chief fear, though, is that China will surge ahead if the frontier labs ease off the pace. But, as Sam Bresnick of Georgetown's CSET observes, is it worth marshaling the full might of the state to shield a handful of companies from competitors cut off from the market solely by their origin?

Proponents of open AI warn: the real threat is not “backdoors” but the fact that Chinese models are becoming the center of the world's research. Already, half the papers American graduate students study come out of Chinese institutions. By analogy with PyTorch, which became the standard precisely because of its openness, leadership goes to whoever's platform the community gathers around.

Hugging Face's head, Clément Delangue, insists that restrictions won't make AI safer — they'll merely hide the risks and concentrate power in the hands of a few. Bresnick sees a more honest way to hold China back: export controls on chips, in particular a halt to sales of Nvidia's H200. And above all, the US would do well to have its own powerful and cheap open models — which, alas, runs counter to the frontier labs' course.

Source: TechCrunch