The first question from a client thinking about AI has long stopped being which model. It is where the data goes. Until now the answer was almost always: somebody else's datacentre across an ocean.
Fact
The facts: on August 11 Mistral announced general availability of its regional endpoints - the customer picks Europe or the US and processing stays in the chosen region. It simultaneously launched the European Compute Units programme, converting long-term commitments into access to compute infrastructure; early participants include ASML, CMA CGM, Amadeus and Caisse des Depots. The stated target is up to 1 gigawatt of capacity by 2030. Per company data, Z.ai's open GLM-5.2 model will also run on its infrastructure.
Their definition of sovereignty comes in four verbs: you control the model, choose where the intelligence runs, own the compute to scale and keep the value with you. It reads like a speech. Underneath it, though, sits something checkable: a processing region you can name in a contract.
For a company here that is the difference between explainable and unexplainable in front of a client. The data stays in Europe is a sentence that goes into a contract and into a regulatory assessment. The data goes somewhere goes nowhere.
One clarification the announcement itself does not make: where exactly this capacity will sit, how much money is going in and under what licence. The gigawatt is for 2030. That is four years and a lot of press releases away.
Sovereignty is not a slogan. It is checked with one question: where does the compute run.
Ask your provider that question this week. If the answer is vague, that is already an answer.
Official primary sources
Mistral AI - Regional inference, open models, new computeThe visual is generated code-art, no third-party images.