Sovereign AI and base models

You hear more and more often that states, companies and other organisations are building "sovereign AI models" or training "their own frontier models". The phrasing sounds as though it named some new category of technology. It is two entirely different things that have ended up in the same sentence.

A base model, or foundation model, says what kind of model it is. Sovereignty says on whose terms it is used. The first is a technical description; the second is a question of ownership and power. A model can be a base model without being sovereign, and a sovereignly controlled model can be any small specialised model.

So it is worth listening to the two words separately. Then you can hear what is actually being promised.

A base model is a foundation, not a summit

Once, an AI model was built for one task: one model spotted spam, another read handwriting, a third forecast consumption. Each started from nothing and each needed its own training material.

A base model turns the order round. First comes one extremely expensive pre-training run on an enormous body of general material, and only afterwards is it decided what the model will be used for. The idea in the name is the foundation of a building: the same foundation carries several different houses.

You recognise a base model by three properties:

• It was trained on general material, not for one task. During training nobody knew what it would end up doing.
• It is adaptable. The purpose is supplied afterwards through prompting, retrieval or fine-tuning.
• Its cost is shared. Pre-training is a large one-off outlay that is worth it only if the same model serves many uses.

Two misunderstandings are worth correcting straight away.

A base model does not mean a language model. Language is the best-known kind of base model, not a synonym for one. The same principle trains models for images, audio, video, protein structures, weather data, sensor measurements and controlling equipment. When somebody says text models are not enough for some purpose, that is not a rejection of base models but a choice of a different kind of material.

A base model is not the same as a frontier model. Frontier means the newest and most capable generation. It is a ranking, not an architecture, and it changes annually. This year's frontier model is next year's ordinary mid-range, even though nothing about it changed.

What sovereignty means

Sovereignty in AI means the ability to use a model, develop it and, if necessary, keep using it without somebody else being able to stop you.

The word states a goal, not an implementation. Put bluntly, sovereignty is not one thing but a list of dependencies, and each of them is separately either in your own hands or not:

• Weights. Do you hold the model's numbers, or only a right to use somebody else's interface?
• Where it runs. Where does the computation physically happen, and whose law reaches there?
• Data. Whose material was the model trained on, and do you have the right to use it next year too?
• Skills. Does somebody know how to fix the model when it stops working, or is the person who does on another company's payroll?
• Hardware. Who makes the accelerators, and who decides to whom they may be sold?
• What survives a cut. Which part of the operation continues if the contract is terminated or exports are banned tomorrow?

That last question is in practice the one all the others merely serve. Sovereignty is measured by what is left when the connection is cut.

Sovereignty is a staircase, not a switch

There are not two options but five, and each step buys more control at some price.

1. An interface from a foreign supplier. The quickest and cheapest way to start. The questions go to another company's servers, and the use ends the day access does.
2. The same model in your own or a domestic data centre. The data does not leave the country and the law is familiar. The model is still somebody else's and you cannot reach inside it.
3. A model published with open weights, run yourself. The model stays usable even if the publisher disappears. The training material and the choices made about it are still unknown.
4. Your own fine-tuning on top of an open model. Behaviour, vocabulary and the peculiarities of a field become yours at a reasonable cost. The foundation is still borrowed.
5. Your own pre-training from scratch. Full control over what the model has seen. It costs more in compute, in material and in people than most organisations expect, and it is not a one-off.

Most talk of "sovereign AI" means some step between two and four. That is not deceit; it is usually the most sensible available choice. The confusion arises because the listener understands step five.

When training your own model is justified

There are a few reasons, and they are concrete:

• The material cannot be sent out. Classified information, patient records or a trade secret does not travel through another company's server, however good the model is.
• The task is not in general material. Rare languages, a narrow specialism, or sensor data nobody has ever published.
• Operational reliability is part of the requirement. The system has to work when connections, contracts or trade policy do not.
• A suitable model simply does not exist. A general model can hold a conversation, but it will not steer a machine in an environment no material has ever described.

One reason is a bad one: "ours is better." Sovereignty is not a measure of quality but a measure of control, and control is usually paid for in quality, speed or both. It can still be the right trade. It is just worth knowing you are making it.

Specialised models also hit the same wall: there is not enough material, because the situations have not been, or could not be, recorded. The usual answer is to generate material by simulation and with synthetic data. Then the old rule applies. The model learns the world the simulation describes. If the simulation is wrong, the model is confidently wrong, and the error is discovered only in real conditions.

Questions that check the claim

Next time somebody says they are building sovereign AI, these six questions establish what is going on:

1. Which part is your own: the weights, the fine-tuning, the application, or only a right of use?
2. What material was the model trained on, and who owns it?
3. Where and on whose hardware does it run?
4. What stops working if the foreign supplier shuts down tomorrow?
5. Who updates the model next year, and where does the money for that come from?
6. How is quality measured, and is it measured in real conditions or in a demo?

The last is the most important. A model that works in a demo is not the same as a model that works when the conditions differ from the training material. This is machine learning's oldest problem, and it does not go away because the model is domestic.

What to take away

In one sentence

A base model says the same model was built as a foundation for many purposes. Sovereign says somebody wants to decide its fate themselves. Neither word says whether the model works.