Self-Sovereign AI

Most programs that write texts or generate images today do not run on one’s own device. They run in the data centers of large companies, mostly in the USA. Anyone who uses them sends their data there and gets a response back. Self-Sovereign AI describes the opposite approach: you operate such systems yourself and retain control over them. This refers to a bundle of conditions — one’s own computers or at least one’s own contracts, disclosed program components, and data that never leaves one’s own house. The term is used both for individual companies and for entire states.

Dependence on a Handful of Providers

Currently, almost all AI usage depends on a handful of corporations. They own the strongest models, the largest data centers, and the specialized chips for them. Anyone who depends on this faces three problems. The price can rise, the offering can be discontinued, and the provider’s rules also apply to you.

For companies, data protection is an additional concern. A law firm or a hospital may not simply send sensitive documents to an external service. Industrial companies are also hesitant, because design data is among their most valuable assets. A self-operated model solves this problem because the data stays within one’s own network.

At the state level, more is at stake. Governments in Europe, India, and Japan argue that AI is becoming critical infrastructure — similar to power grids or mobile networks. Sourcing such infrastructure entirely from abroad is considered a risk by many. As a result, billions are now flowing into national data centers and models.

The Four Building Blocks of Self-Control

Sovereignty does not arise from a single step but on four levels. The first is hardware: computers with suitable graphics cards that one owns or that are located in a data center in one’s own country. The second is the model itself. Here, open models such as Llama, Mistral, or Qwen help, whose blueprint and weights one is allowed to download.

The third level is the data. Training data and user inputs remain within one’s own environment and are not used to improve foreign systems. The fourth level is the surrounding software: tools for operating, monitoring, and adapting the model. If any one of these levels is missing, sovereignty is only partially achieved.

In practice, hardly anyone trains a large model from scratch. That costs tens of millions. Instead, fine-tuning is the norm: you take a finished open model and retrain it with your own documents. This is orders of magnitude cheaper and sufficient for most applications. The price of sovereignty then is one’s own staff and one’s own responsibility for security and up-to-dateness.

From National Data Centers to the Laptop

In the news, the term usually appears in the variant Sovereign AI. This refers to state-funded data centers and models, for instance in France, Saudi Arabia, or India. Chipmaker Nvidia actively promotes this idea, because every national project needs thousands of its chips. Critics call this partly marketing, since the chips themselves still come from just a few factories.

In everyday business, one encounters the concept as a so-called on-premise solution. This simply means that the software runs on servers in one’s own building rather than in the cloud. Banks and government agencies frequently mandate this. Providers such as Aleph Alpha or IBM specifically advertise this mode of operation.

The smallest form can be tried out on your own. With programs like Ollama or LM Studio, you can run an open language model on an ordinary laptop. It responds more slowly and less intelligently than the large services, but without an internet connection. A common misconception is to equate open models with sovereignty. As long as you use them via their manufacturer’s servers, you have gained nothing — what matters is where they run.