On the Lutnick letter, the offshoring fantasy, and why the cage around frontier AI was built on purpose — and not by the people inside it.
On a Friday evening, the US government sent a letter. By the next morning, two of the most capable AI models on Earth were gone — not throttled, not rate-limited, gone — for every user on the planet.
We now have the letter. According to a copy seen by Bloomberg, US Commerce Secretary Howard Lutnick warned Anthropic’s chief executive that the company would need government permission before exporting its Fable 5 and Mythos 5 models to any destination worldwide — or to any foreign national, regardless of where on Earth that person sits — and threatened criminal and civil penalties for non-compliance, citing federal laws that allow export controls on civilian technology an adversary’s military could turn to intelligence use. There is no way to enforce “no foreign nationals, anywhere” selectively at speed, so Anthropic did the only thing compliance allowed: it disabled both models for everyone, globally — paying enterprise customers and its own staff included. In its own public statement, the company said it believed the order was a misunderstanding and was working to restore access as soon as possible.
Restoration is not the point. The point is the demonstration — and the precise legal shape of it. A government just reached into a private company and switched off its best product with a Friday letter, and the instrument it used was export control: the assertion that letting a foreign national so much as use the model is an export requiring a federal licence. Even if access returns tomorrow, what got repriced is jurisdictional risk itself — permanently.
The instinct everyone has
The natural reaction, if you run an AI company, is the same one anyone has when a landlord turns out to hold a key to your flat: leave. Re-headquarter somewhere else. Set up different entities. Get out from under a single government’s whim.
The commentary has already started writing the relocation listicle. Move to Canada. Go to the UK — lighter touch. Find a friendly tax haven outside US reach.
It is a completely understandable instinct. It is also a mirage — and understanding why is the most important thing anyone in this industry can internalise right now.
Two cages, not one
Start by separating two things the headlines blur together. There are two cages here.
The first is access control: who is allowed to use a deployed model. The Lutnick letter is access control in its rawest form — it doesn’t care where Anthropic is incorporated; it cares who is on the other end of the connection.
The second is substrate control: where the model’s weights, and the compute that produced them, can legally travel. This is the deeper, quieter cage, and it’s the one the offshoring fantasy runs straight into.
What makes the letter remarkable is that it welds the two together. By treating “grant a foreign national access” as “export without a licence,” the government collapsed access into substrate — usage into export. That is the deemed-export principle — releasing controlled technology to a foreign person counts as shipping it abroad — applied to a chatbot. And once usage is export, relocation is beside the point: you can move the company, but you cannot move the fact that the model was made of controlled silicon.
The leash is not attached to your address
The United States did not wire its jurisdiction to your headquarters. It wired it to your substrate — the chips, the compute, and the weights themselves. You can move the brass plate anywhere you like. The leash is attached to something you cannot carry in a moving van.
A caveat the lawyers will demand, and one that actually makes the point sharper: this regime is built around the frontier, not around ordinary software. The dedicated control on AI model weights — ECCN 4E091 — only ever reached closed-weight models trained on more than 10²⁶ computational operations, and explicitly excluded open-weight models and anything less capable than the strongest open model; at issuance, regulators believed no model then in existence even met the threshold. (That rule has had its own turbulence — issued in January 2025, then rescinded that May with a replacement still pending — which is precisely why the Lutnick action leaned on broader export authority rather than the weights rule itself.) Not every model becomes American by brushing past an Nvidia box. But the frontier models that matter most are exactly the ones the regime was built to catch — and the chips and compute beneath them stay controlled regardless.
And the cloud door is closing too. The Remote Access Security Act, which passed the House 369–22 and now sits in the Senate, would make the logic explicit: cloud-based access to controlled compute by a foreign person, treated as an export transaction. It is not law yet — but trade regulators are already moving in that direction under existing authorities. Rent the silicon from Dublin or Dubai, and the reach follows it home.
And there’s nowhere neutral to run to
Everything so far is one pole — the United States, following its own companies’ compute and weights wherever they travel. So the natural next move is to run the other way, toward a structure the US can’t see. Here the relocation fantasy hits its second wall, because the rival superpower runs the same instinct.
Manus, a Chinese-founded AI company, ran exactly this experiment: it dissolved its original Chinese entity, moved the core team to Singapore, and built a Cayman/Singapore holding structure — and it got blocked anyway. Not by Washington, but by Beijing, in the first publicly confirmed use of China’s foreign-investment security review to unwind a cross-border AI deal. A different instrument from a US export order, but the same instinct underneath — and it’s the instinct that matters: origin over domicile. Both capitals now look straight through the corporate structure to where the technology actually came from. Offshoring didn’t move Manus out of reach; it moved it into the crossfire. There is no neutral ground between the poles, because both poles run the same look-through.
So watch the relocation candidates collapse
- The UK and Canada are allies — which means they are being pulled into the US regime, not offered as an exit from it. Proposed US legislation would use the Foreign Direct Product Rule to extend jurisdiction over allied producers unless those allies adopt equivalent controls. The point is not that it blocks a lab from moving to London or Toronto; it is that the direction of travel is unmistakable — close ally arbitrage, don’t preserve it. (When Japan tried list-based controls without this mechanism, firms simply shifted production to subsidiaries in South Korea; the new rules exist to shut exactly that move.)
- Tax havens solve the wrong variable. A Cayman structure governs where profit is booked, not where compute lives. Manus was a Cayman/Singapore structure. It got looked through on substance.
- A genuinely non-US compute stack — non-US chips, non-US fabs, non-US capital, non-US persons — is the only thing that would actually work. But that is not a corporate-relocation decision you make over a weekend. It is a decade-long sovereign-industrial-policy programme, and right now the only state seriously attempting it is the one a Western lab cannot defect to, for the mirror-image reason.
Why the weaker model may win anyway
Here is the part the labs should be losing sleep over. Enterprises and governments do not simply buy the smartest model. They buy continuity, legal availability, and procurement certainty. And the Lutnick letter just handed every buyer on Earth a live demonstration that the best model on the market can be switched off — for them specifically — by a government that isn’t theirs.
This does not make foreign or open alternatives suddenly superior. It makes them institutionally less dismissible. A weaker model with sovereign continuity can beat a stronger model with a visible kill switch — especially for the banks, defence contractors, hospitals, and critical infrastructure that cannot build on something a foreign capital can revoke on a Friday. That is the real market repricing, and it does not reverse when the access comes back. A kill switch, once flipped, can’t be unseen.
The cage was the architecture
Here is the part worth sitting with. None of this is an accident of one administration’s mood. It is the design — and it was drawn deliberately, by the architects of export control, around the labs, not by the labs themselves.
If you build your control regime around an entity’s address, the entity moves and you lose. If you build it around the silicon and the weights — the two things a frontier lab physically cannot reconstitute somewhere else — then it doesn’t matter where anyone goes. Jurisdiction stops being a map and becomes a supply chain. The reason you can’t leave is that “leaving” was quietly defined out of existence before anyone thought to try.
And the “backup plan” every AI company is now told it needs? It was never a place. It’s a parallel supply chain — and almost no one on Earth can afford to build one.
You can check in any time you like: take the capital, rent the compute, train the frontier model, become indispensable. That door is wide open, and the welcome is warm.
But you can never leave.
Sources are linked inline throughout. This piece concerns export-control policy and corporate strategy, not legal advice. @blog