Model Sovereignty

Nearly all AI programs known today that write texts or generate images come from a handful of companies in the US and China. Anyone who uses them usually rents them over the internet: the request goes to a foreign computer, the answer comes back. The provider can raise prices, shut down features, or block access entirely. Model sovereignty describes the opposite of this dependency. It refers to the ability to build AI systems oneself, operate them oneself, and decide on their rules oneself. The term is used for states, but also for individual corporations, hospitals, or administrations.

Who holds the switch

An administration that uses AI to pre-sort tax assessments relies on this technology. If access disappears, work grinds to a halt. With electricity or water, no one would accept that a single foreign provider could cut it off at any time. With AI, that is often exactly the case. That is why governments now treat the issue as a matter of security, not just economics.

A second point concerns data. Anyone who sends a request to a foreign service hands over content. With medical records, engineering plans, or investigation files, that is a real problem. European data protection rules often require that such information not leave the premises. A self-operated model solves this problem because the data stays in one’s own data center.

There is also a point that is easily overlooked: models have preferences. They are trained on texts that are predominantly English and reflect a particular view of the world. A system that has barely seen German legal texts argues worse on legal questions. Which topics a model refuses to address is also decided by the manufacturer. Whoever is sovereign can make these decisions themselves.

The four levels of control

Sovereignty is not a switch but a series of stages. The first is access to the model itself. With open models, one may download the trained numerical values and run them on one’s own computers. These numbers are called parameters and store everything the model has learned. With closed models, one only gets an interface — a kind of switch — and never sees the inside.

The second level is hardware. Models need special graphics chips, and almost all of them come from a single US manufacturer. A data center in Europe helps against political access, but does not make one independent of this supplier. The third level is training data. Anyone who only fine-tunes foreign models has never controlled what originally went into them.

The fourth level is human knowledge. Operating a model is considerably easier than training one from scratch. Full sovereignty on all four levels is expensive and currently unrealistic for almost any country. In practice, most choose a middle path: open models, operated in their own country, adapted with their own data.

From Gaia-X to your own chatbot

In the news, the term comes up in connection with state funding programs. France supports the provider Mistral, India and the Gulf states finance their own data centers. In Germany, Gaia-X was an early attempt to build European cloud infrastructure. The EU’s AI Act, the European AI law, is also often mentioned in this context.

For companies, the topic is very practical. Banks and automakers increasingly operate their own models on their own servers, even though rented services would be more convenient. The reason is rarely patriotism but rather contract law and cost predictability. Anyone who does not determine the price per request themselves can hardly plan their product reliably.

A common misconception is that a data center in Frankfurt already means sovereignty. If the server belongs to a US corporation, American laws may still apply to access under certain circumstances. Conversely, the term is sometimes used as a marketing label without real control behind it. It is therefore worth asking, with every announcement: who actually holds which switch here?