Why AI export controls are quietly redrawing the map of who gets to build frontier models

Here’s something that shouldn’t be possible, and yet: in the first half of this year, US export policy on the same company’s AI chips got tighter and looser at the same time, depending on which chip you were asking about.

In January, a Federal Register rule capped China-bound shipments of Nvidia’s H200 at 50% of US volumes, added a 25% tariff routed through Taiwan, and required US-supply certifications. Around the same window, the administration eased broader restrictions on H200 sales — Nvidia’s CEO announced at GTC that the company had resumed taking Chinese orders for it. Then in late May, Commerce moved the other way on Nvidia’s newest Blackwell-series chips: new rules require an export license for any transfer to a China- or Macau-headquartered entity, and close the loophole that let Chinese firms buy through overseas subsidiaries. Net effect by the end of the fiscal year: Nvidia’s most advanced chips were effectively locked out of China’s data center market, while a tariffed, capped version of the previous generation kept moving.

That’s not a policy. That’s a moving target with a tier system attached to it.

Compute access is becoming the real border

I don’t think most people outside the industry have registered what this actually does. Export controls used to be a narrow tool for a narrow category of hardware — munitions-adjacent, dual-use, the kind of thing that shows up in a handful of specialized deals. What’s happening now is different in kind: the compute needed to train a frontier model has become scarce and geopolitically gated enough that access to it is starting to function like a border. Which country’s labs get to work with which generation of chip is now a live policy lever, adjusted quarter to quarter, not a settled fact of the global tech landscape.

The practical effect isn’t that any single country gets locked out entirely — it’s that a tiered map forms. Top-tier chips go to allies and trusted commercial partners. A capped, tariffed, one-generation-behind tier goes to everyone else who can still clear the paperwork. And whoever’s a generation behind on compute access is, functionally, a generation behind on what their frontier labs can even attempt — regardless of the talent in the room.

The instability is the story, not just the restriction

What strikes me most isn’t which chips are restricted this month. It’s that the rules moved twice in six months, in opposite directions, for the same underlying reason: leverage. Chip access is being used as a bargaining chip in a broader negotiation, which means the actual policy on any given day is downstream of things that have nothing to do with AI safety or technical merit — tariff negotiations, diplomatic posture, election-cycle politics.

If you’re a lab, a government, or a company anywhere outside the small set of countries with guaranteed top-tier access, that instability is arguably worse than a clear, stable restriction would be. You can plan around a wall. It’s much harder to plan around a border that moves.

What I’d actually watch

Not the headline chip-ban stories — those get attention but tell you less than they seem to. Watch the tier structure: which countries sit in the “guaranteed access” bucket versus the “capped and tariffed” bucket versus “excluded,” and how often that bucket assignment changes. That’s the real map of who gets a fair shot at building frontier AI over the next five years, and right now, it’s being redrawn in real time.

This is the first entry in The Long View*‘s Social & Geo-Politics section — ongoing essays on the power plays and access questions shaping who actually gets to build what, in AI and beyond.*