Forfeiting the Frontier: How Washington Is Ceding Its Own AI Advantage

Sat, 13 Jun 2026 00:00:00 +0000

Forfeiting the Frontier: How Washington Is Ceding Its Own AI Advantage

On 12 June 2026, Commerce Secretary Howard Lutnick ordered Anthropic to suspend its two most capable models, Fable 5 and Mythos 5, for every foreign national. The ban covered foreign nationals abroad and inside the United States alike, including the company’s own non-citizen employees. No provider can sort a live user base by nationality in real time, so the practical effect was a worldwide shutoff of both models. Anthropic’s statement says the order cited “national security authorities” but gave no written detail, only a verbal account of a narrow technique.

This episode will be litigated and probably resolved; Anthropic expects to restore access. But the important question is not whether one directive survives a court challenge. It is what a pattern of maneuvers like this does to the very thing the United States probably wants to protect: its lead in frontier AI. These maneuvers erode that lead, and that they do so precisely because of what they are: selective, improvised, and aimed at a domestic company rather than a foreign adversary.

How a lead actually erodes

A technological advantage is rarely lost the way people fear it will be, with a rival leaping ahead overnight. It erodes more slowly, from underneath. A frontier lead is not held by any single model or company. It is held by an ecosystem: the researchers willing to build their careers in one country, the investors willing to fund decade-long bets there, and the market on which raw capability turns into products, standards, and dependencies that the rest of the world comes to rely on. Capability itself moves easily from place to place. Ecosystems stay put, right up until they don’t.

What dislodges an ecosystem is not a single shock but a loss of confidence that the place is predictable. Consider the people who sustain a frontier: a founder choosing where to incorporate, an investor pricing a fifteen-year position, a researcher deciding where to spend a career, a finance ministry weighing whether to build national systems on top of an American model. Each of these is a long-term bet. And each depends less on any single government action than on how erratic the government’s behaviour is overall. A country can absorb a great deal of policy that is heavy-handed but predictable and still remain the obvious place to build. What it cannot absorb is the impression that a model serving hundreds of millions of people on a Tuesday can be switched off by letter on a Friday, on contested grounds, against one company. That impression is what dissolves the ecosystem.

These are maneuvers, not policy

It’s true that Mythos is no ordinary chatbot. Anthropic itself describes it as having found previously unknown security flaws across major operating systems and browsers. That is a real capability, and it genuinely alarmed regulators.

But the order fits neither the scope nor the logic of a genuine security measure. The model most users lost was Fable 5, the public and heavily safeguarded one. The stated reason was a “jailbreak” that, by Anthropic’s account, exposed only minor flaws that were already known, the kind of flaws that other public models, OpenAI’s GPT-5.5 among them, reveal even without any such jailbreak. A genuine security measure applies to a whole class of products. This one named a single company while its closest competitor kept operating without restriction.

That selectivity is not an accident; it is the pattern. The dispute goes back to February, when Anthropic refused to let the Pentagon use its models for autonomous weapons or domestic surveillance. The President then ordered every agency to drop the company, and the Pentagon branded it a “supply-chain risk,” a label normally reserved for foreign adversaries. OpenAI secured a Defense Department deal within hours. A federal judge later blocked those moves, finding that they “appear designed to punish Anthropic” and amounted to “classic First Amendment retaliation.”

The clearest sign is what the same government does with the same capability when it serves its own ends. The administration is reportedly running Mythos at the NSA, with Anthropic engineers embedded to support it, and the Pentagon’s own technology chief has called Mythos a “separate national security moment” that the government must harness, even as the supply-chain label remains in place. A capability cannot be too dangerous to show a foreign national in a coffee shop and, at the same time, important enough to run in the nation’s own cyber operations. In fairness, the order followed an ignored request to delay the launch, so the government did have concerns and a channel for raising them.

However, this kind of incoherence is exactly what long-term decisions weigh most heavily. A strict but consistent set of rules is a known cost, one the ecosystem can plan around. A government that bars a domestic company from offering the very capability the government itself runs, and that a rival sells freely, does not impose a known cost. It introduces unpredictability, and unpredictability is what makes a place the wrong one to build in.

Both exits cost the United States

If Anthropic stays put, it operates at the discretion of an administration willing to use its procurement and export powers against a vendor over a contract dispute. That is a precarious position for a company preparing a U.S. listing and financing tens of billions of dollars in computing power. But capability does not stay locked inside one company’s distribution. If it cannot be deployed at scale from within the United States, the frontier simply advances somewhere with fewer constraints. The rational response for any government watching is what is often called “sovereign AI”: the principle that a nation should not build its critical systems on top of technology that a foreign executive can switch off overnight. Every maneuver of this kind strengthens that argument, and funds the non-American alternatives that follow from it.

If Anthropic instead reduced its dependence on the United States, the disruption would be larger still. It would lose the U.S. market the moment a security pretext was invoked. More importantly, the American companies it connects to, the major cloud providers and enterprise channels, would be barred from dealing with a banned foreign entity, and that is a far larger commercial loss than any single contract. The honest counterpoint is that Anthropic’s reliance on U.S. capital and U.S. chips makes a clean exit more theory than plan. But that is exactly the point. A national champion that cannot easily leave is not a reassurance. It is a live demonstration, for everyone still deciding where to build, of how costly it is to be trapped.

What gets priced in

The narrow cost here is the disruption to Anthropic’s customers, which the company says it is working to reverse. The deeper cost falls on the one factor that actually decides who leads a decade from now: the willingness of talent, capital, and allied governments to commit, irreversibly, to the American frontier. That willingness depends on predictability, and predictability is exactly what a series of selective, improvised interventions destroys. Veteran critics have warned of precisely this, about this very action: they expect it to push foreign-born researchers back to their home countries and to make American AI look, to investors, like an unsafe bet. The stated goal of the entire posture is to stay ahead of China. Undermining a domestic frontier lab, while a competitor and the government itself use the same capability, is the surest way to hand that lead away.

Export controls are a tool for denying a capability to an adversary. Aimed instead at a domestic champion, over a capability the same government is running and a rival is selling without restriction, the tool points the wrong way. The capability does not disappear. It moves elsewhere, and the advantage (that was never the capability itself but the conditions that attract it) moves along with it.

Export Controls on Colan Schwartz on Cloud Architecture, Security, Privacy & Startups

Forfeiting the Frontier: How Washington Is Ceding Its Own AI Advantage

Forfeiting the Frontier: How Washington Is Ceding Its Own AI Advantage

How a lead actually erodes

These are maneuvers, not policy

Both exits cost the United States

What gets priced in