The Necessity and Possibility of Sovereign Artifical Intelligence—and a Sovereign Cloud
The Ultimate Battle
On February 10 and 11 2025, Ursula von der Leyen, President of the European Commission, participated in the Artificial Intelligence Action Summit in Paris, France, where she announced a €50 billion investAI initiative and an additional €8 billion to AI Factories during her speech. Paris, February 11, 2025. Source: Dati Bendo / European Union, 2025 / EC - Audiovisual Service
This article, the third in a series of six, is devoted to what is known, mainly in Europe, although it also concerns other countries, as sovereign artificial intelligence.
It was written with Constantin Vaillant Tenzer, a researcher in mathematics applied to cognitive neuroscience and machine learning at Ecole normale supérieure Ulm (Paris Sciences et Lettres University). For more than four years, he has been working on improving AI algorithm training methods in collaboration with French and American companies.
The United States now controls most of the critical layers of digital technology: advanced AI chips, large language models, hyperscale clouds, and extraterritorial legal frameworks (Patriot Act, CLOUD Act). The question is no longer whether Europe should aim for digital sovereignty, but whether it can achieve it in time and with sufficient political consistency. This article argues that ambitious digital sovereignty is technically possible with current tools, provided there is clear, stable, and massive European guidance.
Towards full digital sovereignty: beyond AI, a question of political regime
Much of Europeans’ sensitive data (health, taxation, industry, defense) is now stored or processed on infrastructure controlled by American companies, subject to American law, even when operating from data centers located in Europe. In practical terms, this means that a unilateral decision by Washington can affect access, use, or confidentiality of this data, regardless of the decisions of Member States, to the point of potentially paralyzing the entire European Union.
AI amplifies this problem: not only is the data hosted abroad, but the capabilities for analysis, modeling, and prediction are concentrated in the hands of a small number of non-European companies. In a world where strategic advantage relies on a combination of data, computing resources, and algorithmic mastery (and the structural dependence of their users), not mastering these three elements amounts to accepting a position of structural dependence.
US restrictions on the export of AI chips to China, and the decision to reserve Nvidia’s most advanced Blackwell chips for the US market alone, have shown how digital power can be used as a tool for technological lock-in as well as foreign policy. Even though some decisions have been partially relaxed for intermediate chips (H200), but only for Europe, as Nvidia won’t export anymore to China. The strategic message is clear: the US intends to maintain a lasting lead in aggregate computing power. If Europe continues to depend on US suppliers for its critical infrastructure, it exposes itself to these same levers of pressure, without enjoying the political benefits of technological mastery. We discussed the alternatives and difficulties associated with this manufacturing process in the previous article.
From “The supply chain capitalism of AI: a call to (re)think algorithmic harms and resistance through environmental lens” by Ana Valdivia, Informaiton, Communication and Society, vol. 28. 2025, 12. CC BY 4.0. (https://doi.org/10.1080/1369118X.2024.2420021)
Sovereign and federated cloud
Sovereignty is not limited to hardware: it requires control over the cloud and the legal location of data.
Downstream from GPU production are computing and data centers. Most are located in the United States, and the vast majority of those located in Europe are managed by US companies (Amazon, Microsoft, Google, Oracle). European players such as OVH and Scaleway (Illiad) are struggling to grow, due to the ease of using companies with much more aggressive marketing that also offer much better software integration with their other services. The problem is that, faced with the effective marketing of the American giants, few economic or state players (such as the University Hospital of Montpellier, France or the French social safety net) are concerned about the issues this raises.
Microsoft may be hosted in France, but it obeys US laws.
One way to compete with US players would be to offer partial and flexible allocation of computational resources, which would be a major commercial advantage, entirely possible thanks to MIG (for Nvidia GPUs), corresponding to pay-per-use with dynamic allocation.
The Gaia-X initiative aims precisely to build a European data and cloud infrastructure that is federated, interoperable, and protected from non-European extraterritorial legislation. It proposes certification mechanisms to determine the extent to which a service complies with sovereignty requirements (control of data location, absence of submission to foreign laws, reversibility, etc.). Another similar initiative in France is digital resilience initiative. We recommend that European law include the obligation for health, justice, defense, critical infrastructure, and strategic research data to be stored and processed on certified sovereign clouds (Gaia-X or equivalent), operated by European companies. Digital subsidies and public procurement must be conditional on the use of services certified as “strong sovereignty,” creating structured demand for existing sovereign cloud players (OVH, Scaleway, Arum Technologies, etc.). Finally, legal tools must be provided to enable European states to act quickly, easily, and in a coordinated manner to prohibit the extraterritorial application of foreign laws on European data and to organize a common defense in the event of a dispute.
Furthermore, Bloomberg predicts that only 18% of data centers will be controlled by big tech; this trend toward cloud decentralization should not only concern the United States and China, as is currently the case.
Data sovereignty
In China, all population data serves as a huge laboratory for governance and training for increasingly sophisticated AI tools. It is essential to prevent the Chinese or Americans from doing the same with our data.
The most powerful models derive their advantage from access to massive corpora, often collected without clear public oversight. However, Europe has data of exceptional value: health, climate, industry, transportation, public research, culture. It is necessary to contractually prohibit the use of this strategic data to train non-European proprietary models, except under strict and accountable reciprocity agreements. Above all, we need to enforce these prohibitions.
This risk of exploitation of European data is not just a matter of confidentiality: it is also a major economic issue, particularly visible in the health sector. Europe produces considerable volumes of high-quality medical data (imaging, patient records, genomic data, clinical trials) through its public health systems and research institutions. If this data is used to train proprietary American or Chinese AI models without fair compensation, the economic and therapeutic benefits (early diagnosis, new treatments, optimization of care pathways) will be captured by foreign companies. The European market for AI in healthcare is expected to reach $31.72 billion by 2030, but this value will be largely captured by non-European players if the EU does not secure its data and develop its own models. The European Health Data Space (EHDS) represents an attempt to structure access to the health data of 449 million European citizens in a secure and interoperable manner. But without sovereign AI capabilities to exploit this data locally, Europe risks becoming a mere supplier of informational raw materials for models developed elsewhere, losing out on medical productivity gains, skilled jobs, and associated patents in the process. An expert from EIT Health points out that Europe could become a world leader in health AI if it combines its data via the EHDS with coordinated investments and common standards. This, however, requires a clear political will not to let this value slip away to the American giants.
In addition, many companies use generative models developed by American giants, mainly OpenAI, as the basis for specialized applications in all sectors where textual interaction is possible. This poses serious data security issues, as these models are only weakly protected.
The United States and China are already several steps ahead when it comes to general-purpose language models with several hundred billion parameters, thanks to their privileged access to chips, abundant venture capital, and massive developer ecosystems. Replicating this model identically in Europe would be costly, slow, and likely doomed to remain marginal. Europe’s ambition should instead be to have one or two large sovereign foundation models capable of covering internal strategic uses (administration, defense, health, research) and fully meeting the expectations of the general public, which Mistral unfortunately no longer does, and which would be served in Europe. We need to develop small specialized models that are highly efficient in targeted areas (bio-medicine, industry, defense, robotics, vision, security) for businesses. We are fortunate to know how to do better with less and to be able to stand on the shoulders of the giants who came before us.
Being more efficient, without compromising our economy and the environment, both by developing new hardware entirely from reclycled components and frugal models and training algorithms, is entirely possible in Europe. We have the technologies, the minds and the available capital. It requires a scientific policy and a political will from governments, but also from major companies and institutional investors to make it possible.
Encourage the transition from research to large-scale innovation. Alternatives to LLMs and the example of the American DARPA
It no longer makes sense to talk about fundamental research when discussing LLM or diffusion models, where there have been no fundamental discoveries since 2020 with the PPO (proximal policy optimization) learning process. For the research produced in our academic laboratories, we need to invent things that are radically different from what we are currently doing. That is what we need to fund, not buzzwords. Europeans will undoubtedly make other exceptional discoveries, and we must not miss the boat when it comes to the industrialization and mass commercialization of our future discoveries, in any field whatsoever. The lack of talent is not a problem in Europe, but rather the lack of resources to enable them to practice. In fact, while there are specific research initiatives working on paradigmatically different and much more frugal AI models (e.g., the Miles team at CEREMADE in Paris Dauphine University, Yann Le Cun’s team working on world models), they are not sufficiently coordinated within the academic world, let alone linked to companies of sufficient critical mass. We strongly advocate for a coherent research conversation place between, mathematicians, computer scientists, physicists and computational neuroscientists. It is indeed a question of optimizing for satisfying outputs under the constraints of a low energetic cost and fast answers production, as in the brain and many physical and biological systems. Solving this scientific problem is crucial for research in this field of frugal AI.
As this field is still underdeveloped in the research arena outside France and has not yet been addressed by large companies in the sector, Europe has a role to play here.
The challenge will be to transform the results of research laboratories into innovation, and this is an area where Europe is still lagging behind. In this specific sector of frugal AI, we have a head start. We need to maintain it. But how? Let’s look at the American example for some useful policy ideas.
A first example is that of powerful and autonomous innovation agencies. Created in 1958 in response to the launch of Sputnik, the Defense Advanced Research Projects Agency (DARPA) has been responsible for major innovations (the internet, GPS, lasers, and more recently, autonomous vehicles, etc.). Since then, other civilian agencies, or agencies linked to intelligence services, have emerged in the United States.
An ARPA is characterized by a flexible organizational structure, the identification of technological barriers and market gaps, and the implementation of programs to address them. Project selection and fund management are at the discretion of program managers, who actively direct projects with technological milestones and time constraints. The US DARPA has an annual budget of $3 billion and 100 program managers, who are highly qualified and autonomous in their approach. They play a major role in the agency’s operations. DARPA has been able to adapt its operations to political demands. Program managers maintain strong ties with the scientific community. The organization’s flexibility is based on its independence from the rest of the administration, as it reports directly to the minister (the Secretary of Defense in the case of the US DARPA); a horizontal organization with only three levels of hierarchy (director – six disciplinary offices – program directors), with research operations conducted solely extra muros; the ability to recruit directly from any background with attractive salaries for three- to five-year fixed-term contracts; and, finally, flexibility in research and the allocation of funds. The programs last around three years, with a budget of $30 million per program and around ten winners. ARPA programs enable companies to move forward and bring together communities that rarely communicate with each other. Program directors do not hesitate to take on projects that have been poorly rated by analysts and base their judgment on critical analysis, nor do they hesitate to cancel projects that fail to meet technological milestones in order to increase funding for others.
All of the methods described above are complementary. A project that can be led by an ARPA must include technological innovations that have not been achieved and cannot be achieved without a catalyst. The project must have a concrete and clearly defined economic or social objective. Each agency has its own general objectives. DARPA does neither fundamental nor applied research, but rather something in between, seeking out existing technologies that are under-explored and have high potential, often risky and creating disruptive innovation. The third necessary criterion is that the market is unable to meet the need: no company would produce or finance such research without government assistance, as is the case in the defense sector.
Its antonym is called the free-rider effect, which is often very present in public funding for innovation. This effect has been observed many times in numerous government innovation programs (France 2030 in France, EIC in Europe). The challenges for ARPAs are to measure the long-term impact of their actions and to differentiate between types of funding according to the audience (large companies, universities, start-ups, etc.). It is important to build relationships of trust with researchers, which is a human endeavor, and to bring developed innovations to market. However, the ARPA model is not always the most appropriate for promoting disruptive innovation.
When it comes to government funding specifically for AI, the amounts of public investment and strategies differ radically: the US is focusing on the military and chips, China on technological independence, and Europe on coordination and shared infrastructure. The R&D budget requested for the Department of Defense (DoD) for fiscal year 2026 is $112.9 billion (out of a total federal budget of $181.413 billion for research and development), which includes a massive share dedicated to military AI and autonomous systems (a 23% increase over 2025). This is the main channel of public funding for AI in the US. In addition, the CHIPS and Science Act (still being rolled out in 2026) provides $50 billion to subsidize chip manufacturing plants ($39 billion) and R&D (11 billion), indirectly but massively supporting AI infrastructure. This amount seems modest compared to the private sector, but it targets high-risk fundamental research. This budget line is for non-defense AI R&D, the reason whileit is mostly founded by the NSF and DOE.
China has raised $47.5 billion over five years through the “Big Fund III” (National Integrated Circuit Industry Investment Fund) to counter US sanctions on semiconductors. Although it is difficult to find reliable sources, Chinese investment can be estimated at $125 billion, including $50 billion in public money.
In Europe, the Paris Summit in February 2025 launched the Invest AI initiative, which is expected to come into effect at the end of 2027, to invest €200 billion, including €20 billion for “AI gigafactories” i.e. sovereign supercomputers accessible to European startups in order to reduce dependence on the American cloud. Of this €200 billion, €150 billion is expected to come from private companies that have committed to the initiative. In addition, the annual budget dedicated to digital innovation (the vast majority of which is earmarked for AI and chips) is $1.6 billion per year. However, all these initiatives combined are about one-third of US public spending in this area. Furthermore, there is no guarantee that both public and private funding will be secured.
An investment policy that measures up: patient capital and public procurement
Most of the big successes in American tech accumulated losses for a decade before reaching profitability, supported by massive and patient venture capital. In Europe, funds are smaller, more risk-averse, and exits are more modest, which automatically limits the ambition of projects. Amazon, for example, was not profitable for 10 years before becoming the giant it is today.
Beyond financial and scientific support for disruptive innovation, the most important issue remains investment. While US GDP in 2025 is 1.45 times higher than that of Europe, there is a ratio of 1 to 15 in venture capital investment between the United States and Europe: in 2025, American startups raised more than $340 billion, compared to less than $23 billion invested in Europe (of which 35% came from the United States and 7% from Asia). This poses a serious challenge to sovereignty: not enough funds to develop innovation in Europe and key technologies almost half controlled by non-European players.
Beyond the investment deficit, Europe is facing a brain drain in AI, which directly threatens its ability to build digital sovereignty. A January 2026 Euronews survey reveals that Europe continues to lose more tech talent than it attracts, mainly to the United States. A 2024 report by the organization Interface shows that European countries are “losing significant AI talent, both domestic and international, to the United States”: Germany, France, and even Switzerland are experiencing net outflows of experienced professionals to American and British giants. Paradoxically, a US report from May 2025 notes that even the United States is beginning to lose its historical attractiveness (balance between arrivals and departures), particularly due to federal budget cuts in research (NSF, NIH) and a pivot towards sovereign AI in other countries. However, these cuts mainly affect environmental science research and have little impact on research with high economic application potential, such as machine learning research and therefore the resulting attractiveness. But Europe still does not attract enough of these talented individuals: the lack of massive funding, high-profile industrial projects, and competitive salaries is driving the best European researchers and engineers to leave the continent. For similar engenering positions located in Paris, Mistral AI pays €100,000 less a year than OpenAI. Without a dynamic ecosystem combining cutting-edge research, ambitious startups, and patient investment, Europe is training talents that then enrich American or Chinese AI and scientific capabilities. This paradox: investing in education and then losing the return on investment, directly weakens European technological sovereignty. It is therefore crucial to create attractive conditions (venture capital, large-scale projects such as DARPA, competitive salaries, international visibility) to retain and attract the best global talent in AI.
One or more pan-European sovereign venture capital funds of truly significant size (several tens of billions of euros) must be created. They could be co-financed by the European Investment Bank (EIB), Member States, and private investors, with an explicit mandate: to finance losses for 8-10 years in sovereign digital and AI. This could be extended to other areas of technology. European savings are under-invested compared to US savings. Prudential rules should be adapted to allow a limited but significant fraction of European savings (pension funds, insurance) to be invested in these funds, with partial public guarantees.
European public procurement represents enormous purchasing power, which is still largely focused on non-European solutions for reasons of cost and ease of integration. Using this power as an industrial tool is essential for creating champions that do not depend exclusively on venture capital. A policy of economic sovereignty and group purchasing must be considered. Europe is large enough to offer healthy internal competition.
American and Chinese state investments are not only a matter of economic domination, but also a way of seeking to impose their standards, as they have already done for many technologies (GPS, the internet, cars, weapons, etc.) in order to gain strategic domination and control. The Americans are way ahead, and as we shall see, the issue at stake is not just control of one sector of the economy among others, but something much more significant.
Making digital sovereignty compatible with innovation
A real risk would be to build digital sovereignty at the cost of excessive bureaucratization, which would stifle innovation, particularly for SMEs and start-ups. The AI Act illustrates this risk: necessary in substance, but potentially costly and complex for small players if not accompanied by support measures.
Legal uncertainty over liability in the event of an AI system failure is a major obstacle to its deployment by European players, creating a competitive disadvantage vis-à-vis less scrupulous competitors. The European Commission’s announcement that it will withdraw the AI Liability Directive in 2025 leaves a major legal vacuum. A July 2025 study by the European Parliament highlights that the lack of harmonized rules creates fragmentation between Member States, with differing interpretations of negligence, variable standards of proof, and legal uncertainty that weighs heavily on both victims and businesses. In concrete terms, a European SME wishing to deploy an AI system in healthcare, transportation, or industry is hesitant: who will be held liable in the event of a serious error in the model? The user company? The model developer? The data provider? The cloud provider? This rational legal fear is prompting many European administrations and companies to simply give up on experimenting with AI, or to outsource to American giants who assume (partially) assume these risks through their terms and conditions. Meanwhile, American and Chinese players, operating in more permissive or clearer legal frameworks, are deploying AI on a massive scale without these constraints. This regulatory asymmetry is becoming a major strategic handicap for European democracies, which must clarify liability regimes without creating crippling bureaucracy.
The objective of AI legislation must be twofold: to protect citizens and democracies from the extreme risks of AI, while allowing European innovators to experiment quickly and, above all, not creating excessive bureaucracy that would paralyze small structures to the benefit of large ones.
Ambitious European digital sovereignty does not mean breaking with the United States or China, but negotiating from a position of relative strength rather than dependence. Europe can cooperate on standards, fundamental research, and certain common uses (climate, global health), while protecting its strategic assets. A competitive but cooperative digital world, where Europe has an autonomous technological base, is preferable to a simple technological “parochialization” of the Union.
European digital sovereignty is not an unattainable “techno-nationalist” dream, but a project compatible with the current state of technology and the economy, if we accept that digital technology is a matter of essential sovereignty, including and especially when it is made available to the general public, and if we commit to a long-term vision (10-20 years) to protect public and private capital, while recognizing the urgency of the situation. In view of the risks, this sovereignty must be achieved by the end of 2026, which is possible provided that all stakeholders are mobilized. Nevertheless, it must be assumed that, in certain segments (large generalist models), Europe will arrive late, but that it can win in others (sovereign cloud, efficient processors, small specialized models, robotics, vision), provided that it concentrates its resources rather than dispersing them.
It is not just a question of “catching up” with the United States or China, but of defining a European digital path: open, democratic, but capable of defending its autonomy in a world where power is now also measured in teraflops, data centers, and the quality of AI models. As for regulation, it is important, but it must be designed by people who understand AI and not be a brake on innovation through heavy-handed bureaucracy, but rather provide effective protection, which is not the case with the AI Act.
We Europeans have lost the game on large LLM models, but we can win it on small, highly specialized and lightweight models, including in the field of defense. These are models that require little infrastructure, but are adaptable and highly accurate for specific tasks, including as intelligent espionage micro-systems, or even tools inspired by microbiology (see the work of the bio-electronics team at the University of Bordeaux) for the creation of intelligent micro-systems. The medical imaging sector still has major European players. In the field of robotics, the Chinese clearly have a head start in robot manufacturing. Nevertheless, INRIA’s Willow team is behind most of the software currently used in robotics. Intelligent drones are in full development, thanks to our Ukrainian allies. We should also take full advantage of this technology.