AI without US vendors: what actually works in 2026 — and what does not

The question now comes up in every other conversation: can we do this in Europe too? The answer splits awkwardly in two, and most writing on the subject mentions only one half.

The good half: the model provider

Moving from OpenAI to a European provider is technically trivial. Mistral, IONOS and Scaleway all offer OpenAI-compatible endpoints. In most applications two lines change: the base URL and the key.

For the tasks companies actually have — summarising, drafting, classifying, extracting, translating, asking questions of your own documents — the available European models are sufficient. Not in every benchmark, but in practice. The gap to the frontier is widest on long chains of tool calls and on programming, and narrowest on anything that is language work in the narrower sense.

For embeddings and reranking, the parts every document search actually hangs on, Berlin's Jina AI is a serious option. For images, Freiburg's Black Forest Labs is competitive. For translation, DeepL has been better than the American competition for years.

The uncomfortable half: the compute

All of these models run on NVIDIA accelerators. There is currently no European alternative at relevant scale, and anyone claiming otherwise is selling either subsidies or hope.

Which means: you can change the jurisdiction your data is processed in. You cannot change the supply chain that processing rests on. That is a real difference, but it is a different difference from the one the CLOUD Act is about.

For most companies the first half is the one that matters. Access to your prompts and your documents is a concrete risk. An export ban on data centre GPUs is a macroeconomic risk that no single company can do anything about anyway.

What we concretely recommend

Provide an endpoint before you forbid anything. Every ban policy we have seen produced shadow IT. The tools are reachable through a browser; a ban only moves usage to where nobody can see it.

Point that endpoint at a European provider. If you are building a central endpoint anyway, the European variant costs nothing extra and satisfies the AI Act's documentation duty at the same time.

Split the cases. For research and drafting, model quality matters more than jurisdiction. For anything touching customer data, personnel data or engineering data, it is the other way round. That split is cheaper than one uniform rule that is either too strict or too loose.

Run open weights where it genuinely counts. A mid-sized model on your own hardware inside your own network is usable for narrow tasks and answers the sovereignty question completely. It is not the answer for a general assistant, and it should not be sold as one.

The honest state of play

European AI is no longer a sacrifice. It is also not the frontier. For the tasks that make up 90 per cent of usage in a normal company, the difference is by now smaller than the difference between a good integration and a bad one.