What is Sovereign AI?
Sovereign AI is the ability to run and govern AI systems on your own terms: knowing what you depend on, and having real alternatives when those dependencies change. It operates at two levels. For a state, it covers national compute programmes, regulation such as the EU AI Act, and export controls on advanced chips. For an organisation, it covers open-weight models, self-hosting and on-prem inference. The two are linked: policy shapes what is available to run, and engineering decides what you can actually run yourself.
The term is used loosely, and often to sell something. It is worth separating what it describes from what gets attached to it.
Sovereignty here is not a state you achieve. It is a set of dependencies you can name, and a set of options you keep open. Most organisations sit somewhere in the middle of the scale and move along it deliberately, rather than all at once.
The two axes
Two questions do most of the work. They are independent of each other, and confusing them is where most of the muddle starts.
Who governs the model
Who decides what the model is, how it behaves, and whether you can carry on using it? This covers the licence on the weights, whether the weights are published at all, which jurisdiction the developer answers to, and what happens to your access if the terms, the pricing or the politics change. A model you can download and keep is governed differently from one you reach through an API, wherever the servers happen to sit.
Who runs the inference
Who operates the hardware the model runs on, and under whose law? The options run from a hosted API, through a sovereign or regional cloud, to inference on machines you own outright. This is an engineering and cost question before it is a political one: throughput, latency, GPU supply, and the staff to keep it all running.
What Sovereign AI is not
Not the same as data residency
Storing data in a particular country is a narrower claim than it sounds. Residency says where bytes are held. It says nothing about who governs the model, who can withdraw your access, or which jurisdiction can compel disclosure from the operator. A model served from a datacentre in France, by a company incorporated elsewhere, under a licence that can be revised, satisfies a residency requirement without satisfying much else.
Not the same as using open weights
Open weights are necessary for some kinds of sovereignty and sufficient for none. Downloading weights you cannot afford to serve, cannot evaluate against your own tasks, and cannot keep current leaves you dependent in a different direction. The weights are where the work starts rather than where it finishes.
Why it is contested now
Three developments moved this from conference talk to procurement documents.
- Regulation arrived. The EU AI Act phases in obligations for general-purpose models and for high-risk uses, which pushes organisations to state plainly what they are running and why.
- Compute became industrial policy. National and EU-level programmes now fund AI infrastructure directly, on the reasoning that a country which cannot train or serve models depends on the ones that can.
- Export controls made hardware political. Restrictions on advanced chips established that the supply of compute can be turned into leverage.
None of this settles the question. It does mean the answer now has a budget line attached to it.
Questions to ask about your own stack
Sovereignty is easier to reason about as a set of specific questions than as a position to hold.
- If your main model provider changed its terms or its pricing tomorrow, what would you do, and how long would it take?
- Which of your workloads genuinely need a frontier model, and which are already served well by a smaller open-weight one?
- Could you evaluate a replacement model against your own tasks, or would you be taking someone else’s benchmark on trust?
- Where does your data go during inference, who can compel its disclosure, and does your contract say so in writing?
- If you moved inference in-house with Ollama or vLLM, what would the hardware cost, and who would run it?
Most teams find the answers are uneven. Some parts of a stack move easily and some do not, and knowing which is which is most of the value.