America is losing the AI race. They just haven't realised it yet.

In June, the United States government ordered Anthropic, America’s most valuable AI company, to withhold its most powerful model from every foreign national on earth. The company could not enforce that line cleanly through a live product, so it did the only thing available to it. It switched the model off for everyone, including Americans. This happened while the same company was preparing to go public at a valuation near a trillion dollars.

Almost everyone read the episode as a blunder, an own-goal, or one more reason not to trust Washington. I think that reading is comfortable, and I think it is wrong. The ban was not a mistake. It was a worldview doing exactly what it was built to do, and in the process it exposed that the worldview is false.

Here is the worldview, stated plainly. The United States believes a frontier AI model is a weapon. China believes it is infrastructure. The world runs on the technology it can depend on, not the most powerful one, and a weapon you forbid people to buy is just an expense.

Everything that has happened since February is downstream of that single category error.

The weapon theory

Look at what the weapon model produces. A national security narrative built around raw capability. Export controls. A vetted access list of approved organisations. A flagship model deliberately engineered to refuse questions about cybersecurity, biology and chemistry. And, the moment an exploit surfaced, an off switch. This is the enriched-uranium mindset: scarce, lethal, licensed, denied to adversaries. It is the chip-control playbook applied to software that copies itself for free.

If you accept the premise that AI is a strategic munition, every American action this year becomes rational. The refusal to let its leading lab keep the technology out of mass surveillance and autonomous weapons. The decision, when that lab pushed back, to brand it a national security supply-chain risk, a designation a federal judge found to be pretextual and an act of unlawful retaliation. The export order in June. All of it follows cleanly from one idea.

The idea is false.

Nobody asks whether their electricity has a PhD

The value of AI to the economy has nothing to do with possessing the single most capable instrument. It is realised by pouring the technology into the ordinary machinery of business, the invoices and the claims and the correspondence and the support queues, the work that is counted in payroll rather than in benchmark scores. There is an enormous amount of that work, and almost all of it can be done by a model that is merely good rather than exceptional. The frontier, the genuine lead the American labs still hold, is irrelevant to it. Nobody asks whether their electricity has a PhD. They ask whether it is on.

Availability is the property that decides the market, and it is the one property the weapon theory cannot deliver. A company does not run AI in a demonstration. It wires the technology into the operations it depends on every day, and no serious operator will build a production process on a system that can be switched off for weeks because a foreign government is arguing with itself. After June, that risk is no longer hypothetical. It has a date and a timestamp.

A capability you can be denied is not infrastructure. It is a liability with good marketing.

China built the grid on purpose

China made the opposite choice, deliberately, as state policy. Its current five-year plan names artificial intelligence a national priority on the order of defence, and sets targets for the technology to reach most of the economy by the end of the decade. The method is commoditisation. Release strong models with open weights, under the most permissive licences available, at prices that fall toward the cost of the electricity needed to run them. Make the technology abundant, ambient and impossible to own.

In the same fortnight that Washington pulled its model, several Chinese laboratories shipped open-weight models. The strongest was published in full, under a licence that lets anyone on earth download it, run it on their own hardware, and keep running it forever. No government can recall it, because there is nothing to recall. It is not as capable as the best American model. It does not need to be. It is good enough for the work that pays, it costs a fraction as much per token, and it cannot be taken away.

This was already working before the ban. By spring, on the largest neutral marketplace through which developers reach models, a clear majority of all usage already ran on Chinese open-weight systems, and the open-weight standard of the Western world two years ago had dropped off the board entirely. The strategy did not need a catalyst. What the United States supplied in June was not the defeat but the proof, and the one argument powerful enough to move a cautious enterprise off a familiar supplier. The incumbent is not reliable, and the alternative cannot be confiscated.

The weapon that pays for itself

There is a reason this will not be corrected, and it is not stupidity. It is money.

The proposition that justified the ban, that the model is so capable it threatens national security, is the same proposition that justifies the trillion-dollar valuation. The danger narrative and the investment narrative are one sentence. A lab that tells the world its model could supercharge a cyberattack is also telling its investors that it owns something scarce and strategically indispensable. So the American sector is commercially committed to insisting its product is too dangerous to be free, and that insistence is precisely what drives the world’s buyers toward the people giving the technology away. The story that makes America rich is the story that makes America lose.

Unwinding it would mean deflating the valuations built on top of it. An entire industry, and a great deal of the market’s faith, now rests on the premise that frontier capability is a scarce strategic asset. The market is demonstrating, quietly and at scale, that for almost all real work it is an abundant commodity. Both things cannot be true, and only one of them has a trillion dollars riding on staying false.

What America actually lost

So let me say it without the hedging that usually surrounds this subject.

This is a Chinese victory. Not a victory of capability, where China remains the runner-up and the American labs are genuinely ahead, but a victory of strategy. China understood that the value of this technology would come from spreading it rather than hoarding it, built for that future on purpose, and waited for the United States to disqualify its own best product. It did not need to win the frontier. It needed America to keep believing it was guarding a weapon, and to act on that belief until the weapon became something no one could depend on.

The off switch was never the disease. It was the symptom of believing you own something that, by its nature, cannot be owned. China did not beat America to a better weapon. China understood that the thing was never a weapon at all.

What remains genuinely open is whether the world this produces is the better one. An AI layer that is cheap, open and beyond any government’s power to recall is, in the same breath, more resilient and less governable. The capabilities that frightened Washington into reaching for the off switch are real, and in an open world no one holds the switch, not Washington, not Beijing, not a court. That is the trade the next decade turns on, and it deserves a serious argument.

But it is no longer America’s argument to settle alone. That is the true consequence of June. Not that the United States lost a product for a few days, but that it forfeited the authority to set the terms, and proved, to anyone watching the wiring instead of the benchmarks, that it had mistaken the nature of the thing it was trying to win.

Written from the perspective of someone who builds with these systems rather than trading in them. The only question that has ever mattered to me is whether the technology works on Monday, and is still there on Tuesday.