CNAS Insights | A Closing Door on Open Weights? Implications for AI Sovereignty | CNAS

CNAS has published an insight titled "A Closing Door on Open Weights?" and the answer, per the think-tank-industrial complex, is: yes, and sooner than the conference circuit would like. Foreign countries and companies can likely no longer count on the regular release of highly capable open-weight models, which is a polite way of saying the free samples are being quietly withdrawn from the counter. Note the phrase "AI sovereignty" doing the heavy lifting — a proud national capability consisting, in practice, of downloading a file that somebody in California or Beijing was gracious enough to hand over.

The number everyone will quote comes from Epoch AI: China's leading open-weight model, GLM-5.3, trails the U.S. frontier in cyber capabilities by just six to seven months. Six months. That is not a moat, that is a delivery window. And of all the yardsticks on the shelf, they picked cyber — the capability nobody wants to be second in and everybody insists they are only researching responsibly. The word "just" in "just six to seven months" is carrying more editorial freight than the rest of the paragraph combined.

Then what? Everyone who built a national strategy on the assumption that somebody else's weights would keep arriving on schedule now has to ring the frontier labs and ask nicely, which is how "sovereignty" ends up meaning a licence agreement with a company in another jurisdiction. The door, incidentally, is being pushed shut by the same labs whose safety documents insist that openness is the risk — an argument with the considerable virtue of also being the invoice. Cheer up, though: six to seven months is a lifetime in this industry, provided anyone will define the term before the next release.