In a single month, Washington switched one company's frontier model off for the entire planet, then asked a second company to hold its next one back and clear users one at a time. Anthropic lost Fable 5 and Mythos 5 to an export-control letter in June. Two weeks later OpenAI agreed to ship GPT-5.6 as a vetted preview, with the government signing off on access customer by customer. Two very different episodes, one underlying instrument: a discretionary power the executive holds over who may run the most capable models, used without a statute, without a published finding, without any way to argue back.

That instrument isn't going anywhere. The legal hook already exists, ECCN 4E091 put closed-weight models trained above 10^26 operations on the control list at the start of 2025, and the broader draft rules reported in March would make the US a gatekeeper for nearly every high-end processor sold by an American firm. So the useful question isn't whether the government keeps its hand on the dial. It's what the dial does to the industry between now and the middle of 2028, now that everyone building or buying frontier AI has to assume it's there.

It gets written down, but only just. The unwritten version is too useful to abandon and too awkward to leave naked. A power exercised through letters fired off on a Friday afternoon invites lawsuits and bad press every time it's used, so the pressure to dress it up is real. Expect a regime that looks like licensing without admitting to it: mandatory pre-release sharing windows, the thirty-day notice the December executive order already floated, attestations, geofencing the labs can't actually enforce. Nobody will call it a license, because a license implies criteria a court could test. It'll behave like one while keeping the discretion that made the original letters work. The legal authority for all this is plausible while the facts stay murky, and that gap, real authority paired with evidence nobody outside the room gets to test, is exactly the space a formalized version will try to keep open.

The labs self-censor well before anyone asks. None of it leaves a paper trail, which is what makes this the hardest shift to prove and the one I'd bet on hardest. Once a company knows a model judged "too capable" can be pulled over a weekend on secret evidence, the rational move is to ship under that ceiling. Trim the demos, hold the riskiest capability back, route the spicier stuff through enterprise contracts where it's quieter. Anthropic spent the spring asking for a coordinated brake on frontier development and then learned what a brake feels like when someone else works the pedal. Every competitor watched that happen. The chilling effect won't show up in a press release. It shows up as a frontier that quietly advances slower in public than it does in the lab.

Allies stop waiting, slowly. The shutdown hit exactly the governments whose hospitals and researchers were mid-sentence when Fable went dark, and the reaction in Brussels and London was immediate talk of technological sovereignty. For those capitals the foreign-access ban was the moment America dropped its hands-off pose, and that kind of grievance funds budgets. By 2028 I'd expect real mid-tier domestic capability in the EU and a serious UK design play, plus a great deal of sovereign-AI theater that amounts to renting the same American chips behind a national flag. The mismatch I keep running into is that a directive takes an afternoon and a frontier lab takes a decade. Money chases the rhetoric faster than it closes that gap.

Open weights become the substrate nobody can switch off. A power aimed at denying capability to foreign nationals pushes those exact users toward models the agency can't reach. The cheap, capable, open-weight releases coming out of China have no off switch a US official can flip, and neither does a model running locally on someone's own hardware. I expect demand for the uncontrollable option to rise precisely because the controllable one became political. That isn't a win for anyone's security. It's capability sliding toward the places US leverage can't reach, which leaves Washington with less sight of frontier risk, not more.

Then the market splits into an approved frontier and everything beneath it. Picture two tiers by 2028: a small set of vetted, gated, government-cleared top models available to the US and a short list of close allies, and a much larger field of slightly-less-capable systems that ship the normal way. The gap between them stops tracking who has the best researchers and starts tracking who cleared the process. Build a sector on that and the most capable tier competes on clearance instead of merit, which is a peculiar incentive to hand an industry that sells itself on speed.

I could be wrong about the codification. An unwritten power that works in an afternoon is worth far more to the people holding it than a statute they'd have to stand up and defend in court, so maybe it never gets written down at all, and the whole thing just runs on letters and quiet phone calls indefinitely. I genuinely can't tell which way that one breaks. The direction underneath both versions is the part I'd put money on: the default is moving from "ship it and see" to "ask first," and the asking goes to an office that owes nobody an explanation, on evidence nobody outside the room gets to see. Two more years of that and asking won't feel like an interruption to how frontier models reach people. It'll just be the shape of the thing.

Sources: