Inference never crosses the border because the model is a container next door

A notary chain in Lyon wants a document assistant with one condition: no inference outside France. Every hosted-model contract they've seen fails it somewhere in the subprocessor annex.

You rent a box in Paris. One container runs Ollama as a daemon, with the model weights on its disk. The agent container points its provider config at that container's URL, bring-your-own-provider style, with no gateway in the path and no upstream vendor receiving a prompt.

Prompt, retrieval, generation, and the deed being reasoned about are all on one machine, in one datacenter, whose city is in the rental listing. The subprocessor annex for AI inference has nothing in it.

The model lives where the documents live, so the question of transfer never comes up.