Promoted by the boss of Nvidia, since the end of 2023, the concept of sovereign AI has attracted an increasing number of states. However, it covers a reality not always in line with the concept.
With global spending in the infrastructure of the games to reach $ 100 billion by 2028, according to an IDC study; American pressures to curb the advance of China on AI; And European capitals that take out the big game to attract investments around this technology, AI is increasingly imposed as a major weapon in the competition that the great powers are engaged.
In this context, it is hardly surprising that a new concept is currently on the rise: that of a sovereign AI, or the idea that each power has its own AI system, trained on national data, aligned with the country’s values and turning on local and sovereign infrastructures. Branded by Jensen Huang, the boss of Nvidia, during an event at the end of 2023, the concept has been gaining from popularity since in line, in line with previous technological waves (data sovereignty, sovereign cloud, etc.).
A thousand nuances of sovereignty
States thus multiply the programs in this direction. France cajole its nugget Mistral, who, with Bpifrance, the investor Émirati MGX and Nvidia, undertakes to build the largest European campus dedicated to AI. In the United Kingdom, the University College London (UCL) has teamed up with Nvidia to design a large model of language (LLM, for Large Language Model), Britllm, trained on British data and mastering the different idioms of the Kingdom. English, of course, but also Welsh and Irish. He was trained on Isambard-Ai, a superoder based in Bristol and fruit of the British government’s investment in AI. “Our goal is to get out of the competitive and open-source British LLMS,” says Pontus Stenetorp, Deputy Director of the UCL Artificial Intelligence Center.
Europeans are not the only ones wanting a sovereign AI: the concept is also gaining popularity in the Gulf countries. In May, Saudi Arabia launched human, a company whose objective is to build AI infrastructure in the kingdom. The United Arab Emirates have invested $ 20 billion in the Stargate UAE project, announced with great fanfare during a recent visit by Donald Trump. Are associated with Openai, Nvidia and the local company G42, partially held by the sovereign fund of the Emirates. If the servers will be built using American technologies, the Emirates will have their hands on data governance and model training. The data will not leave the territory and the LLMS will be led to Arabic first.
Even Venezuela recently announced a plan to build infrastructure at the service of a sovereign AI, with the help of Chinese companies.
The promises of sovereign AI
The main argument of the promoters of sovereign AI is naturally the exercise of greater control over a fundamental technological brick, on the way to reshaping a large part of the company. And thus the prospect of reducing external dependencies, especially in the United States and China, at the same time promoting the growth of a local start-up ecosystem around AI, likely to gain international market share. Faced with the prospects for using AI in areas as sensitive as health and the police, exercising better sovereignty over it is also seen as an essential tool to protect the privacy of citizens.
Developing your models at home, on local data centers and using data on which we exercise total control, is also a way to better combat cyberattacks, at a time when Gartner predicts that by 2027, 40% of data leaks will be due to errors in cross -border data transfers at the service of generative AI.
For others, it is a question of causing an AI which reflects the language and the local culture, in the face of the fear that models mainly trained in English and in an Anglo-Saxon context would hardly be adapted to regional particularisms. This is one of the main arguments put forward by India to promote its Bharatgpt. But also by Spain for its own LLM Alia.
Finally, a sovereign AI can be put at the service of research and government for uses that go beyond the logic of profit. Gefion, a Danish national supermarket launched in 2024, is thus used for applications put at the service of the common good, such as the discovery of new drugs and a better prediction of climate events.
On the business side, sovereign AI is a juicy business. For Nvidia, recently become the first company to exceed $ 4,000 billion in valuation, sovereign AI programs constitute as many opportunities to sell its graphics cards for local data centers. The American investment bank Jefferies estimates that sovereign initiatives could generate $ 200 billion in additional income for Nvidia. The United Arab Emirates alone wish to import half a million Nvidia flea each year, serving their sovereign AI program.
The limits of sovereign AI
However, the concept can be misleading. Indeed, from the moment when it is impossible to cause large language models without the chips of Nvidia (unless you use those of AMD, also American), all these solutions are not truly sovereign. Especially since the servers housing these chips are also mainly built by two American companies, Dell and Supermicro.
If we are not talking about total sovereignty, these projects may however have the merit of better guaranteeing confidentiality and security of data, in particular against American extraterritorial laws such as the Cloud Act or the FISA 702. However, the major players of the Cloud also offer to host data on servers located in Europe, with the possibility of quantifying the data on the customer, the key remaining in Europe and outside the control of the American suppliers. But these solutions are not really a bulwark at the American Cloud Act.
In addition, “a European AI laboratory can host its models on a data center located in France, but it works thanks to American hardware, software and intergiche. The illusion of control masks an interdependence canvas,” says Nathan Benaich, founder of the Air Street Capital Investment Fund, specialized in AI. For him, sovereign AI constitutes “political marketing and not a technical reality. It inscribes AI in a rhetoric of national autonomy, even though the underlying systems remain manufactured abroad and taken in globalized logics.”
To find sovereignty over AI, the only solution would therefore be to have more European champions likely to serve negotiation levers in the face of American power, on the Dutch ASML model, without whom TSMC would be unable to melt the chips of Nvidia.
The Ardian Semiconductors investment fund was launched in 2024 to respond to this ambition, by supporting European actors in the semiconductor. “Our goal is to seek companies in Europe that are exposed to the most interesting megatencies, such as data centers for AI and the smart car. And thus to promote the emergence of industrial champions, which allow us not to become self-sufficient on semiconductors, but to make our partners in the world cannot do without Europe”, explains Thomas Pebay-Peyroula, Co-founder of Silian Partners.
