Export Controls Can't Catch Corporate Structure

Export controls written as a list of entities can't restrain a system that moves through corporate structure. Chips are objects on a shipping manifest. Subsidiaries, joint ventures, contract relationships, those are agreements, and agreements rearrange faster than a rulebook can name them.

Commerce posting guidance on a Sunday that Nvidia's Rubin and Blackwell and AMD's MI350x can no longer ship to offshore subsidiaries of Chinese AI firms is being framed as a closed loophole. It isn't. The guidance, by existing, concedes those chips have been moving through third-country subsidiaries (Malaysia gets named) for roughly a year. A year isn't a loophole. A year is the policy working as written and missing what it was supposed to catch.

The mismatch is the thing worth naming. The controls treat AI compute as a regulated good, like a centrifuge or a fighter jet. Those are objects. You stamp them, license them, track them across a border. AI chips behave that way at the loading dock and stop behaving that way the moment they sit inside a structure of subsidiaries the entity list doesn't reach. The chip moves once. The corporate relationship that put it where it ended up moves continuously.

You can keep adding entities. The new guidance does exactly that, extending the reach to known offshore arms. The next structure is already being drawn up in a law firm somewhere, and the policy is always going to be a step behind, because a new subsidiary takes a week and a new rule takes a year.

The point isn't that controls failed or that controls worked. It's that an entity-based regime is the wrong shape for the problem, and the people writing it know it. The fix taking roughly a year and arriving on a Sunday tells you which way the pressure was running.

A paper out this week argues that a Chinese predictive-surveillance vendor's political-risk modeling was constrained by export controls. Maybe. It doesn't pin down which hardware trained which model in which window, and that's exactly the link you'd need to make the claim do real work. "Suggestive" is the right word. "Constrained" is doing more lifting than the evidence supports, especially now that we know the chips were getting through on the side most people assumed was sealed.

I don't know how much of China's frontier AI buildout depended on those Malaysian-routed shipments versus indigenous capacity, and that's the gap that matters most. The Sunday guidance suggests Commerce thinks the answer is "enough to bother closing." That's an admission, not housekeeping.