Frontier AI looks less like software and more like uranium with an API.

Frontier AI looks less like software and more like uranium with an API.

China is reportedly considering restrictions on overseas access to its most advanced AI models, including unreleased systems from firms such as Alibaba, ByteDance and Z.ai.

The proposed direction is familiar: keep basic tools open enough to spread influence, but treat frontier capability as a national asset.

The US has already shown the same instinct from the other side. When Washington temporarily restricted foreign access to Anthropic’s Fable and Mythos models, it made the precedent visible: frontier AI access can be switched off by state action, not just vendor choice.

For a while, Chinese AI firms used open and low-cost models as a distribution strategy. Good enough, cheap enough, widely available enough to win global users while the American labs fought over subscription tiers and compute bills.

Now Beijing appears to be drawing a harder line. Advanced models, training knowledge, data, talent and acquisitions are being pulled into the same national-security frame. The model is no longer just a commercial product. It is strategic infrastructure.

So the global AI market is not simply converging around the best model at the best price. It is fragmenting into zones of permission.

Open enough to attract users.

Closed enough to protect advantage.

That matters for businesses because many AI strategies still assume the model layer will remain broadly available, globally accessible and commercially neutral. That assumption is getting weaker by the month.

If a system is important to your business, ask a dull but necessary question: what happens if access changes?

Not because China might block something. Not because the US might. Because both already understand the same thing.

The most capable models are no longer treated as ordinary software.

They are leverage.

And if your workflow depends on someone else’s leverage staying available, you do not have an AI strategy yet.

You have an exposure.

That exposure cannot be fixed by shouting for European sovereignty after the fact. Europe does not yet have a frontier AI stack at US or Chinese scale, and slogans do not create one.

So the practical question is not “sovereignty, yes or no?”

It is much duller, and much more useful:

Which parts of your AI stack could you move if the rules changed tomorrow?