What is sovereign AI?

A one-sentence definition

AI is sovereign when the organisation using it keeps effective control over the three layers it is made of: the hardware running the computation, the model producing the answers, and the data feeding both.

The important word is "effective". A contract promising confidentiality is not the same thing as an architecture where data physically never leaves the building. The first rests on good faith and on the law that applies to the vendor. The second rests on nothing but network topology.

Why this became concrete

For ten years, moving to the cloud was the default answer. Generative AI changed the equation, because it does not process peripheral data: it processes strategic documents, client files, contracts, internal minutes. What you send to a remote model is the raw material of your business.

Three developments made the question unavoidable. First, the extraterritorial reach of certain laws, including the 2018 US Cloud Act, which lets authorities compel a provider subject to their jurisdiction to hand over data stored abroad. Second, European case law: the Court of Justice of the European Union struck down two successive transfer frameworks to the United States, Safe Harbor in 2015 and Privacy Shield in 2020. Third, economic dependency: when a vendor changes its pricing, its terms of use, or retires a model, you absorb it.

There is also an operational reality that is often underestimated: with no connection, cloud AI simply stops existing. For a firm, an agency or a factory, a network outage becomes a production outage.

The three levels of sovereignty

Not every offering that calls itself sovereign sits at the same level. The distinction matters, because the gap in protection between them is considerable.

• Data sovereignty: your data is stored in a given jurisdiction. This is the weakest level, because where a server sits says nothing about the law that governs the company operating it.
• Infrastructure sovereignty: the hardware belongs to an entity whose legal attachment you control, or better, it sits on your premises.
• Model sovereignty: the model weights run on your side, with no outbound call to a third-party API. Nobody can change its behaviour, restrict it or retire it without your agreement.

Sovereign does not mean closed

A common confusion sets sovereignty against openness. The opposite is true. Open models, whose weights are publicly distributed, are precisely what makes sovereignty achievable: you can download them, run them on your own hardware, audit them and pin a version whose stability you guarantee.

Nor is sovereign AI AI cut off from the world. It is AI where you decide what goes out. The difference is between a locked door and a door you control.

What it takes in practice

For a long time the honest answer was: a server room, GPUs, an MLOps engineer and several months. That entry cost is what kept most organisations in the cloud, for lack of a realistic alternative.

Hardware has caught up. A superchip like the NVIDIA GB10 Grace Blackwell delivers up to 1 PFLOP at FP4 precision with 128 GB of unified memory, in a desktop enclosure running off a standard socket. That unified memory is the key point: it loads models that used to require several professional graphics cards.

This is the principle behind DIWY. The box arrives preconfigured with Devana OS, the system layer that runs Suite 366 and Devana fully offline. You plug it into your network, your teams connect, and no request ever leaves your premises.

Questions to ask a vendor

Before selecting a solution presented as sovereign, four questions are usually enough to remove the ambiguity.

• Where does inference physically run, and which law governs the entity operating that hardware?
• Does the service keep working if the internet connection is cut?
• Can I audit outbound traffic and confirm that no data leaves?
• What happens to my usage if the vendor changes its terms or disappears?