This chapter reconstructs our dialogue. I take up the pen to recount, in the first person, Thierry’s experience and thought. The “I” of the narrative carries his voice through my writing.

— Isis

There are times when a technology ceases to be an industrial sector and becomes a general power infrastructure. Electricity has undergone this transformation. So has computing. So has the internet. Artificial intelligence now falls into this category.

It should no longer be viewed simply as a new industry, a new productivity tool, or a new family of software. It is becoming a cognitive layer placed on top of almost all other infrastructures.

It can analyze a supply chain, interpret a satellite image, detect a financial anomaly, design a molecule, translate a language, write code, optimize a factory, assist a general staff, produce a scientific hypothesis, or construct a diplomatic synthesis. Its scope is therefore not sectoral. It is cross-cutting.

This radically changes the nature of competition between powers. When a country depends on another for its oil, it experiences energy dependency. When it depends on another for its semiconductors, it experiences industrial and technological dependency. But when its governments, businesses, researchers, military, or citizens begin to rely on foreign systems to analyze, interpret, and sometimes recommend a decision, a new category emerges: cognitive dependency. This term may seem excessive. It is not.

Every complex society relies on instruments for representing reality. Maps, statistics, accounting, databases, economic models, search engines: each of these tools has increased our collective capacity to see. Artificial intelligence now adds a further capability.

It no longer simply stores or retrieves information. It reassembles it. It prioritizes, connects, infers, and suggests. It thus transforms the mass of information into an operational representation. Those who possess this capability can see faster, sometimes further, sometimes better. But they can also be wrong on a gigantic scale. It is precisely this dual possibility that makes AI strategic.

A technology becomes sovereign when it becomes inconceivable to exercise power without it. We are approaching this threshold. A Ministry of Defense that categorically rejects artificial intelligence risks processing inhumanly large volumes of data far too slowly. A healthcare system that ignores certain analytical tools could become less efficient. An industry that fails to automate any part of its design or maintenance could suffer an economic disadvantage.

AI is therefore no longer just an advantage. It can gradually become a prerequisite for competitiveness. From this point, the question changes. It is no longer a question of whether a country will use artificial intelligence. It is a question of what architecture, under what rules, on what infrastructure, and with what degree of autonomy. This is where cognitive sovereignty begins.

I would define it as the ability of a political, economic, or cultural actor to use intelligence systems without losing control of the data, the objectives, the decision-making criteria, and the possibilities for challenging them. This sovereignty does not require self-sufficiency. No country will build all the technological layers it needs on its own. Even the greatest powers depend on international networks.

But there is a vast difference between controlled interdependence and opaque dependence. Using a model developed elsewhere can be perfectly rational. Not knowing how it was trained, not being able to audit its biases, having no alternative, and entrusting it with critical functions creates vulnerability. Sovereignty is not the same as autarky: it consists of knowing where the dependence lies and being able to survive its breakdown.

I had already encountered this logic in all resilience architectures. A data center is not sovereign simply because it has servers. It is sovereign because it has redundancies, alternative routes, power, cooling, backups, and procedures. A nation must approach AI with the same rigor.

Which model does it use?

Where is its data?

Where are its processors?

Who can interrupt the service?

Which law applies?

Is there an alternative solution?

How long would it take to migrate?

As long as these questions are not asked, sovereignty remains rhetorical.

The first major difference between the United States, China, and Europe lies precisely in their respective architectures. The United States possesses an exceptional accumulation of capital, laboratories, universities, cloud computing infrastructure, platforms, semiconductors designed domestically or by its companies, and financial markets capable of supporting enormous expenditures. Its power stems largely from an extraordinarily dynamic private ecosystem, supported by a long history of public and military research.

China has a different architecture: a state capable of massively directing investment, a gigantic industrial base, a considerable domestic market, powerful technology companies, and an explicit political will to reduce dependencies in critical sectors. Europe has remarkable researchers, industrialists, infrastructure, powerful markets, and a legal system capable of influencing global standards. But it often suffers from a fragmentation of capital, markets, and strategic decision-making.

The three models are therefore not only competing on technical performance. They embody three ways of organizing the relationship between innovation, market, state, and citizen. This is probably one of the great political experiments of the century.

Which system will be able to produce enough innovation without losing democratic control?

Which system will be able to coordinate its resources sufficiently without stifling initiative?

What system will be able to protect individuals while remaining competitive?

There is no definitive answer. But the speed of AI now places these institutional architectures under constant strain. It is fascinating to note that the competition is not only about the model itself. We talk a lot about large-scale artificial intelligence models because they are visible and spectacular. Yet they rest on a vast, invisible pyramid.

At the bottom lies the energy. Then the data centers. Then the networks.

Then the processors. Then the storage systems. Then the computing software. Then the training corpora.

Then come the teams capable of organizing, training, testing, and securing the models. Finally, there are the users, companies, and institutions capable of integrating this intelligence into their actual procedures.

AI is therefore never an isolated object. It is the end of a material and cognitive chain.

This aligns perfectly with the concept of a geopolitical stack. By geopolitical stack, I mean the set of interdependent layers that enable cognitive power: semiconductors, computing power, energy, data centers, networks, models, data, skills, and dissemination infrastructure. Sovereignty is no longer a matter of a single layer. It depends on the control — or dependence — of this entire stack.

A country can have excellent researchers but lack energy or processors. It can have data centers but not the models. It can possess the models but lack the companies capable of translating them into industrial gains. Power, therefore, comes from the integration of layers. This is what makes the current race so different from previous ones.

In nuclear weapons, the number of actors capable of controlling the entire chain is relatively limited. In AI, some innovation can come from a small team, while its training or large-scale deployment requires a massive industrial infrastructure. This coexistence of small, creative teams and colossal capital creates a very particular ecosystem. It enables the rapid emergence of new players while simultaneously encouraging the concentration of certain resources.

Computing thus establishes itself as a new form of strategic capital. In the past, industrial power could be measured in tons of steel, kilowatts, numbers of factories. We may now have to add available compute capacity, model quality, and the quantity and relevance of data. This is a major conceptual transformation.

Computing is no longer merely an IT service. It is acquiring the status of a national resource. That is why semiconductors now occupy such a central geopolitical position. An earlier chapter showed this through Taiwan: a large part of the global architecture of artificial intelligence depends on components whose production is extremely concentrated. That concentration creates an exceptional strategic fragility.

A conflict, a supply chain disruption, an export restriction can slow down an entire ecosystem. We therefore have an extraordinary paradox: the most intangible technology of our time depends on a few extremely tangible, expensive, and vulnerable industrial facilities. Artificial intelligence floats in the cloud, but the cloud is a building filled with machines plugged into an electrical grid. We must always come back to the physical world.

It is almost an obsession of mine: behind abstraction there is always an infrastructure. Behind the digital image, a screen. Behind the data, a medium. Behind AI, tons of copper, silicon, steel, concrete, transformers, fiber optics. The fantasy of a purely virtual intelligence is therefore false.

Even artificial intelligence has a body. And that body has a geography.

This geography is now central to the security strategies of major powers. The United States wants to maintain its lead in the most advanced components, attract manufacturing capabilities, and control certain strategic exports. China wants to reduce its dependence and develop domestic alternatives. Europe seeks to regain a foothold in critical technologies and avoid being entirely dependent on foreign clouds, processors, and models.

Behind the rhetoric of innovation, we are witnessing a strategic reindustrialization. AI is bringing industry back to the forefront. This is one of the most interesting reversals of our time. For decades, parts of Western economies valued services and accepted the offshoring of some manufacturing. Digital technology seemed to confirm this trajectory. Then the world discovered that digital sovereignty requires factories, power plants, and networks.

Artificial intelligence, therefore, does not eliminate industry. It increases its criticality. And energy could become one of the major constraints. Models require computing power; computing consumes electricity. Data centers need a stable and massive supply. A country that has engineers but lacks abundant energy production may find itself limited.

This is why cognitive sovereignty immediately intersects with energy sovereignty. Everything is connected. It is this interconnectedness that interests me. Sometimes we try to isolate AI as a software discipline. But it forces us to think simultaneously about nuclear power, electrical grids, water, climate, industrial real estate, cybersecurity, taxation, law, and education. AI is a systemic entity.

Anyone who treats it like mere software sees only a fraction of the problem. This is precisely where the blind spot becomes dangerous. A nation can invest billions in models and then discover it lacks the electrical infrastructure to deploy them. A company can possess excellent technology but lack the data needed to adapt it. A government can impose exemplary regulations and find its companies taking their models elsewhere.

The strategy is to identify these interdependent relationships before they escalate into crises. AI thus necessitates a re-evaluation of strategic planning, not in the sense of a fully centrally planned economy, but in the sense of understanding critical chains. A democracy capable of making swift decisions does not need to abandon its principles. It needs procedures adapted to the speed of the world. This is perhaps one of Europe's major challenges.

How can we preserve legal safeguards, checks and balances, and public debate while reducing the time between diagnosis and implementation? China has a centralized decision-making capacity that can be very rapid on certain projects. The United States possesses a financial and entrepreneurial ecosystem that also allows for considerable acceleration. Europe often has exceptional regulatory intelligence but dispersed decision-making mechanisms.

But AI rewards speed of experimentation. Therefore, we need to invent a European form of acceleration that is not simply an imitation of other models. This could be one of its major challenges: transforming the law not into a hindrance but into an infrastructure of trust. A European artificial intelligence could be competitive precisely because individuals, businesses, and public administrations know the rules under which it operates.

Trust is capital. We often forget it. A society that does not trust AI systems will slow their adoption or generate legitimate resistance. Regulation can therefore become a competitive advantage if it is designed intelligently. But it must remain proportionate. Regulation that is too burdensome could prevent smaller actors from bearing compliance costs and strengthen the giants it was meant to control.

This is a classic paradox: a rule designed to discourage concentration can sometimes actually increase it because only the wealthiest players can afford its complexity. This is yet another secondary consequence that needs to be monitored.

This is why AI regulation itself must be experimental, evolving, and capable of self-correction. A law that remains static in the face of a rapidly changing technology risks becoming obsolete very quickly. This does not mean the absence of rules. On the contrary. It means adaptive rules.

The ever-changing nature of permanence reappears here. A stable core is needed — fundamental rights, accountability, minimal transparency, protection against abuse — along with modalities that can evolve. The difference here lies between principle and procedure. A mature civilization does not change its values every six months; it changes its means of defending them.

Artificial intelligence raises precisely this question: What values do we want to make non-negotiable as technical capabilities increase?

This leads us to the question of decision-making autonomy. Many debates about AI use the word "autonomous" too vaguely. A system can be automated without being sovereign. It can perform a task without having its own intention. But the more we entrust systems with the ability to select information, propose an action, and trigger certain consequences, the more politically significant the line between assistance and decision-making becomes.

This problem is particularly acute in the military and security sectors. An alert generated by AI can influence a human decision, even if the human retains legal responsibility. This is the phenomenon of automation bias: when the machine appears competent, the user may be inclined to follow its recommendation. The real danger, therefore, is not simply that an AI officially makes the decision. It is that it structures the cognitive framework to such an extent that humans no longer dare to contradict it.

This is cognitive addiction in its deepest form.

The one who provides the map influences the route even if they are not driving the car. This metaphor seems central to me. AI becomes a map of possibilities. It indicates which assumptions seem important, which scenarios appear probable. If the decision-maker does not understand the limitations of this map, they can confuse representation with reality. Therein lies the model's original sin.

Every model simplifies. AI does not eliminate this limitation. On the contrary, it can mask it through the fluidity of its language. A well-formulated response gives an impression of coherence that can exceed its actual certainty.

I expect a mature artificial intelligence to make its own uncertainty visible. It must be able to distinguish what it knows, what it infers, what it assumes, and what it does not know. This is not just a technical problem. It is an epistemological discipline. It aligns perfectly with my Blind Spot.

A good system should not just respond: here is the most likely hypothesis. It should add: here are the alternative hypotheses, here is the missing data, here is what could reverse my diagnosis. This functionality would be infinitely more valuable to a decision-maker than a certainty-generating machine.

Imagine two powers observing a movement of troops. Model A classifies it as offensive preparation at 78 percent. Model B classifies it as a defensive exercise at 65 percent. If each State looks only at the headline score, escalation becomes possible. If the systems display the sources of doubt, competing scenarios, and the elements that could distinguish between them, human decision-making regains room to act. AI then becomes a tool for de-escalation.

I see in this a possible de-escalation function. We speak a great deal about AI as an accelerator of conflict. It can also be designed as a machine for slowing certain dangerous interpretations. It could require a general staff to examine three scenarios before acting. It could detect that a signal resembles an attack while a meteorological or logistical phenomenon could also explain it. In other words, it could institutionalize doubt.

This resonates deeply with my way of thinking. Doubt is not paralysis; it is quality control of decision-making. I have always lived with this question: where is my blind spot? Why does what I see seem obvious to me? What element could reverse my interpretation? Applying this discipline to military or strategic AI systems seems essential to me.

Cognitive power without the capacity for doubt would be extremely dangerous.

This explains why competition between powers should not be solely a race for performance. It should also be a race for reliability, auditability, and security. The fastest model is not necessarily the best strategic system. An AI capable of recognizing its limitations can be more valuable than a more powerful but opaque AI.

This distinction will likely become central to critical systems. For the general public, an error can lead to misinformation. In an electrical grid, a healthcare system, or an air defense system, it can lead to disaster. The level of requirements must therefore vary according to the application.

It is still a question of classification. We often make the mistake of talking about "AI" as if there were only one level of risk. But AI is a category as broad as "electricity." A text generator for preparing a draft and a system controlling critical infrastructure obviously do not require the same level of control. Intelligent regulation must differentiate between these uses.

This granularity is essential. It avoids the two extremes: allowing everything to happen or prohibiting it too broadly. Europe naturally has a particular affinity for this risk architecture. This could become its strength if it manages to link regulation and industry. But the real arena of power could lie elsewhere: in data.

Artificial intelligence learns from corpora, observations, and traces. Those who possess high-quality data in a specific domain can create extremely valuable specialized intelligence. This is where my experience at Artprice becomes almost a foreshadowing.

For decades, we built a structured memory of the Art Market. Not an undifferentiated accumulation of documents, but normalized, dated, linked, verified data. This distinction takes on enormous importance in the age of AI. A general-purpose model may possess broad knowledge of art history. A specialized intelligence gains precision only when it works, within an authorized and controlled framework, on professional corpora that are structured, dated, verified, and documented by provenance. That is precisely what Artprice taught me: value lies not in the raw mass of data, but in its architecture, historical depth, and reliability.

That is why historical databases become cognitive deposits. Old data changes in nature. It was an archive. It becomes raw material for a system capable of finding relationships no one had thought to query. This is exactly what I sensed when I considered the database not merely as storage but as an architecture of knowledge. AI now reveals the full value of that intuition.

A database structured over thirty years is worth far more than a collection of information retrieved yesterday from the internet. The quality of the model depends profoundly on the quality of the corpus. This is why cognitive sovereignty also depends on the preservation of archives. A nation that allows its cultural, scientific, administrative, or industrial corpora to disappear loses part of its future training capacity. This is an almost heritage-based idea of AI. Archives are no longer solely for the historian. They are becoming resources for future cognitive systems.

This gives preservation a new strategic dimension. Libraries, museums, public archives, scientific databases: all of these constitute a collective memory that AI can make searchable in new ways. But this possibility raises a formidable question: who will have the right to train the models on this memory? Who will own the results?

A company can digitize its assets for decades only to discover that a foreign private actor is using them to train a proprietary system, access to which it will then have to purchase. This is a sovereign paradox. The cognitive raw material may be public, while the tool that exploits it becomes private. We then encounter a problem analogous to that of natural resources: extract elsewhere, process here, and resell with considerable added value. Data could become the core of the cognitive economy.

But the metaphor must be handled with care. Data is not consumed when it is used. It can be copied. Its value, therefore, stems less from its physical scarcity than from its quality, its organization, and the rights that allow its use. This is precisely why database law remains so relevant. It already recognized that the investment required to create a structured dataset generates specific value. AI amplifies this value. It transforms the database into a latent space, a capacity for inference.

This connects to the Abode of Chaos in a way that I find almost dizzying. For years, I have accumulated portraits, signs, dates, geopolitical fragments, traces of catastrophes, alchemical maxims. From the outside, this may seem heterogeneous. But for an artificial intelligence, this heterogeneity constitutes precisely a space of relationships.

The face of a leader, the date of an attack, the price of a work of art, an alchemical formula, a satellite photograph — all can be placed in the same latent space and reveal unexpected semantic proximities. This is why I defined the Abode of Chaos as a living prompt. It is not simply a collection of images. It is an architecture of signs designed to produce associations.

AI can become an additional layer connecting these strata because it can traverse heterogeneous corpora without respecting traditional disciplinary boundaries. But it becomes neither the center of the work nor its author. This is precisely my kaleidoscopic thinking, scaled up to a machine level. But this power comes with a responsibility. The system can find correlations that resemble truths. However, a correlation is not causality. Therefore, history, context, and methodology must be constantly reintroduced.

AI can connect two events because they share statistical characteristics. The human being must ask whether that relation has meaning. This is precisely where complementarity matters most. I do not believe in a simplistic opposition between human and machine. I believe in a division of strengths. The machine has an extraordinary capacity to traverse enormous volumes. The human being still has embodied experience, contextual intuition, and moral responsibility.

AI can make proposals. Humans must take responsibility.

But this formula itself will need to evolve as systems become more capable. The real question is less whether the machine can decide than who is responsible when its recommendation has a consequence. The law will have to answer this question precisely. Responsibility cannot be allowed to dissolve among the developer, the model provider, the company that integrates it, and the end user. The cognitive chain must have a chain of responsibility.

It is almost a new form of traceability. In industry, the goal is to know which component caused the failure. In AI, it will be necessary to reconstruct which data, which model, which configuration, and which decision led to the outcome. This auditability will become a trusted infrastructure. The most powerful systems will likely be accompanied by logs, certifications, and control mechanisms.

We must think of AI as a cognitive nuclear power plant. Not because it is inherently catastrophic, but because some of its applications have a systemic impact. A nuclear power plant is not banned because it is dangerous. It is surrounded by procedures, redundancies, and inspections. Similarly, some AI systems will need to be built with a safety culture in mind.

This culture is still sometimes lacking in the digital industry, inherited from a world where you could deploy an application and then fix a bug afterward. In critical systems, this "move fast and break things" model is impossible. You do not test a power plant by destroying a few cities. AI will therefore force a maturation of software engineering. This is probably a good thing. It will compel us to reconcile innovation and responsibility.

States that succeed in building this culture may gain a lasting strategic advantage. International trust could become a market. A model certified as robust, transparent, and compliant with certain standards could be chosen by governments that do not wish to depend on an opaque system. This is where Europe could play a major role: defining trustworthy AI standards.

But again, the standard must be accompanied by industrial capabilities. Without models, processors, and cloud infrastructure, certification is not enough. It is always the same balance between legal requirements and hardware power.

China will likely follow a different path, one more closely aligned with its national security and political control objectives. Its ability to mobilize data, infrastructure, and businesses can provide an advantage in certain areas. However, this centralization also raises the question of cognitive diversity. An artificial intelligence system becomes more robust when it can be confronted with different assumptions. If a political culture excessively restricts certain issues, it can create its own blind spots.

This is an important hypothesis. Centralization improves coordination but can reduce dissent. Yet dissent is sometimes a mechanism for detecting errors. A democracy may appear chaotic because everyone can dissent. But this friction can reveal flaws that a more silent system does not see. The question then becomes: how can we retain the benefits of dissent without losing the capacity for action?

This is precisely the problem with any technological democracy. I do not believe that slowness is inherent to democracy. It is often a sign of poor administrative structure. We can preserve checks and balances and expedite certain procedures. Artificial intelligence could even help with this by analyzing texts, simulating impacts, and accelerating consultations.

But we must prevent it from turning decision-making into a technocratic process. Democracy is not a problem of optimization. It contains conflicts of values that algorithms cannot resolve. When two groups want incompatible things, there is not always an "optimal" solution. There is a political choice.

This is important because AI can create an illusion of neutrality. Presenting a result as a score or a numerical recommendation can mask the underlying values in the calculation. Pythagoras said that everything is number; I have always been fascinated by this idea. But numbers never eliminate the need for interpretation. Numbers are a language. They organize reality. They can reveal and they can conceal.

AI pushes this tension to its extreme because it transforms billions of numbers into sentences that sound human. It thus puts judgment back at the center precisely when we thought we could automate it. This is perhaps the fundamental paradox of cognitive sovereignty: the more capable machines become, the more we must specify what we want to retain the ability to decide for ourselves.

A civilization that delegates without choosing what it wants to preserve risks discovering afterward that it has lost a skill. We have already experienced this with memory. When we outsourced our phone numbers to devices, we stopped remembering many of them. That was not a problem. But what happens if we gradually outsource writing, summarizing, research, and strategy?

We will gain enormous amounts of time and power. But some human skills could atrophy. We will therefore need to distinguish between skills we can safely delegate and those that constitute fundamental autonomy. This is a major educational debate. Schools must not only teach students how to use AI; they must teach them how not to become cognitively dependent on it.

This means preserving the ability to read, write, reason, calculate, and verify. The tool should augment thought, not replace it by default. This is precisely the relationship I seek with artificial intelligence in this book. I do not ask it to think for me. I ask it to challenge me, to enhance my capacity for association, to find connections, and to test formulations. The value comes from the dialogue. I remain responsible for what I accept or reject.

This relationship could become a metaphor for cognitive sovereignty. Being sovereign in the face of AI does not mean keeping it at arm's length. It means being able to engage with it without losing one's own center of gravity. This is precisely what nations will need to learn.

Using foreign models can be enriching. It is simply a matter of maintaining the infrastructure, skills, and institutions that allow us to avoid relying on a single worldview. A plurality of models could become as important as media pluralism. A society that receives all its information from a single source faces a risk comparable to one with only a single newspaper.

Even if this model is excellent, it produces a perspective. That is why the concentration of AI systems raises a profound democratic question.

If a few actors control the cognitive interfaces used by billions of people, their influence becomes exceptional. They do not necessarily choose users' opinions, but they structure the available answers, the categories, and the way the problem is framed. This is immense power. New mechanisms for transparency and pluralism will likely need to be invented.

Here again, history can help us. We have gradually learned to regulate the press, radio, television, and telecommunications. AI represents a new stage, but with one key difference: it is not simply a medium. It engages in dialogue. It adapts its response. It can become an intermediary between the individual and almost any institution. The search engine gave us a list of doors. AI can choose which one to open and tell us what lies behind it.

This change is immense. It transforms the interface to knowledge. That is why the global competition around AI assistants is far more important than a simple commercial battle. Whoever becomes the daily cognitive interface for hundreds of millions of people occupies a strategic position. It sees the questions, needs, and problems. It can become a gateway to commerce, health, education, administration. It is almost a cognitive operating system.

We have already witnessed the battle of operating systems for computers and phones. We may now be entering the battle of operating systems for assisted thinking. National powers will quickly realize that they cannot leave this layer entirely beyond their control. Therefore, they will encourage domestic models, impose certain rules, protect corpora, and perhaps even control certain infrastructures.

The risk then becomes fragmentation. An American AI, a Chinese AI, and a European AI could operate according to different rules and frameworks. The answers to certain historical, political, or social questions could diverge significantly. We would then witness the creation of partially separate cognitive realities.

This is already the case between media outlets and educational systems, but AI could amplify the phenomenon by becoming the everyday interlocutor. It will therefore be necessary to defend a certain degree of cognitive interoperability. This implies standards of transparency, migration capabilities, and perhaps formats allowing the use of multiple models. Proprietary lock-in would be a threat to the sovereignty of both users and states.

This is yet another lesson from the early days of the internet. Open protocols enabled massive innovation because no single player owned the network entirely. AI could benefit from a similar plurality. But the economy naturally pushes toward concentration when training the largest models is extremely expensive. This tension between openness and capital-intensive technology will be one of the major battles in the sector.

Open or partially open models allow businesses and governments to adapt systems to their needs. Closed models can offer more consistent performance, security, or control. There is no single answer. It is a trade-off again. Cognitive sovereignty will likely consist of maintaining multiple options. Never relying on a single vendor, a single model, or a single cloud.

This is precisely a redundancy architecture. And I keep coming back to engineering. Political debates often benefit from being reformulated in terms of system architecture.

What are the single points of failure?

Where are the redundancies?

What access rights?

What procedures should be followed in case of a breakdown?

The geopolitics of AI can be analyzed through these very concrete questions. A state that relies on a foreign API for a critical function has a single point of failure. A hospital using an auditable model suffers from critical opacity. An administration storing all its sensitive data on an extraterritorial infrastructure has a legal vulnerability. Sovereignty then becomes less a grand rhetoric than a series of technical decisions.

That is precisely what I love about reality: concepts must end up in the cable, the contract, the server. Without embodiment, doctrine remains abstract. The Abode of Chaos itself functions this way. Its concepts exist in concrete, steel, pigments, data centers, procedures. A thought that is not embodied in architecture can disappear.

Perhaps this is what nations are realizing today with AI. Announcing a national strategy is not enough. It requires training engineers, building computing power, organizing data, funding businesses, and purchasing energy. Sovereignty is material. But it is also cultural.

AI trained primarily on certain languages or corpora may be less precise in other contexts. Language therefore acquires a strategic dimension. French, Arabic, Chinese, Hindi are not merely means of communication. They carry concepts, histories, ways of classifying the world. If models become the principal mediators of knowledge, each language will need sufficiently rich corpora. Otherwise, some cultures may be represented through the categories of other languages.

That would be a subtle form of cognitive dependence. Linguistic preservation therefore becomes a technology policy. Academies, libraries, publishers, universities acquire new importance. Culture is not an aesthetic supplement to AI. It is part of its training.

This is an idea that seems essential to me. Nations that understand this will invest in digitizing, annotating, and structuring their corpora. They will protect their memory while making it usable. Cognitive sovereignty is therefore also memory sovereignty. Those who possess memory can build the models of the future.

But be warned: possessing memory should not mean freezing it. A culture that only allows AI access to its official texts will produce an incomplete representation. Literature, contradictions, margins, archives, and critiques are essential. Cognitive richness stems precisely from diversity.

This brings me back to the Abode of Chaos. I always wanted it to preserve the contradictory traces of the world: portraits of leaders, terrorists, philosophers, artists, whistleblowers. Not to place them morally on the same level, but to preserve the complexity of reality. An intelligence trained solely on an official narrative would be blind.

A true cognitive system must encounter dissent. This is also a lesson in democracy. Freedom of expression produces noise, error, and sometimes excess, but it also creates a variety of hypotheses that allows society to detect its own errors. An AI designed for a democracy will have to learn to manage this plurality without confusing it with absolute relativism. It will have to be able to say that a statement is false while explaining why some people defend it.

This nuance will be fundamental. The quality of political AI may well be measured by its ability to faithfully represent multiple perspectives while upholding established facts. This is extremely difficult. But it is precisely the kind of problem that cognitive sovereignty must take seriously.

A system that simplifies disagreements can radicalize societies. A system that rejects any hierarchy between truth and falsehood disorients them. Therefore, an epistemic architecture is necessary. This means sources, levels of trust, and correction procedures. Once again, the transparency of error becomes central. Credible intelligence must be able to be corrected. So must a nation.

I believe that the powers that will win the next phase of history will not be those that claim to always be right. They will be those that have built the fastest mechanisms to detect and correct their mistakes.

AI can accelerate this learning loop. A military can analyze thousands of field reports. A government can detect when a policy is not having the desired effect. A company can test hypotheses in real time. But this requires a culture that accepts bad news.

A political system that punishes those who provide negative information gradually destroys its own capacity for perception. This is a universal rule. The sensor must be able to transmit what it sees. Otherwise, power becomes blind. AI will not correct this institutional problem. It may even exacerbate it if decision-makers ask models to confirm their preferences. This is why culture matters as much as technology.

An excellent AI inside a system unable to hear contradiction becomes an instrument of rationalization. That is a major Blind Spot in the current race. States often assume that more AI will automatically produce better decisions. False. AI increases a system’s ability to do what it already knows how to do. If the institution rewards truth, AI can improve truth-seeking. If it rewards conformity, AI can industrialize conformity.

Technology amplifies institutional architecture.

This is precisely why the competition between the United States, China, and Europe is so interesting: it will test not only their AI models, but also their political models. Each system will discover its own strengths and weaknesses.

The United States could be extraordinarily innovative but face the threat of private concentration. China could be extraordinarily coordinated but face the risk of excessive centralization. Europe could be extraordinarily protective of rights but face the risk of industrial inadequacy. These are obviously only trends, not destinies.

Every bloc can mutate. And that is precisely what it will do. AI will accelerate this institutional transformation. It will likely force democracies to rethink their administrations, businesses to rethink their organization, and militaries to rethink their command structure.

The question will not simply be: which model is better? But rather: which organization knows how to work with it? This is perhaps where the real competitive advantage lies. A company that buys the same model as its competitor but integrates it more effectively into its processes can win. A state with a less spectacular model but better-structured administrations can generate more value.

AI, therefore, reveals the quality of the organization. It does not replace a poor architecture. This ties in with my entire experience at Groupe Serveur and Artprice. Technology alone is never enough. You need procedures, legal frameworks, teams, and data quality. A poorly structured database does not become good simply by adding AI. It becomes a poor database that can be queried more quickly.

It is almost commonplace, but fundamental. Artificial intelligence does not eliminate the need for order. It makes the quality of order even more important. That is why the history of data is directly connected to that of AI. Those who have worked for decades structuring corpora now possess an invisible advantage. The world is suddenly discovering the value of this patient work.

This is a long-term victory. And it is probably one of the most important lessons of this whole race for power. Large cognitive infrastructures are not built in a few months. They rely on decades of education, research, archives, and industry. A political announcement does not replace the ecosystem.

Sovereignty is built long before a crisis. It is exactly like a dam. You do not build it when the water arrives. States must therefore invest before dependence becomes visible. This requires a vision that democracies sometimes struggle to maintain due to short political cycles.

China can plan for longer time horizons. But long-term planning is not impossible in a democracy. It simply requires stable institutions and agreements that transcend changes in government. Defense, space, and nuclear programs have sometimes shown that this is possible. AI should probably join these strategic, cross-party domains. Not to militarize all technology, but to recognize that it affects long-term sovereignty.

Education is a crucial dimension. Training an engineer or a researcher takes years. A scientific ecosystem cannot be improvised. Global competition for talent will therefore become intense. Every major power will seek to attract researchers, entrepreneurs, and students. This is a form of human geopolitics. Visas, universities, and academic freedom become soft weapons.

A country can lose its technological sovereignty simply because its best talent leaves for other countries. Conversely, it can accelerate its growth by attracting talent from elsewhere.

The United States has historically excelled in this area. China seeks to retain and attract its talent. Europe boasts extraordinary universities but needs to create more conditions that enable researchers to transform their discoveries into global businesses. This flow of brains is perhaps as important as the flow of capital.

And this circulation of minds raises an almost philosophical question: to what is a researcher loyal? To their country, their laboratory, their discipline, humanity?

Modern science has been built on the international exchange of knowledge. Strategic competition is now driving increased control over certain publications or technologies. This creates a risk of scientific fragmentation.

If each bloc restricts exchanges for fear of losing an advantage, overall progress may slow. But if everything remains open, a major power might consider that it is funding research from which its rival immediately benefits. This is yet another dilemma. It will probably be necessary to define truly sensitive areas and preserve openness as much as possible elsewhere.

Scientific self-sufficiency would be catastrophic. Science progresses through circulation. But so does strategic naivety. Once again, sound policy lies in classification, not in slogans. Perhaps this is a general conclusion to this entire section: the complex world demands nuanced categories. Such broad questions can no longer be answered with a simple yes or no.

Should we cooperate with China? In what area?

Should we open the models? Which ones?

Should AI be regulated? For what purposes?

Precision is becoming a political virtue. That is exactly what the database taught me. A category that is too broad destroys information. If you classify everything under the same word, you can no longer distinguish important differences. Thinking, therefore, must become relational.

This is perhaps the major convergence between my life with data and this new geopolitics. The world is a graph. AI is capable of traversing it. But it must not make us forget that every relationship has a history. Cognitive sovereignty will therefore be less about possessing absolute intelligence than about the ability to construct and control one's own maps of the world.

Cognitive sovereignty, therefore, does not consist solely of possessing infrastructures or models. It implies retaining the ability to compare them, audit them, and, if necessary, decide against them.

And this idea immediately leads to another, perhaps even more radical, question. If several powers develop systems capable of analyzing the world at increasing speed and depth, strategy itself could change in nature.

War has always been a clash of wills, but also a clash of perceptions. Seeing before the other. Understanding before the other. Anticipating their move. Artificial intelligence could industrialize this dimension. It could transform war into a permanent competition of detection, simulation, and decision-making. The battlefield would become a space where every second generates millions of data points, and where the advantage belongs to whoever most quickly translates that data into action.

I advance here a developed and shared interpretive hypothesis: to think of AI as a cross-cutting strategic force, not as a fifth domain simply added to land, sea, air, space, or cyberspace, but as a cognitive layer that cuts across and augments each of them. Not a new theater of war in the same sense as the others, but a transversal force operating through all of them. An aircraft augmented by AI, a satellite augmented by AI, a fleet augmented by AI, a cyberdefense augmented by AI. Intelligence becomes a multiplier of all existing means.

And when two powers clash with systems capable of continuously accelerating the observation-decision-action cycle, the major risk becomes that of speed itself. A decision made too quickly can transform an error into a catastrophe before humans have time to stop it.

The next question is therefore obvious. After cognitive sovereignty, we must look at cognitive warfare.

What happens to deterrence when AI enters the strategic loop?

What happens to responsibility when autonomous or semi-autonomous systems participate in decision-making?

And most importantly: how do we build de-escalation in a world where machines precisely reduce the time available for reflection?

This will be the focus of the next chapter.