Imagine that your company has spent the past year getting used to a remarkably capable assistant. It helps write reports, searches your documents, explains awkward bits of code, prepares presentations and turns a week’s tedious preliminary work into an afternoon’s. It is not quite a colleague, but it has become part of how the place thinks.
Then, one morning, it is gone: not because the company has missed a payment, not because the service has failed and not because anyone has done anything wrong. It has disappeared because a government in another country has decided that people with your nationality should not be allowed to use it.
Something close to this happened earlier this month. The US government instructed Anthropic to suspend access to its newest models, Fable 5 and Mythos 5, for foreign nationals, including foreign nationals inside the United States. Anthropic could not easily distinguish everyone affected, so it removed the models from all customers.
The restrictions were later partly relaxed. Mythos 5 was allowed back for a selected group of approved American organisations, while Fable 5 remained subject to negotiation. That did not really undo the point. It made it clearer. The US government had not merely regulated a product. It had decided which organisations could use a particular form of intelligence, and which could not.
Then OpenAI announced that it too would delay the general release of its newest model, GPT-5.6. At the request of the US government, it began with a limited preview for a small group of trusted partners whose participation had been shared with the government. The details were different from the Anthropic case, but the direction was the same. A frontier model was no longer simply something a company developed and sold. It was becoming something that first passed through the hands of the American state, which might decide who received it, when and on what terms.
This should not really surprise us. The cloud has never been the floating, placeless realm suggested by its name. It is somebody else’s computer, as the old joke has it, but that is only half the story. It is somebody else’s computer in somebody else’s country, under somebody else’s laws.
We have been quietly accepting this arrangement for years. European firms store data on American cloud services. Their staff communicate through American platforms, their customers pay through American systems and their phones depend on American or Chinese app stores. Much of the machinery through which modern institutions function is privately owned and geographically located somewhere else.
Usually that does not matter very much. A cloud provider may raise its prices, alter its terms or suffer an outage, but the result is mostly inconvenience, expense and several meetings in which people discover that nobody has kept a proper copy of anything. Firms can move their data, rewrite systems and find alternatives, eventually.
AI is different because it is beginning to sit closer to the centre of everyday work. A capable model does not merely store files or host a website. It helps people write, search, analyse, translate, code, plan and make sense of the things they already know. Once it is woven into an organisation, removing it is not like losing a piece of software. It is more like discovering that an entire junior department has been sent home.
That does not mean the American government is uniquely wicked for exercising power over technology developed in America. China is hardly famous for its airy indifference to where its digital infrastructure lives. The point is not that states have laws. It is that we have spent years pretending that digital services somehow existed above them.
The Fable and GPT-5.6 episodes matter because they make the dependency visible. A company or university using an American frontier model has not acquired that capability. It has acquired permission to use it, for as long as the provider, the US government and the wider political climate allow.
That may be perfectly sensible for many tasks. The best model is often the best model, and there is no virtue in refusing useful technology because it comes from abroad. Europe does not need to build its own Fable 5 in every capital city. But it does need options.
A serious institution should be able to move between models without rebuilding everything it does. Its documents, prompts, tools and internal knowledge should belong to it rather than disappearing into a supplier’s private machinery. Important work should have a local or regional fallback, even if that fallback is slower or less capable than the best American service. Governments, universities and large firms should be helping to build computing capacity that remains usable when a foreign provider changes its terms, its prices or its government.
This is where open models matter. Downloading a model does not conjure up the hardware, electricity, skills or data needed to use it properly. A locally run model may be weaker and less polished than the best service available from California. But there is a profound difference between a capability you possess and one you rent.
The first personal computers were not as powerful as the machines in universities and corporations. That was not their point. Their importance was that ordinary people, small firms and schools could finally own useful computing power rather than applying for time on somebody else’s mainframe. AI may follow a similar path. At first, independent capacity will belong mainly to states, universities and large organisations. Later, smaller businesses, schools and perhaps households may be able to keep a capable machine of their own.
That is not a future in which every family trains a frontier model in the attic beside the Christmas decorations. It is a future in which people have enough local intelligence to remain useful, informed and independent when a distant service is no longer available.
The cloud was never neutral. It was always somewhere, and somewhere has a flag.