If you lost your LLM tomorrow, would you still have a business?

On a Saturday in the middle of June, Anthropic switched off its two newest models for every customer on the planet. Not throttled. Off. The US Department of Commerce had decided that foreign nationals should not have access to Claude Fable 5 and Mythos 5, and since nobody can verify a user’s nationality in real time, the only way to comply was to turn the lights off for everyone. They came back on 17 days later.

So if I were starting a European AI company today, my first question would not be about the product. It would be what happens the morning I cannot reach the model, the whole thing runs on. A year ago, that question would have sounded like the sort of thing a risk committee asks to justify its existence. It does not sound like that now.

“AI sovereignty” is a Brussels phrase. It sounds like a working group, and it usually is one. But strip the politics out, and what is left is a question any competent engineering lead should be able to answer on the spot: what breaks if this supplier stops answering the phone, and how long does it take us to find an alternative?

None of this is an argument against using them. OpenAI and Anthropic have built genuinely remarkable things, and two people with an API key can now ship in a quarter what used to take a floor of engineers a year. I would use them too, and I do.

The problem starts when the whole company is downstream of one account with one supplier.

Governments restrict things. Providers retire models, reprice them, or quietly change what you are allowed to do with the output. None of that requires anyone to have a grudge against you. It only requires you to be a rounding error in somebody else’s compliance decision, which, if you are a 15-person company in Zürich, you are.

The level of vigilance required may depend on how heavily regulated the industry is. In financial services, for example, a brokerage that cannot explain what happens when a supplier vanishes may find itself having a short and educational conversation with its regulator. That makes it important to retain the ability to run models on internal infrastructure, test what is currently available rather than assume last year’s solution still applies, and plan for the possibility that something in use one day may no longer be available the next.

Open weights make that possible in a way it was not two years ago. Mistral in France, Gemma from Google, DeepSeek and Qwen out of China, all downloadable, all runnable on hardware you can point at. The quality gap to the frontier is real. It is also narrower than it was last time you checked, and for a great many jobs, it does not matter at all.

Almost no European startup should be training a foundation model. The numbers do not work, and the ones telling you otherwise are usually capital raising. What founders should be able to do is name their dependencies out loud, and say what they would run instead, and roughly how long the swap would take. That is not a strategy. It is housekeeping.

Building the alternatives here is harder than it should be, and it fails in the same three places every time. Money. Somewhere to plug it in. People who will stay.

I have been on the wrong end of the first one. In 2020, I co-founded a company in Europe and spent longer trying to raise money here than I did building the product. In the end, we went to the US, where the meetings were shorter, and somebody said yes. The gap matters more with AI than it did with whatever I was selling, because training a frontier model means spending an extraordinary amount before anyone can tell you whether it works.

The second one is duller and worse. An AI data centre is a shed full of very hot metal that needs an enormous and uninterrupted supply of electricity, and in most of Europe, getting a grid connection of that size is a multi-year exercise in planning law. Ministers can call AI strategically important as often as they like. The transformer still arrives when it arrives.

To be fair, something is being done. The EU now has 19 AI Factories and 13 Antennas attached to the EuroHPC supercomputers, explicitly prioritising access for startups and SMEs that would otherwise never get near that kind of compute. It is a genuinely sensible policy, which I appreciate is not a sentence I write often about EU industrial strategy.

Which leaves the people. London and Zürich have serious AI clusters. The problem is not that Europe fails to produce good researchers. It produces them reliably and then watches a decent share of them take a package in California, or stay put and take one from the European office of an American company, which amounts to much the same thing from a sovereignty point of view.

We train them. Somebody else books the revenue.

Europe has no shortage of AI startups. Most of them will never train a frontier model, and there is no reason they should.

The question is how much of the thing they actually own. If the company has a nice interface, some prompt engineering, and a wrapper around an API that anyone can also call, then two questions follow, and neither of them is comfortable. How long would it take a competent team to rebuild this? And what is left of the product on the morning the model underneath it goes away?

For a lot of narrow jobs, a smaller model is fine. For some of them, a smaller model is better, because it is cheaper and you can run it on a machine you own, in a building you can walk into.

Some companies must do the harder work of building models so that Europe is not just a customer. This requires money, grid connections, and people. This list will be the same in five years unless something changes. Europe also needs companies that are attractive enough to keep researchers from considering a flight to San Francisco as the obvious career move.

None of this requires anyone to be dramatic about it. It requires Europe to have enough of its own options that a decision taken in Washington on a Saturday is an inconvenience rather than an outage. For founders, it starts somewhere much smaller than that. Ask what you would run tomorrow if the model you use today stopped answering.

If the answer takes longer than a minute, you do not have a fallback. You have a supplier and a hope.