Sovereign AI: what the word really covers

The word doesn't mean the same thing depending on who says it

A hosting provider selling sovereign AI is talking about where the machines run. A software vendor selling the same thing is sometimes talking about the contract that stops it reusing your data. A model maker is talking about where the model comes from. These are three different promises, and none of them contains the other two.

Four questions separate these promises. Where do my documents live. Who analyses them. What leaves my premises when someone asks a question. And who builds the model that answers. A provider can answer the first three perfectly and not the fourth at all. That's our case, and we'd rather write it down here.

Where your documents live

Every company has its own instance: its own machine, its own database, its own files. Nothing is shared between two clients, not even the search index. One instance does not know the others exist.

These machines are in Europe, in Helsinki, with a German host. Not in France, and we do not pretend otherwise: we chose price and reliability over the flag. If your internal policy requires French soil or your own server room, the instance can be installed wherever you want, at the same price. It is a move, not a surcharge.

Who analyses your documents

When you connect a documentation source or upload a procedure, it has to be split up and turned into vectors for plain-language search to work. That work is done on your machine, by an engine installed on it. The 27,231 knowledge passages on our first firm instance were all computed on site.

Result ranking too: the model that decides which passage best answers the question runs locally. None of your documents leave to be read elsewhere in order to build the index. This is the level of sovereignty that gets forgotten most often, and it's the one that covers the largest volume of data.

What leaves the instance when someone asks a question

The question, and the extracts the local search judged useful. Not the knowledge base, not the whole document, not the company's history. Those elements go to the configured AI provider, today Anthropic, whose programming interface does not use what is sent to it to train its models.

Anthropic is an American company. That is the one point we cannot claim, and there is no reason to dress it up. In exchange, what leaves the instance is cut back to the bare minimum, and the provider is not carved into the product: the code calls an abstraction, not a brand.

Fully local, and what it really costs

The question comes up at every meeting: why not run the model in-house, and cut every link with the outside world. Technically, it's ready. We measured it in September 2026 on our standard machines, four cores and eight gigabytes, no graphics card: a seven-billion-parameter model produces three to six words per second.

Six words per second means a two-hundred-word answer takes more than half a minute. Nobody uses that twice. For chat to be pleasant, you need a graphics card, and the bill changes order of magnitude. So our position is pragmatic: today, local makes sense for background tasks that run overnight, sorting messages, preparing summaries, not for conversation.

This calculation will change. Small models are improving fast and hardware is getting cheaper. The day fully local becomes sensible, the switch is already built into the architecture, and it will not mean changing product.

The six questions to ask before you sign

They apply to us as much as to any other vendor. Is my knowledge base shared with other clients. Where is the machine physically. Is the indexing of my documents done on my side or somewhere else. What exactly is sent to the model, and in what form. Is what gets sent used to train anything. And if I leave, what do I take with me.

On the last one, our answer is simple: the instance is yours, the database is yours, and a full backup is handed over on request. A provider who hesitates on that question has answered all the others for nothing.