The World Wants American AI, Just Not on a Kill Switch

French President Macron and Indian PM Modi both brought up the same uncomfortable topic at the G7 summit last week: what happens if the U.S. decides to pull the plug on American AI services?

It’s not a hypothetical anymore. The Anthropic blackout earlier this month showed exactly how fast access can vanish — no warning, no recourse, just a sudden silence from the model you were relying on. That incident turned a theoretical concern into a practical one for governments and businesses outside the U.S.

Macron framed it as a sovereignty issue, which it is. If your entire AI infrastructure depends on APIs or cloud services from American companies, you’re effectively renting intelligence from a foreign power. Modi went further, suggesting that nations need their own AI stacks — not because they don’t want American technology, but because they can’t afford to be held hostage by it.

The irony is that American AI is genuinely good. OpenAI, Anthropic, Google, Meta — they’ve built models that most countries can’t replicate easily. The training costs alone are prohibitive, let alone the talent and compute required. So the world wants access. They just don’t want the access to come with a kill switch.

This isn’t a new tension. We’ve seen it before with cloud infrastructure, with semiconductor supply chains, with GPS. But AI feels different because of how quickly it embeds into critical systems — healthcare, finance, military logistics. Once you integrate a foreign AI into your national infrastructure, you’re not just using a tool. You’re depending on someone else’s goodwill.

The Anthropic blackout was a wake-up call. It wasn’t politically motivated — it was a technical failure, which almost makes it worse. It proved that even without malicious intent, access can disappear. If a bug can take down your AI assistant, what happens when a sanctions list gets updated?

Some countries are already moving. The EU is pouring money into open-source models and local compute clusters. India has its own AI mission with a focus on building foundational models trained on Indian languages. France is backing Mistral and other European players. But none of them are at parity with the U.S. yet, and they know it.

What I find interesting is that nobody at the G7 proposed cutting ties with American AI entirely. That’s not realistic. Instead, the conversation has shifted to redundancy and resilience — can you run your critical workloads on a U.S. model while maintaining the ability to switch to a local alternative if needed? That’s harder than it sounds, because most AI workflows are deeply integrated into specific ecosystems.

The real question is whether the U.S. will treat this as a diplomatic problem or a business opportunity. So far, the response from American AI companies has been defensive — arguing that their models are the best and that concerns about access are overblown. That might have worked before the Anthropic blackout. It rings hollow now.

I don’t think the answer is for every country to build its own GPT-5. That’s wasteful and probably impossible for most. But there’s a middle ground: shared governance, local fine-tuning, guaranteed access clauses in contracts. The technology exists. What’s missing is the political will to treat AI access as critical infrastructure rather than a consumer product.

Macron and Modi are right to raise the alarm. The U.S. has the best AI in the world, but that doesn’t matter if the rest of the world can’t trust it to stay on.