Sovereign AI

Computer programs that write texts, generate images, or understand speech do not run on ordinary laptops. They need large halls full of specialized computers, so-called data centers. Most of these halls are located in the USA and belong to a handful of corporations. Sovereign AI is the political demand that a country own and operate such technology itself. This includes three things: its own computers, its own data, and its own technical expertise. The term has appeared in almost every government speech on digital policy since around 2023.

Dependence on a few providers

Whoever does not own a technology themselves is dependent on the conditions set by others. A provider can raise prices, shut down features, or exclude entire countries. With oil and gas, Europe experienced what such dependence means in a crisis. With AI, the concern is similar, except that here it concerns administration, medicine, and the military.

A second point is data. When a German government agency uses an American AI program, its queries often travel to servers abroad. Different laws apply there, for example regarding access by intelligence agencies. For tax data, patient records, or police records, this is a real problem.

Third, it is about language and culture. Models trained predominantly on English texts have poorer knowledge of German legal terms. They also pass on American perspectives without anyone having decided that this should happen. Domestic models are meant to counteract this.

The three building blocks: chips, data, expertise

The first building block is computing power. This requires graphics chips, or GPUs for short, i.e. special processors that can carry out very many calculations in parallel. Almost all of them come from the US company Nvidia, manufactured in Taiwan. A country can buy these chips and install them in its own data centers. It would only be truly independent if it could develop and manufacture them itself.

The second building block is data and models. A language model learns how language works from huge amounts of text. Anyone wanting their own models must collect German-language texts, prepare them, and clarify the rights to them. Some countries train completely from scratch, others take a freely available model and adapt it. Freely available means: the finished building blocks of the model can be downloaded publicly.

The third building block is people and rules. Experts are needed to operate such systems, and laws are needed to regulate how they are used. An important distinction: sovereign AI does not mean isolation. The goal is usually freedom of choice, that is, the ability to switch to one’s own solution in an emergency.

Billion-dollar programs and cloud offerings in Europe

In the news, the term appears mainly in connection with large funding programs. The EU is building supercomputers through the EuroHPC program, such as the Jupiter computer at the Jülich research center. France, India, Japan, and Saudi Arabia have announced their own multi-billion-dollar projects. Nvidia CEO Jensen Huang has openly promoted the slogan Sovereign AI for years, because his company sells the chips for it.

The term also appears on product packaging. Cloud providers advertise data centers that are located exclusively in Germany and staffed by German personnel. Cloud here simply means: the computers are located at the provider’s site, and you rent them via the internet. Whether the offering is truly independent depends on who owns the company and whose software runs on it.

A common misconception is that sovereign AI is automatically better or safer. A domestic model can make just as many mistakes as any other. Sovereignty says something about control, not about quality. For investors, the term is nevertheless relevant because it channels government money into data centers, chips, and energy supply.