China’s next-generation large model companies may not open-source their models.

SemiAnalysis’ Dylan Patel recently mentioned that several Chinese model labs have told inference service providers that the next model will switch to a licensed model rather than being open-sourced. His view is that open is dying quickly.

If only one or two companies move toward closed source, that’s not unusual. What’s truly worth watching is that the interests across the entire industry chain are slowly aligning on the same side.

Model labs naturally want to sell APIs and collect licensing fees; inference service providers prefer models that others can’t easily deploy; downstream application companies also want the capabilities they pay for to be sufficiently scarce.

Once a model becomes fully open source, prices get undercut, migration costs fall, and many so-called “barriers” disappear.

So now, discussions about open-source AI are getting more and more interesting. Security, misuse, and leaks of model weights are all real risks.

But at the same time, the idea that “open source is dangerous” is increasingly aligning with the commercial interests of everyone across the industry chain.

Over the past two years, China’s model community has directly open-sourced top-tier model capabilities, driving API prices down step by step—and forcing U.S. model companies to lower their prices as well.

If the next generation of Chinese models really begins to shut its doors collectively, the open-source boom that everyone takes for granted today may have originally been only a very brief window.