America the AI Gatekeeper: How the US Controls Frontier Models

Talk of an ignominious peace deal with Iran signalling American decline misses the more consequential shift: the US government has just demonstrated it can decide who gets to use frontier AI. On June 12 the Trump administration ordered Anthropic to block foreigners from Fable and Mythos, its latest and most capable frontier models.

What happened on June 12 and why it matters

The administration said it was responding to a supposed jailbreak for Fable — a prompt that circumvents defences against harmful uses such as hacking computers or making bioweapons. The chances are that officials wanted Anthropic to switch off the models for everyone, and that targeting foreigners was a means to an end.

That is what Anthropic did, while saying the concern about its model was overblown. The legal basis of the order remains unclear, and the ban seems unlikely to last.

What matters is the precedent: global access to the best AI may come down to a decision in the Oval Office. In March the administration showed it is prepared to pressure the frontier companies when it designated Anthropic a “supply-chain risk.” Now it has shown it is prepared to press users, too.

Why Washington is now the decisive AI gatekeeper

America must decide how to wield this vast new power. The rest of the world must decide what to do about it. Even as governments plan for a less reliable America on defence and trade, they now face a new form of dependence on the world’s biggest economy — through AI models and the computing power needed to run them.

A short history of US controls on frontier technology

This is not the first time America has tried to restrict access to frontier technologies. After the second world war it stopped helping Britain’s nuclear-weapons programme. When modern cryptography emerged in the 1970s, it blocked exports before accepting the trade-off between keeping allies secure and using secrecy to boost its own offensive capabilities.

Washington still withholds its best military equipment even from close allies. America kept the F-22 fighter for itself; allies received the F-35. That pattern — selective sharing tied to strategic risk — is the template for AI.

When controls work and when they fail

Whether control endures depends on the technology and on rivals’ ability to develop it independently. Nuclear co-operation with Britain resumed in the 1950s after Britain developed its own capability; with other countries, America relied on the Nuclear Non-Proliferation Treaty of 1968. Cryptography methods could not be contained and eventually went public. Many countries can already mount cyber-attacks, limiting how far any single export control can reach.

What makes frontier AI different — and familiar

Frontier AI echoes all those cases. If the very best models can disable crucial infrastructure or help users create pandemic-ready pathogens then, like nuclear weapons, they are too dangerous for unrestricted public use. Yet as with cryptography, it will be hard to ensure that advanced proprietary capabilities are never copied.

Open-weight models, which anyone can download, could advance and proliferate. In cyber-security a small imbalance can be decisive: if an attacker has version 5 while the defender is stuck with version 4, and the better model uncovers just one more vulnerability, the weaker party will be compromised. As AI is embedded in military hardware, a similar logic may apply on the battlefield.

The economic pressure to keep selling AI abroad

America also has a huge economic interest in leading AI and selling it overseas. Many Anthropic staff are not American and were hit by the ban; freezing research at America’s leading lab would be self-defeating. The firm says 80% of its consumer use is overseas.

As American technology has boomed over the past decade, Europe’s payments to America for intellectual-property products have risen fivefold. Washington has a clear incentive not to give the rest of the world a reason to team up with China, the second-ranking AI power.

A hierarchy for access: from guarded to handicapped models

The result may be a hierarchy of access. The best capabilities will be closely guarded by America to provide an edge in cyber-offence and military use. The next-best alternatives may be made available to allies — the equivalent of the F-35. A sufficiently handicapped model, with the strongest safety precautions its makers can design, may then be sold to the wider world.

Why allies are exposed — and where leverage remains

Such a future would be uncomfortable for America’s allies, which are nowhere close to rivalling Anthropic. They are already vulnerable to President Trump’s bargaining on trade, alliances, the dollar system and more. AI could become the most important lever of all.

Some dependence is two-way: America needs Dutch lithography machines and Taiwanese fabs. And foreigners could shun Anthropic or OpenAI in favour of good-enough open-weight alternatives. Those options limit, but do not erase, US leverage.

The compute gap that locks in dependence

Running frontier models requires computing power, which America has perhaps 15 times more of than Europe, plus much larger ongoing investment. If SpaceX, whose share price surged after listing on June 12th, realises its vision of data centres in space, the gap is unlikely to close.

“Europe 2031”, a gloomy essay on AI’s future, imagines vassal status for Europe as its cyber-security, defence and swathes of its economy come to depend on American models and compute. The warning is less about any single model than about control of the stack that makes advanced models usable at scale.

How Europe and East Asia can build bargaining power

Many countries will conclude they need America more than ever. Even so, they can strengthen their position. Compute is so scarce that it makes sense to build data centres almost anywhere, yet many governments show little sense of what is at stake.

In Britain, that demand rose from 41GW to 125GW in those seven months, more than twice peak power demand. OpenAI paused a data-centre project there in April, citing regulation and costly energy.

Throughout Europe, more energy, easier planning and looser rules for modelmakers would help, rather than token gestures. In East Asia, Taiwan, Japan and South Korea should integrate rather than duplicate their efforts. All governments should avoid the false promise of centrally steering investment.

The goal should not be protectionism but ensuring America is not the only economy where the AI ecosystem can thrive. As in trade and defence, the response to Washington’s transactional turn is not to lament alliances but to build strength.

Figures on shipments, models and grid connections are as reported in the sources cited at the time of publication.

Report a correction