The emerging geopolitics of artificial intelligence has produced a genuinely novel diplomatic problem: allied nations want American AI models—they are, after all, the best in the world—but they are increasingly unwilling to accept the terms on which Washington insists they be deployed.
The core issue is sovereignty, though nobody wants to use that word. When a European ministry or Asian central bank integrates an American large language model into critical infrastructure, it does so knowing that the US government retains theoretical authority to restrict, modify, or terminate access. This was an abstract concern until recently. It is no longer abstract.
The kill-switch question
American AI companies have built their enterprise offerings around cloud-hosted inference, which means the model runs on servers controlled by the provider. This architecture is efficient, secure in certain respects, and allows for rapid updates. It also means that a sufficiently motivated US administration can, through export controls or direct pressure, degrade or cut off access to any foreign customer.
Allied governments watched the semiconductor export restrictions imposed on China and drew the obvious conclusion: today's partner is tomorrow's potential target. The Trump administration's recent moves against certain Anthropic model deployments—ostensibly over content-policy concerns—only sharpened the anxiety. If Washington can dictate what an AI model may discuss, it can dictate rather more than that.
The result is a quiet but accelerating demand from European, Japanese, and Gulf state governments for what might be called sovereign deployment: models that run on local infrastructure, with weights held domestically, and without any technical mechanism for remote American intervention.
The American position
Washington's counterargument is not unreasonable. Frontier AI models are dual-use technologies with genuine national-security implications. Allowing weights to proliferate beyond American oversight creates risks that range from the mundane (intellectual-property theft) to the catastrophic (weaponization by hostile actors who obtain the weights through allied intermediaries). The kill switch, in this framing, is not imperial overreach but responsible stewardship.
American AI companies are caught in the middle. They want the revenue that comes from foreign government contracts, but they operate under US jurisdiction and cannot simply ignore Washington's preferences. Some are exploring compromise architectures—on-premises deployments with contractual rather than technical restrictions—but these satisfy no one fully.
The fragmentation risk
The deeper danger is that this standoff accelerates the fragmentation of the global AI ecosystem. If allied nations conclude that American AI comes with unacceptable strings, they will invest more aggressively in domestic alternatives or turn to Chinese providers who, whatever their other drawbacks, do not answer to the US Commerce Department.
This is already happening at the margins. France's push for European AI sovereignty, Japan's substantial investments in domestic model development, and the UAE's courtship of Chinese AI firms all reflect the same calculation: dependence on American AI infrastructure is a strategic vulnerability.
Our take
The United States built the world's best AI industry and now faces the classic hegemon's dilemma: how to maintain influence without provoking the resistance that erodes it. The kill switch may be technically necessary and strategically rational, but it is also a constant reminder to allies that their digital infrastructure ultimately answers to Washington. At some point, the reminder becomes the problem. The smarter play would be to negotiate explicit, treaty-level guarantees about when and how access restrictions might be imposed—turning an implicit threat into a bounded, predictable framework. But that would require treating allies as partners rather than customers, and American AI policy has not yet made that leap.
