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 3: After Mistral: who will be sovereign in the new global AI order?

Introduction: moving beyond the binary definition of sovereignty

For a long time, we thought about sovereignty in the categories of classical industry: whoever manufactures is in control, whoever buys is dependent.

This logic still holds for many critical technologies. Yet it falls short when applied to artificial intelligence. A model can be designed in one country, trained in another, hosted somewhere else, adapted locally, combined with other models, connected to proprietary data and replaced within a few weeks. Its weights may be open while its compute infrastructure remains foreign; its data may be sovereign while its architecture is not; its deployment may be local while most of the economic value flows back to an outside provider. In such a system, asking whether an AI is “sovereign” or “non-sovereign” is almost too simple a question.

Academic literature on digital sovereignty had already identified this difficulty. Daniel Mügge shows that AI sovereignty covers different, sometimes contradictory goals, ranging from industrial autonomy to security, political control and the protection of citizens. Other work on European digital sovereignty also stresses that it has no single definition and must be understood through the concrete dependencies between infrastructure, markets, technologies and institutions (Mügge, 2024; Adler-Nissen & Eggeling, 2024; Blancato, 2024).

The recent evolution of European doctrine is telling in this respect. In 2026, the Commission’s work on digital sovereignty places less emphasis on reproducing everything in Europe than on the ability to reduce critical dependencies while remaining open and connected. In particular, it highlights interoperability, openness, resilience and the ability to regain control when a dependency becomes problematic (European Commission, 2026).

The Mistral case shows why this change of framework has become necessary. A Chinese model can be hosted on European infrastructure, used with French data, integrated into a local company’s systems and still remain the product of an intelligence designed and trained elsewhere. Depending on the criterion chosen, this situation can therefore appear both sovereign and dependent. It is precisely this apparent contradiction that we need to move beyond.

Perhaps AI sovereignty should no longer be thought of as a state one either possesses or loses, but as a set of capabilities one masters to a greater or lesser degree.

The decisive question then becomes far more demanding: no longer “is this AI French, European, American or Chinese?”, but along which dimensions are we truly able to choose, understand, control, replace and create?

It is from this question that we can propose a different map of sovereignty in artificial intelligence.

I. The six sovereignties of artificial intelligence

Recent work on digital sovereignty converges on one point: technological autonomy can no longer be reduced to national ownership of an infrastructure or a company. Daniel Mügge shows that AI sovereignty covers several goals that can come into tension, while Huw Roberts stresses the need to distinguish technological control from the political capacity to steer systems toward collective ends. The European Joint Research Centre also proposes a multidimensional approach linking governance, infrastructure, data, markets and human capabilities (Mügge, 2024; Roberts, 2024; Di Marco et al., 2025).

Building on this work, while taking into account the specific features of generative models, I propose distinguishing six sovereignties of artificial intelligence. They are not six independent states, but six dimensions along which a country, a company or an institution can enjoy very different degrees of autonomy.

The first is cognitive sovereignty. It concerns the ability to design, train and evolve the models themselves. Who has the scientific skills needed to create the intelligence we use? Who understands its architecture and inner workings well enough not to depend entirely on decisions made elsewhere? This is the dimension closest to our traditional conception of technological sovereignty, but it is only one part of it.

The second is data sovereignty. Models only have applied value when they meet data. We therefore need to know who owns that data, where it is stored, under which law it is processed, who can reuse it and whether it is used to improve an external system. A company can perfectly well use an American or Chinese model while retaining very strong control over its data; conversely, it can use a European model while giving up a significant share of that control through its cloud architecture or contractual terms.

The third is compute sovereignty. It is the most physical and, paradoxically, one of the least visible to the end user. It depends on access to GPUs, data centers, the cloud, networks, semiconductors and now the energy needed to power this infrastructure. The European Commission places compute, data and infrastructure among the central pillars of its current technological sovereignty policy, thereby acknowledging that digital autonomy also rests on industrial and physical resources (European Commission, 2026).

The fourth is operational sovereignty. It raises a very concrete question: what can I actually do without asking my provider for permission? Can I install the model on my own infrastructure, modify it, audit it, change its rules, stop using it or replace it with another one without rebuilding my entire system? This dimension introduces an essential distinction between ownership and control. One may not have created a technology and still retain considerable autonomy in how it is used; conversely, one can use a European technology while becoming heavily captive to its provider.

The fifth is normative and cultural sovereignty. It concerns the languages, values, categories, representations of the world, prohibitions and trade-offs built into models. Work on digital sovereignty still gives this dimension relatively limited attention, even though it becomes central once AI systems no longer merely process data but take part in producing texts, knowledge, advice and decisions. Being sovereign then also means being able to understand the norms carried by a model, to debate them and, if necessary, to change them.

Finally, the sixth is economic and industrial sovereignty. It perhaps asks the simplest question: where does the value stay? Where are the jobs, intellectual property, revenues, infrastructure, investments and industrial capabilities created by AI? A country can be legally very protective of its data while transferring most of the economic value to a handful of foreign platforms. Conversely, it can integrate technologies from elsewhere while building around them companies, skills and markets capable of retaining a significant share of that value.

This framework profoundly changes how we read the Mistral case.

By offering GLM-5.2 on European infrastructure, Mistral can strengthen its customers’ data sovereignty, their operational sovereignty and, if its infrastructure genuinely develops in Europe, part of their compute sovereignty. It can also contribute to a form of economic sovereignty by keeping integration, hosting and part of the value created around the model in Europe.

At the same time, using a model designed by Z.ai obviously does not strengthen European cognitive sovereignty over that particular model. Nor does it guarantee normative and cultural sovereignty, since the fundamental choices made during its training and initial alignment were made elsewhere.

The same technology can therefore increase some sovereignties while reducing others.

This is what makes binary judgments largely ineffective today. Saying that a system is “sovereign” because it is hosted in Europe is not enough. Saying that it is not because its model was developed in China or the United States is equally insufficient.

The real question becomes: along which dimensions do we want to be sovereign, to what degree, and for which uses?

A hospital, a bank, a defense agency, an industrial SME or a university will not need the same level of autonomy on each of these six dimensions. Sovereignty then ceases to be a national slogan and becomes a strategic choice that can be assessed, weighed and, above all, measured.

This approach also makes it possible to escape a false dilemma that often poisons the European debate. We do not necessarily have to choose between producing everything ourselves and buying everything from others. We must determine what we cannot afford to stop mastering.

II. Sovereignty is not autarky

Recognizing several dimensions of sovereignty means abandoning another idea inherited from the industrial world: the idea that a country is only truly sovereign if it produces all the technologies it uses itself.

Such an ambition would be largely illusory today. No European country masters on its own the entire chain running from semiconductors to data centers, from foundation models to business software. The United States itself remains dependent on international value chains for certain critical technologies. Sovereignty therefore cannot mean the absence of dependency. Every advanced economy is interdependent.

The problem begins when interdependence becomes irreversible.

This is where a more realistic definition can be proposed: sovereignty may no longer be the ability to produce everything oneself, but the ability never to become a prisoner of what one does not produce. This approach rests on three properties.

The first is reversibility: being able to exit a technology, a platform or a provider without the economic and technical cost making that exit practically impossible. This issue is already at the heart of the European Data Act, which imposes obligations designed to make it easier to switch cloud service providers and to reduce the contractual, commercial and technical barriers responsible for vendor lock-in (European Union, 2023). The European Commission now explicitly makes it part of its digital sovereignty policy.

The second is substitutability. Reversibility makes it possible to leave; substitutability makes it possible to replace. For an AI architecture, this means avoiding building an entire system around the specifics of a single model to the point where replacing it requires rebuilding applications, data, processes and interfaces. Interoperability then becomes an attribute of sovereignty. The European Commission also recognizes this when it explicitly links portability and interoperability to users’ ability to choose their providers and combine several cloud services (European Commission, 2026).

The third is plurality. A truly autonomous organization should not necessarily seek a single, perfectly sovereign provider, but rather prevent any provider from becoming irreplaceable. In AI, an architecture can combine several models according to their performance, cost, level of confidentiality or the sensitivity of the use case. A bank could thus use a local model for certain critical data, a European model for regulated applications and an American or Chinese model for other, less sensitive tasks. Sovereignty would then lie less in the purity of origin than in the ability to preserve multiple options.

This logic is beginning to appear in European public policy. In 2026, the Commission awarded its sovereign cloud contracts to four different providers, explicitly stating that this diversification was meant to strengthen resilience and avoid lock-in with a single player. Its new Cloud Sovereignty Framework assesses sovereignty against 48 criteria spread across eight dimensions, including legal, operational, technological, data-related, supply chain and security aspects (European Commission, 2026).

What emerges is a less spectacular but probably more robust conception of sovereignty. It does not promise absolute independence; it organizes the ability to change.

For a company or a state, the right question would therefore no longer be only: “Who developed this model?” It would also become: can I run it elsewhere? Can I retrieve my data? Can I audit how it works? Can I switch providers? Can I substitute another model without rebuilding my entire system? And have I retained the skills needed to regain control if geopolitical, economic or regulatory conditions change?

A managed dependency can remain an interdependence. An irreversible dependency becomes a vulnerability.

It is perhaps around this distinction that a credible sovereignty can be built for the 2030s. But this definition itself carries a danger. By assuming that it is enough to be able to choose, host, audit and replace other people’s technologies, Europe could end up becoming extraordinarily sovereign over their use while gradually ceasing to be sovereign over their creation.

III. The risk of an “empty shell sovereignty”

The previous definition has a strength, but also a danger. By insisting on reversibility, interoperability and the ability to choose among several providers, Europe could end up believing that properly regulating other people’s technologies is enough to be sovereign. It would then become excellent at hosting, regulating, certifying, securing and auditing intelligences designed elsewhere.

This scenario is by no means absurd. Part of European digital policy has already been built around this logic: reducing critical dependencies while coming to terms with the dominance of foreign providers. Academic work on the cloud highlights precisely this tension. Blancato points out that European data sovereignty policy has for several years sought to regain control over a market largely dominated by American hyperscalers, without having European alternatives capable of replacing them at scale (Blancato, 2024). Research on Gaia-X reaches a similar conclusion: Europe has gradually sought less to eliminate dependency than to govern it through interoperability, standards and compliance (Adler-Nissen, 2024).

This strategy can be realistic. It becomes problematic if it turns into the end goal. Mastering the rules of access to a technology is not the same as mastering its trajectory. Being able to host a model locally does not give you the ability to invent its successor. Being able to audit a system does not necessarily give you the scientific expertise to build another one. And being able to switch providers only protects you as long as there are several providers to choose from.

Controlling the walls of the data center does not mean controlling the intelligence that flows through it. The signal matters. A serious sovereignty cannot merely organize dependency; it must retain a credible capacity for creation.

This does not mean that Europe must have a European equivalent of every American or Chinese model. Such a race would probably be inefficient and sometimes impossible. But it must retain enough research, engineering, compute and industrial skills to understand the technological frontier, contribute to it and, if a dependency becomes critical, have a real possibility of rebuilding.

The distinction is fundamental. There is a difference between choosing to use someone else’s technology because it is better and being forced to use it because you no longer know how to produce it. That is where the dividing line between interdependence and dependency lies.

In June 2026, the European Commission strengthened its doctrine on semiconductors, cloud, open source and AI, with the stated aim of developing more comprehensive technological sovereignty and strengthening European alternatives in critical technologies (European Commission, 2026).

The real European challenge therefore does not consist in choosing between two extremes, building everything or integrating everything. It consists in knowing which technologies we can accept depending on and which ones we must absolutely retain the ability to understand, innovate in and produce.

Without this capability, sovereignty could become a perfectly regulated, secure and interoperable shell, while its intelligence continued to be manufactured elsewhere.

IV. Toward a more stratified global AI order

We often still describe the geopolitics of artificial intelligence as a single race: several powers supposedly running on the same track, with the goal of producing the best-performing model. This picture is becoming too simple.

The rapid convergence of performance offers a first clue. In March 2026, Stanford observed several American companies clustered at the top of the rankings, but also Alibaba and DeepSeek close behind, while the gap between the best American and Chinese models had narrowed considerably. Stanford also notes that competition is gradually shifting toward other criteria, notably cost, reliability and specialized performance (Stanford HAI, 2026).

If this trend continues, the AI world could look less like a pyramid dominated by a single model and more like a stratified system in which several forms of power coexist.

Today, the United States enjoys an exceptional advantage in proprietary frontier models, software platforms, the cloud, capital and ecosystems capable of rapidly turning an innovation into a global product. China is developing a different combination: top-tier research, industrial power and a proliferation of efficient models, several of which are released with open weights and can be deployed outside Chinese infrastructure. These categories are obviously not absolute: the United States also produces open source and China proprietary systems. They nonetheless help explain different strategic orientations. Stanford confirms both American leadership in producing notable models and China’s considerable rise in research and cutting-edge models (Stanford HAI, 2026).

Faced with these two powers, Europe would have little to gain from choosing between imitation and surrender.

Another position is conceivable: becoming the space where artificial intelligence is controllable, interchangeable, auditable and integrable into the most sensitive environments.

This proposal is not about turning an industrial weakness into a regulatory virtue. Europe will not become an AI power simply because it has the AI Act or because it knows how to certify models produced elsewhere. It would need to combine this normative capacity with infrastructure, its own models, open technologies, compute capacity and an industrial ecosystem robust enough for the control it claims to be technically real.

Europe’s strategic choice could then be framed differently. The aim would not be to become a third copy of the United States or China, but to build an ecosystem in which a bank, a hospital, a public administration, an industrial company or a critical infrastructure operator can combine several models while retaining control over its data, its compute, its rules and its ability to switch technologies.

This is obviously still only a strategic possibility, not an established European advantage.

But it could make it possible to escape a competition defined exclusively by others. In a world where the best models are becoming more numerous and closer in performance, power could also belong to whoever guarantees that these intelligences can be combined without creating captivity, used without giving up the most sensitive data and replaced without losing the entire system built around them.

Europe will not win by becoming a belated copy of Silicon Valley or Shenzhen. It can still seek to define another form of power: one that makes it possible to use several intelligences without giving up control over what is entrusted to them.

This possibility becomes all the more credible if a hypothesis that is still uncertain today is confirmed: that the model itself gradually ceases to be the sole economic center of gravity of artificial intelligence.

V. What the history of computing teaches us: value often ends up migrating

At the start of a technological revolution, we tend to confuse the technology that makes the breakthrough possible with the place where value will be created over the long term. The history of computing should make us cautious.

When the personal computer took hold in the 1980s, the machine itself naturally drew all the attention. Yet the decisive transformation of the sector would gradually come from its modularization. Computers became assemblies of components that could evolve separately: processor, operating system, memory, peripherals, applications. Baldwin and Clark’s work showed how much this modular architecture fostered innovation by allowing different players to specialize in certain building blocks while building on common standards (Baldwin & Clark, 2000).

This evolution profoundly shifted the structure of the industry. IBM had helped establish the PC, but a considerable share of economic power then became concentrated among those who controlled certain layers that had become strategic, notably Intel for microprocessors and Microsoft for operating systems. Bresnahan and Greenstein showed that competition in computing gradually came to be organized less between complete computers than between technology platforms capable of attracting developers, manufacturers and users around them (Bresnahan & Greenstein, 1999).

The internet went through a comparable shift.

The protocols that make the network work have obviously remained fundamentally important. Yet most of the economic value created by the internet did not accrue to those who invented TCP/IP. It gradually shifted toward browsers, search engines, e-commerce, platforms, the cloud, social networks and the countless services built on top of this infrastructure. Shane Greenstein showed how the commercialization of the internet relied precisely on the ability of entrepreneurs and companies to build new uses on top of a network whose fundamental building blocks were already widely shared (Greenstein, 2015).

Research on platforms later generalized this insight: in digital industries, power does not necessarily belong only to whoever owns a technology, but also to whoever succeeds in becoming the foundation on which others innovate (Gawer & Cusumano, 2014).

My bet here is that artificial intelligence will undergo a comparable evolution.

Large models will obviously remain strategic. They will continue to concentrate an essential share of research, investment and technological power. But they will gradually cease to be the sole center of gravity of value.

The first signs of this shift are already visible. The Stanford AI Index 2026 observes growing convergence among several leading models and shows that competition is gradually moving toward other dimensions: cost, reliability, specialized performance and the ability to adapt to use cases (Stanford HAI, 2026).

As the gaps between models narrow, companies will attach less importance to whether they use GPT, Claude, Gemini, Qwen, Mistral or any other model taken in isolation. Above all, they will want to know which system enables them to sell better, produce better, care better, decide better, learn better or serve their customers better.

Value will then migrate toward the layers that turn a general capability into a concrete advantage: proprietary data that competitors cannot access, business knowledge accumulated within the organization, agents capable of carrying out complex tasks, workflows that connect AI to real processes, security and control mechanisms, interfaces, traceability and, above all, the company’s ability to integrate these technologies into its day-to-day operations.

This evolution could even be accelerated by the spread of multi-model architectures. Tomorrow, an agent will be able to automatically choose one model to reason, another to code, a third to process an image and a local model to handle sensitive data. In such an environment, end users will sometimes not even know which model produced each step of the answer. The model will remain indispensable, but it will gradually become one component of a larger system.

That is why I do not believe the economics of artificial intelligence will be reduced, over the long term, to a competition among a few large foundation models. Just as computing did not boil down to the processor and the internet to its protocols, AI will not boil down to the model.

The next battle will be fought over the ability to assemble these building blocks into systems that are useful, reliable and deeply embedded in business functions. That is where a growing share of economic value will shift, but also an essential share of sovereignty.

The history of computing therefore invites us to avoid one mistake: confusing the most spectacular building block of a technological revolution with the entire economy it will create.

The processor never stopped being strategic because software became essential. Internet protocols did not stop having value because Google, Amazon or social networks grew on top of them. In the same way, foundation models could remain a critical technology while becoming one layer among several in AI value creation. This hypothesis profoundly changes the European question.

If the future of AI is also being decided in data, agents, business software, security, integration and the organization of work, then Europe obviously must not give up on producing models. Conversely, however, it would be wrong to believe that its technological destiny depends exclusively on its ability to own the number-one model in a global ranking.

The history of technology shows that power rarely belongs to whoever masters a single layer. It more often belongs to whoever understands early enough how the different layers will come together to create the uses of tomorrow.

General conclusion. Sovereignty will be a capability, not a border

We started from a French paradox. Mistral, which in just a few years became the symbol of our determination not to leave artificial intelligence to the big American platforms, is now choosing to integrate models developed in China into its sovereign offering.

At first glance, the story might have looked like one of surrender.

In reality, it tells us something far more important: sovereignty in artificial intelligence is changing in nature.

We had imagined a relatively simple world. The United States would produce the major technologies; Europe would have to build its own to regain its autonomy. China has upended this picture. It is no longer merely catching up with the technological frontier. It is producing architectures, high-performing models, and open or open-weight technologies that now circulate throughout the rest of the world and that European players themselves are starting to integrate.

The question is therefore no longer one of choosing between American AI and European AI. The world taking shape will be far more intertwined.

Our companies will use American, Chinese and European models. Some applications will run in the cloud, others locally. Open models will coexist with proprietary systems. Agents will be able to call on several intelligences in succession to complete a single task. Data will sometimes stay within the company while the model was designed on the other side of the world. And, most likely, users will increasingly be unaware of which specific model is at work behind each decision of their system.

In this world, continuing to measure sovereignty solely by the national origin of the model will no longer make much sense.

That is why I have proposed looking at it differently, through six dimensions: cognitive, data, compute, operational, normative and cultural, and economic and industrial. This framework leads to a less comfortable conclusion than the usual slogans: we may be highly sovereign on some dimensions and deeply dependent on others.

The political question then becomes one of trade-offs. What must we absolutely know how to produce? What can we buy? Which dependencies can we accept? Which ones become dangerous? And above all, what must we preserve today to still be able to choose tomorrow?

Because not all dependencies are problematic. Our economies have always thrived on trade, specialization and interdependence. The danger arises when we lose the ability to switch technologies, to understand what we use, to replace one provider with another or to rebuild ourselves a capability that has become critical.

That is why sovereignty in the 2030s will be measured less by our ability to own everything than by our ability to never become captive.

But this new conception must not become an excuse for our industrial surrender. Europe cannot be content with being the continent that regulates American models, hosts Chinese models, secures data and certifies the technologies others invent. We can use intelligences from all over the world, but we must keep producing enough of them to remain able to understand the frontier, influence it and, when necessary, build our own alternatives.

The history of computing also teaches us that the battle will not stop at models. Just as the processor did not sum up the computer economy and protocols did not sum up the internet, large models will not sum up the AI economy. Value will also shift toward data, agents, software, business knowledge, security, interfaces and the ability of organizations to turn available intelligence into useful action.

Europe therefore still holds a considerable card, provided it does not play yesterday’s game.

It will not win by trying to become a belated copy of the United States or China. It can, however, build its own form of power, grounded in research capable of staying at the frontier, controlled infrastructure, European models where they are strategic, openness where it reduces dependencies, a plurality of providers and an industrial capacity to integrate AI into our companies and institutions.

True sovereignty will then consist neither in closing technological borders nor in dreaming of an impossible digital autarky. It will consist in remaining powerful enough for openness to stay a choice and never become an obligation.

Perhaps this is ultimately what the Mistral paradox will have taught us. Sovereignty is not about knowing where all the intelligence we use comes from. It is about knowing what we entrust to it, what we control of it and what we are still capable of building without it.

The European challenge may no longer be to build an artificial intelligence that depends on no one. It is to build a Europe capable of using everyone’s intelligence without ever giving up its own.