THE MISTRAL PARADOX: 3 ACTS TO RETHINK SOVEREIGNTY IN THE AGE OF AI

By Dr. Tawhid CHTIOUI, Founding Presidentof aivancity School of AI & Data for Business & Society; selected by Keyrus as one of the 25 most influential global figures in the field of AI and data (January 2025).

---

ACT 2: France knows how to train talent. Why does it struggle to build power?

Introduction: the French paradox

France’s relationship with artificial intelligence is marked by a paradox it can no longer afford to accept for long. It has renowned researchers, demanding scientific schools, high-performing laboratories and entrepreneurs capable of building ambitious tech companies. Yet when it comes to turning that excellence into global industrial power, the picture changes abruptly.

The numbers show the scale of the gap. In 2025, American companies attracted nearly $286 billion in private AI investment, compared with $4.36 billion for France. In generative AI alone, the United States accounted for $163.6 billion in private investment, while China and Europe combined totaled $4.7 billion. The OECD reaches the same conclusion using a different method: American companies captured roughly 75% of the global value of venture capital invested in AI in 2025, compared with 6% for the European Union as a whole (Stanford HAI, 2026; OECD, 2026).

These gaps do not mean that Europe has stopped innovating. They say something more worrying: in today’s AI, scientific quality and technological power have become two different things.

Mario Draghi framed the problem particularly clearly in his report on European competitiveness: Europe lacks neither ideas, nor researchers, nor entrepreneurs; it fails far more often to turn innovation into global commercial success, notably because its companies struggle to secure funding and scale up in a market that is still fragmented (Draghi, 2024). Two years later, the European Commission itself acknowledges that the continent still lags behind in computing capacity, semiconductors and certain infrastructure essential to AI development (Commission européenne, 2026).

This distinction is essential. For a long time, we thought of scientific power in terms of universities, laboratories, publications and talent. Artificial intelligence forces us to look at the entire chain. Turning a discovery into power now requires the ability to rapidly mobilize billions in capital, access tens of thousands of advanced processors, secure the energy needed to run them, attract and retain the best researchers, have a market large enough to grow in, and be willing to fund several years of uncertainty before a technology becomes profitable. The OECD also shows that computing and hosting infrastructure has become, since 2023, the leading destination for AI venture capital, with $109 billion invested in 2025 alone (OECD, 2026).

In other words, the new technological frontier no longer rewards only those with the best ideas. It favors those who have the ecosystem capable of scaling them.

This may be where the real fragility of France and Europe lies today. For a long time, we asked our educational and scientific system to produce knowledge and talent. We must now ask what happens to that intelligence once it has been produced: who funds it, who gives it the means to compute, who industrializes it, who opens a market to it and, ultimately, in which country it creates value.

A nation can train excellent researchers while underfunding their ambitions. It can produce discoveries on which others will build companies. It can even become a formidable talent pipeline for the very technological powers it then seeks to compete with.

France’s challenge, then, may not be to produce more intelligence. It is to understand why we find it so hard to turn the intelligence we produce into power.

It is this mechanism that we now need to take apart.

I. The great confusion: talent does not mean sovereignty

For too long, we have confused two things: having talent and having technological power. The two are obviously linked, but they are not the same. A country can train excellent mathematicians, publish leading research and host renowned laboratories without being able to build the companies, infrastructure and market positions that would allow it to control, over the long term, the technologies born of that research. The problem could almost be summed up in a deliberately provocative equation: talent + research ≠ technological sovereignty.

What is missing between the two is an entire transformation chain.

In today’s artificial intelligence, a scientific discovery must be able to meet sufficient capital, immediately available computing power, digital infrastructure, energy, quality data, engineers capable of turning research into a product, entrepreneurs capable of building an organization, and customers large enough to let it scale quickly. That growth must then be financed for several years, new talent must be attracted, access to critical components must be secured, and the company must take on competitors whose resources are sometimes out of all proportion to those of a young startup. Speed itself becomes a strategic resource.

An excellent innovation that takes five years to reach the market can be overtaken by a slightly less advanced technology that is industrialized in eighteen months. At the AI frontier, where model generations now follow one another within months, the time between research and industrialization becomes almost as important as the quality of the research itself.

Data from the Stanford AI Index illustrate this new reality. Global computing capacity dedicated to AI has reportedly grown by an average of 3.3x per year since 2022. Meanwhile, more than 90% of the models considered notable in 2025 were produced by industry, even though universities continue to play a major role in fundamental research and influential publications (Stanford HAI, 2026).

This shift is considerable. It means that the innovation frontier increasingly lies within organizations capable of bringing together science, capital, engineering and infrastructure at the same time. Knowledge remains indispensable, but on its own it is no longer enough to reach the technological frontier.

Yet the literature on national innovation systems identified this difficulty long ago. The work of Lundvall and Nelson showed that innovation never results simply from the excellence of a single laboratory or company. It depends on the interactions between universities, companies, funders, public authorities, markets and institutions capable of turning knowledge into economic activity (Lundvall, 1992; Nelson, 1993). AI is now pushing this logic to an extreme, because the cost and complexity of certain infrastructure make this interdependence far more visible.

The European Commission itself seems to have gradually embraced this change in perspective. Its plan for an “AI continent” does not rest solely on training researchers. It explicitly combines computing power, access to data, financing, AI adoption by businesses and skills development. In particular, the EU plans up to five AI gigafactories, nineteen AI Factories and a target of mobilizing €200 billion in investment. Implicitly, this program acknowledges a shortfall: the Commission itself describes Europe’s lack of large-scale computing infrastructure as a bottleneck for the continent’s competitiveness and strategic autonomy (Commission européenne, 2025-2026).

To this chain must be added a player whose role is sometimes underestimated in Europe: the customer.

Governments can help create a market for their own technologies through their purchasing, industrial policies or major programs. Private companies can do the same when an economic ecosystem is large enough to offer new players immediate outlets. A technology does not become strategic simply because it is excellent; it becomes strategic when it finds users quickly enough to fund its improvement, accumulate data, attract investors and gradually impose its standards.

Public procurement, in particular, is therefore not just a budgetary instrument. It can become an instrument of sovereignty when it allows emerging technologies to find their first major domestic markets. This is where the comparison with the United States and China becomes illuminating. Their advantage does not simply lie in having more researchers or entrepreneurs. They have developed, following very different logics, systems capable of connecting research, financing, infrastructure, industry and markets.

Europe has long thought of innovation as a matter of knowledge. The major technological powers also treat it as a transformation chain.

Sovereignty, then, does not mean owning each link separately. It requires that capital can meet talent, that talent can access compute, that compute can become a product, that this product can find a market, and that the company that created it can grow without having to leave its ecosystem to find the means to match its ambition.

In other words, the real question is no longer just how much talent we train. It is what collective machine we have built to turn their intelligence into power.

This is where the American, Chinese and European trajectories diverge.

II. Three different systems for producing power

The United States, China and Europe do not build their technological power in the same way. Their economic, political and institutional models differ profoundly. Yet the comparison reveals an essential distinction: the first two powers have each succeeded, according to their own logic, in organizing a chain capable of quickly turning research into technologies, and then technologies into markets.

In the United States, this chain works first and foremost as a capitalization machine. Leading universities produce research and attract international talent; venture capital funds highly uncertain projects very early on; Big Tech then provides computing capacity, cloud infrastructure, data, distribution networks and financial firepower that are hard to match. Finally, acquisitions make it possible to quickly fold emerging innovations back into large platforms.

The financial scale of this system has become spectacular. In 2025, US private investment in AI reached $285.9 billion, more than twenty-three times the amount recorded in China, while 1,953 new American AI companies received funding, more than ten times the number in the next country (Stanford HAI, 2026). These figures do not mean that all this money necessarily produces better innovations. They do, however, show the exceptional depth of an ecosystem capable of simultaneously funding thousands of experiments and a few gigantic industrial bets.

China has built a different mechanism, which could be described as an industrialization machine. The state sets strategic priorities and invests in infrastructure, semiconductors, computing power and the spread of AI across the economy. But it would be wrong to imagine a system simply run from the top down. Alongside giants such as Alibaba, Tencent, Huawei, Baidu and ByteDance, particularly intense competition has developed among DeepSeek, Z.ai, Moonshot, MiniMax and many other players.

Recent work by the Stanford Institute for Human-Centered AI thus describes a Chinese ecosystem far more diverse than is often perceived in the West, in which more than a dozen organizations are developing high-performing open or open-weights models. The authors highlight the role of public support but point out that it does not, on its own, explain this momentum: competition between companies, the pursuit of computational efficiency and the drive to make models easier to deploy also play a major role (Meinhardt et al., 2025).

China’s official strategy also reflects this determination to move AI quickly from the lab to the real economy. The “AI+” initiative explicitly calls for its integration into industry, research, services and production chains in order to shorten innovation cycles and accelerate industrial modernization (State Council of the People’s Republic of China, 2025).

Europe presents a third face: that of fragmented excellence.

It has leading researchers, renowned universities, large industrial companies, a market of nearly 450 million consumers and considerable savings. Yet these resources still struggle to flow between countries, financial markets, universities and companies.

The Draghi report sums up this difficulty bluntly: no EU company created in the last fifty years has reached a market capitalization above €100 billion, while the six American companies valued at more than $1 trillion were all created during that period. In addition, nearly 30% of the unicorns founded in Europe between 2008 and 2021 moved their headquarters abroad. Draghi stresses the central cause: Europe produces many ideas but struggles to turn them into global commercial successes, notably because of the fragmentation of the single market and the financing of scale-ups (Draghi, 2024).

This weakness is now widely enough recognized that the EU itself is seeking to correct it through new policies on financing, AI infrastructure and simplification of the internal market. But the diagnosis remains harsh: Europe has many of the resources needed for technological power without yet managing to give them the same continuity as its competitors.

Europe may not be short of building blocks. What it still lacks is a machine capable of quickly assembling them into power.

It is this difference that changes how we must now judge our educational and scientific systems. Training more talent will remain essential, but it will not be enough if the ecosystem in which that talent operates does not then allow it to research, start companies, raise funding, industrialize and grow on the same scale as its American or Chinese competitors.

III. Rethinking our definition of educational excellence

This situation should lead us to question an indicator we are particularly attached to: the amount of talent we train.

We rightly celebrate our engineers, mathematicians, researchers and graduates when they join the world’s best laboratories. But if a growing share of that talent goes on to build American tech champions, if their discoveries are industrialized elsewhere, or if the innovations coming out of our laboratories find the capital and markets they need to grow abroad, then our measure of excellence remains incomplete.

This is obviously not about calling international mobility into question. Science advances because researchers move around, test their work against others’ and join the environments that allow them to go further. A sovereignty policy that sought to keep talent within national borders would end up weakening the very research it claims to protect.

The question is a different one: what remains in the ecosystem that helped train this talent when it succeeds elsewhere? And above all, is that ecosystem attractive enough for them to come back, start businesses, transfer their discoveries or build their companies there?

This is where higher education’s responsibility changes. Training excellent scientists remains its primary mission, but it must also learn to better connect research, entrepreneurship, industry and society. Creating spin-offs, partnering with companies, technology transfer, joint patents, mobility between the lab and industry, and researchers’ ability to become entrepreneurs are no longer peripheral to the academic mission. They are now part of the path that leads from knowledge to impact.

The European Commission itself seems to recognize this change in perspective. In 2026, it proposed a new framework for measuring “knowledge valorisation” that no longer stops at publications or patents. The proposed indicators now include co-patents between universities and companies, venture capital invested in spin-offs and the number of startups founded by PhD holders. The stated aim is precisely to better measure research’s capacity to generate economic and social value and to contribute to Europe’s strategic autonomy (Commission européenne, 2026).

Academic work on technology transfer also shows that this transformation cannot simply be decreed. Wolszczak-Derlacz’s (2025) analysis of more than 400 European universities highlights in particular the importance of university–industry relationships in the ability to produce and share intellectual property. In other words, scientific excellence has greater impact when it is connected to an environment capable of extending it beyond the lab.

This leads us to broaden our definition of university performance. A grande école or a university should no longer only be able to say how many graduates it trains, how many of its researchers publish or how many of its alumni work at Google, Meta or OpenAI. It should also be able to show how many companies were born in its laboratories, how many technologies have reached the market, how many researchers move between academia and industry, how many graduates become entrepreneurs, and what scientific, economic or social value this intelligence produces in its environment.

Educational sovereignty is therefore measured not only by the amount of talent we train, but by our collective ability to turn that talent into scientific, economic and social power.

This shift in perspective is essential. It is no longer about pitting universities against companies, fundamental research against entrepreneurship, or international mobility against the national interest. On the contrary, it is about building more bridges between these worlds. The real failure would not be that our best researchers one day work at Stanford, at OpenAI or in a Chinese laboratory. It would be that they never find in France or Europe the conditions that make them want, one day, to build something just as ambitious there.

IV. The China shock reveals a European mistake

For a long time, Europe thought about its technological lag almost exclusively in comparison with the United States. The equation seemed simple: on one side, American platforms, their capital, their cloud and their proprietary models; on the other, a Europe trying to preserve its autonomy without the same resources.

China’s rise is changing this geography.

China is not simply pitting new champions against OpenAI, Google or Anthropic. Part of its ecosystem offers a different combination: models now close to the technological frontier, often lower costs and, above all, a far more proactive policy of releasing open-weight models that can be downloaded, adapted and run outside their developers’ platforms.

Stanford HAI thus considers that Chinese open-weight models have now caught up with, and in some uses even surpassed, several of their international competitors, and stresses that their global spread could reshape relationships of technological dependence between states (Meinhardt et al., 2025). Qwen is already widely used by developers outside China, while DeepSeek has shown how quickly a Chinese model can spread worldwide.

This dynamic is particularly important for countries that have neither the capital nor the infrastructure needed to train their own frontier models. The Carnegie Endowment notes that another competition is emerging, particularly in the Global South: it is no longer just about who will build the most powerful model, but about who will offer the most accessible, adaptable and easy-to-deploy technologies. After its launch, DeepSeek became one of the most downloaded apps in more than 140 markets, including Brazil and India (Sheehan et al., 2025).

The reasoning of a government, a university or a company can then change profoundly. Why depend exclusively on a closed American API, whose provider retains control over the model, its evolution and often its terms of use, if a sufficiently capable Chinese model can be installed on locally chosen infrastructure, adapted to the country’s needs and, within certain limits, controlled by whoever operates it?

This obviously does not mean that the Chinese model has become “sovereign” merely by being installed locally. Its design, its training and certain technological dependencies remain external. Questions of security, licensing, auditing, culture and governance remain, as we have seen. But the nature of the dependency is no longer the same.

This may be one of the most important geopolitical innovations of China’s strategy. China can spread its technological influence without necessarily requiring the world to come and use a platform based in China. Its models can travel independently of their original infrastructure.

This strategy resonates with particularly strong demand in emerging economies: having high-performing tools without entirely giving up control of their data, being able to adapt models to their languages and uses, and avoiding exclusive dependence on the major American platforms. Carnegie also points out that the US-China competition for AI in the Global South is increasingly being played out on the terrain of accessibility, cost, infrastructure and local adaptation, and no longer solely on the absolute performance of models Lu & Winter-Levy, 2025).

Europe may have made an analytical mistake by viewing AI for too long as a race in which it simply had to catch up with the United States. China is gradually transforming the race itself.

It is not only seeking to shift the center of global innovation to Beijing, Hangzhou or Shenzhen. It may also be seeking to make its technologies the building blocks on which other countries will build their own artificial intelligence.

If this strategy prevails, the geopolitics of AI will no longer be just a competition over who has the best model. It will become a competition over whose models will become everyone else’s cognitive infrastructure.

And it is this new landscape that forces us to reexamine Mistral’s strategy with greater nuance.

V. What if Mistral were making a rational choice?

Yet there remains a possibility that the debate over Europe’s “falling behind” tends to overlook: what if Mistral’s evolution was not only forced upon it, but also reflected a clear-eyed reading of how the AI market is transforming?

Shahin Vallée, a researcher at the German Council on Foreign Relations, frames the hypothesis in a deliberately provocative way. In his view, Europe does not currently have the resources needed to win the frontier model race on its own and should prioritize AI adoption over the sole pursuit of frontier innovation. He therefore sees Mistral’s repositioning toward inference, infrastructure and enterprise deployment as a potentially rational strategy, provided that high-performing open models remain available (Vallée, 2026).

The hypothesis is uncomfortable because it can be read in two ways.

The first is pessimistic: Mistral would be acknowledging that it is becoming extremely difficult to keep up, over the long term, with the financial and computational pace set by the American and Chinese giants, and would be gradually shifting its center of gravity toward activities less exposed to this race.

The second is more strategic: the economic battle for AI may not be won solely by owning, at every moment, the number-one-ranked model.

As the performance of the best models converges and high-performing open-weight models become accessible, a growing share of value may shift toward what companies build around them: the infrastructure they run on, the ability to choose and orchestrate multiple models, their adaptation to proprietary data, security, compliance, agents capable of acting within information systems and, above all, their real integration into the organization’s processes.

This is clearly the direction in which Mistral is now broadening its offering. The company no longer presents itself solely as a frontier model lab. It now offers a much broader stack combining models, compute infrastructure, agent-building tools, model customization and alignment, applications and applied AI services. Its official strategy explicitly aims to enable organizations to build custom systems around their own knowledge and to retain control over their infrastructure and their “learning loop” (Mistral AI, 2026a).

This shift could become significant if foundation models follow a path similar to that of other technologies that gradually became components of a larger system. When several engines become sufficiently powerful, competitive advantage no longer necessarily lies in the engine alone, but in the vehicle built around it, in how it is used and in how it integrates into a particular environment.

It would nevertheless be premature to conclude that the model battle no longer matters. Whoever controls the frontier retains a capacity for innovation, negotiation and independence that the user of other people’s models does not fully possess. Vallée himself acknowledges the strategic cost of his scenario: prioritizing adoption means, at least in part, accepting that others will continue to push the technological frontier (Vallée, 2026).

This is precisely why it would be too simplistic to choose right now between two narratives: that of a Mistral in retreat and that of a Mistral that has discovered, ahead of everyone else, where value will lie tomorrow.

Perhaps we are witnessing both movements at once: a relative retreat in the race for raw model power, but an attempt to stake out a position on another frontier, that of their adoption by companies, institutions and states.

If this hypothesis is confirmed, it once again changes the European question. It will no longer just be a matter of whether we can produce the best model in the world, but of determining which layers of the intelligence chain we absolutely must control so as not to become dependent on those who produce it.

Conclusion. A rational strategy for Mistral is not necessarily a sufficient strategy for Europe

One might conclude that Europe should simply accept the new international division of labor: the United States and China would produce the most advanced models, while we would specialize in hosting, securing, orchestrating and integrating them into businesses.

This strategy may be economically rational. In the short term, it may even significantly accelerate AI adoption in a Europe where part of the economy still lags in usage. Shahin Vallée is right to point out that a technology only has impact when it is actually adopted, and that sovereignty is not built in laboratories alone (Vallée, 2026).

But turning this observation into an industrial doctrine would be dangerous.

Because there is a fundamental difference between making smart use of other people’s innovations and gradually losing the ability to produce them yourself.

If Europe abandons the model frontier on the grounds that others know how to build models faster and more cheaply, it risks becoming extraordinarily skilled at integrating technologies whose trajectories it will no longer control. It will be able to choose between several providers as long as several providers exist, benefit from open models as long as they remain open, and build its infrastructure around innovations whose main directions will be decided elsewhere.

Such sovereignty would be paradoxical: we would increasingly master the use of an intelligence whose making we would master less and less.

The problem, then, is not choosing between innovation and adoption. Lasting technological power must be able to do both. Adoption without innovation leads to dependence; innovation without adoption leads to scientific excellence without economic impact. The real challenge lies in the ability to sustain the loop that connects research, industry, uses and new research.

This is where the Mistral case regains its full significance. If its repositioning allows the company to build a powerful European player in infrastructure, orchestration and applied AI while maintaining a credible capacity for model research and development, it could represent a new form of technological power. If, on the other hand, the “model” side gradually becomes secondary because decisive innovations systematically come from elsewhere, we will have to be honest enough to admit that this is no longer quite the sovereignty we thought we were building in 2023.

The debate obviously goes beyond a single company. It concerns the collective choice that France and Europe must make in the face of a historic change of scale.

The Draghi report already stressed the need to close the gap between invention and commercialization, but the AI challenge adds a further difficulty: we must learn to industrialize more without giving up on remaining a place where the next generation of technologies is created (Draghi, 2024).

This means no longer pitting against each other policies that have too often been designed separately: academic excellence, industrial policy, capital, energy, compute infrastructure, public procurement, entrepreneurship and the growth of tech companies must become components of a single strategy.

France can continue to train excellent researchers. Europe can multiply funding programs and infrastructure. Mistral can become a formidable integrator of global models. All of this is useful. But none of it yet fully answers the question of sovereignty.

Because being sovereign means neither producing everything yourself nor simply knowing how to use what others produce. Between autarky, which is impossible, and dependence, which is dangerous, we still need to define what we absolutely must control, what we can share, what we can buy, and what we must never lose the ability to rebuild ourselves.

This may be where the real question begins.

No longer: “Can Europe still win the AI race?”

But rather: which forms of sovereignty do we really need to remain in control of our choices in a world where no intelligence will be entirely national?