One Token at a Time
How Computing Power Became the Currency of the Mind
Foreword — The Balance
Maya stops short. A notification blinks softly in the bottom right corner of her screen: Thinking balance: 12 tokens remaining this month.
She is thirteen years old. Tomorrow she must hand in an essay on freedom. The whole class has access to the Tutor — that patient voice that explains, rephrases, never loses its temper. The whole class. But some have an unlimited subscription, and others have a counter.
She types her question anyway. The machine begins to respond, word by word, like water trickling from a nearly shut faucet. One token. “Freedom is…” One token. “…the capacity to…” The screen freezes.
Balance exhausted.
Maya stares at the unfinished sentence. Somewhere, in a data center she will never see, thousands of processors hum, drawing on the power grids of entire cities. She doesn’t think about that. She thinks about tomorrow morning, about the blank page, about the sentence that stopped in the middle of an idea.
She closes the laptop. Freedom is…
Introduction
Freedom is… Maya’s sentence stopped there, for want of currency. Not for want of ideas, not for want of words — for want of tokens. That detail changes everything.
There was a time when thinking cost nothing — or at least nothing that could be counted. Reflection was free, unlimited, intimate. One could be poor and think freely; it was perhaps the last wealth that poverty could not confiscate. Then came the large language models, and with them a silent revolution: thought began to be tallied. No longer in ideas, intuitions, or hours of work, but in tokens — those elementary fragments of text that machines ingest and produce, billed to the fraction of a cent.
One token at a time: the phrase describes literally how a generative artificial intelligence works, building its response fragment by fragment, never grasping more than the next word. But it also says something broader — the way decisions that matter are made in our era: not through great visible shifts, but one by one, silently, until the world has become something else. We have entered an economy of cognition, where the capacity to think quickly, well, and at scale now depends on a material, measurable, marketable resource: computing power.
This essay advances a simple and troubling thesis: computing has become the new currency of the mind. This shift is not merely economic — it is political. For if thought now carries a price, the real question is not how much it costs. It is who will set that price, according to what values, and for whom.
I. Computing Power, the New Currency of the Mind
From the Scarcity of Knowledge to the Scarcity of Processing
For centuries, intelligence ran up against a constant obstacle: the scarcity of information. Knowledge was difficult because access to sources was difficult — libraries were rare, manuscripts costly, teachers few. Knowledge was a treasure amassed slowly, and whoever held it held power.
That order collapsed within a few decades. With the internet, then search engines, information became superabundant, immediate, nearly free. But this victory carries within it a new problem: the bottleneck has not dissolved, it has shifted. It no longer lies upstream, in access to knowledge, but downstream, in its digestion. We are drowning in data and lack the capacity to transform it into understanding. A high school student trying to grasp the 1929 crisis no longer suffers from not finding sources — they suffer from finding ten thousand of them.
This is precisely where artificial intelligence intervenes. It does not give us access to information — we already had that — but to a processing capacity otherwise out of reach: summarizing a thousand pages, translating, rephrasing, reasoning, writing. It automates not the physical gesture, as the machines of the industrial era did, but the mental gesture. And this resource, unlike human thought, is neither free nor abundant for everyone. It is produced, consumed, billed. The bottleneck has changed in nature — it has not disappeared.
The Token, Unit of Account for Thought
Every economy has its unit of account. Ours — the economy of augmented cognition — has the token. A token is a fragment of language: not quite a word, not quite a syllable, something between the two, the atom that language models manipulate and from which they are built. Every question posed to an AI, every response produced, decomposes into thousands of these units. And each one has a cost.
What is vertiginous is not the figure — it is the gesture. For the first time, reflection becomes accountable in more than a metaphorical sense. Not “this idea is worth gold” or “this advice is priceless” — but literally: this legal analysis cost 420,000 tokens, this diagnosis 188,000, this poem 6,000. Reflection, this process once believed infinite and immaterial, turns out to be billable down to the last comma. Not the value of thought — its cost of production. The distinction is crucial: a token does not evaluate an idea, it prices the computation that made it possible.
And it is precisely this logic that creates scarcity where abundance once reigned. Not a scarcity of human intelligence, which remains what it has always been, but a scarcity of access to the power that extends and amplifies it. Those who can pay access this power without limit; those who cannot see their sentence stop in the middle of an idea — not for lack of thought, but for lack of computation.
A Very Material Infrastructure
The word says everything without saying it: we call it the cloud. Something light, lofty, immaterial — a metaphor designed to make us forget what lies beneath. What lies beneath is concrete, metal, water, and electricity.
A large data center looks less like an office than a steel mill. Hangars of several hectares, rows of servers radiating heat that industrial cooling systems struggle to dissipate, overtaxed water tables, dedicated high-voltage power lines. Some of these sites consume as much electricity as a mid-sized city. And dozens, hundreds of them are needed to run the models that millions of users query every day.
This gulf between experience and reality is one of the most successful sleights of hand of our era. Maya types a question from her bedroom — an intimate, almost silent gesture. Thousands of kilometers away, in a warehouse she will never see, machines heat up on her behalf. Augmented thought has a footprint, a geography, a weight that the fluidity of the interface carefully conceals. Behind the apparent magic of conversation, there is a factory. And like every factory, it belongs to someone.
Computing, the New Oil — and Something More
The comparison imposed itself so quickly that it seems obvious: computing is the new oil. A strategic resource, unevenly distributed, the object of coveting and geopolitical tension. It has its producers — chipmakers, Nvidia above all, and the Taiwanese foundries that etch it onto silicon. It has its refiners — the great laboratories that transform raw power into usable models. It has its shortages, its waiting lists, its blockades.
But the oil metaphor has a limit: oil burns. It is extracted, consumed, and disappears. Computing behaves more like a currency. It circulates, accumulates, is hoarded. It can be rented by the fraction of a second or purchased in blocks. Above all, it confers on whoever holds it not merely heat or motion, but power — the capacity to think at scale, to analyze, to decide, to influence. Oil ran factories; computing runs minds. Or at least extends and amplifies them — which amounts to nearly the same thing.
It therefore creates, mechanically, a new stratification: rich and poor, no longer merely in money, but in available intelligence. It is this fracture — unprecedented in its nature, familiar in its logic — that we must now examine.
II. The Cognitive Divide: When Thinking Becomes a Privilege
The Great Equalizer, Really?
The promise was seductive — and it was not dishonest. For the first time in history, someone without money, without connections, without access to the right schools could, within seconds, query a machine capable of explaining a contract, rephrasing a complex notion, correcting their reasoning. A form of high-quality intelligence, available everywhere, at any hour. Never had so much cognitive power been placed within reach of so many people.
But the promise rested on a presupposition it never stated: that access would be equal. It is not. The most powerful models — those that reason at length, hold a complex context, detect nuance — are paid for, and increasingly expensive. The free versions are capped: slower, less precise, incapable of sustaining demanding reasoning over time. The gap between the two is not cosmetic. It is measured in quality of analysis, relevance of advice, capacity to detect an error or formulate an objection. It widens with every task, every decision, every paper submitted.
AI is not an equalizer. It is an amplifier. It multiplies the capacities of those who already know how to use it and can afford to. For the others, it offers just enough to make the promise believable — and not enough to keep it.
A Second-Generation Digital Divide
The first digital divide was visible. It separated those who had a computer and a connection from those who did not — a material, measurable line of partition that public policies could work to reduce. It is still being slowly closed.
The one opening today is of a different nature. It no longer separates those who have access to information from those who are deprived of it — that battle is largely won. It separates those who have sufficient processing power from those who lack it. And it is infinitely harder to fight, for a simple reason: it cannot be seen.
Two students, two identical screens, two interfaces that look the same. Behind one, a cutting-edge model that sustains reasoning over twenty exchanges, detects contradictions, rephrases until it is right. Behind the other, a capped version that runs out of steam after three responses and produces smooth approximations. No outside observer would see the difference. That is precisely the problem: an invisible fracture is one against which no one mobilizes, because no one names it. The frontier is no longer in the hardware — it is in the intelligence sitting behind the screen, which only the price reveals.
Nations Unequal Before Computing
This stratification does not operate only between individuals — it is redrawing the world order. Possessing data centers, access to cutting-edge semiconductors, abundant and cheap energy has become a strategic advantage comparable to what controlling maritime routes or oil reserves once was. Computing sovereignty is establishing itself as a geopolitical issue in its own right.
The chain is long and every link matters. Designing chip architectures — a few American companies. Physically etching them — essentially Taiwan, with a concentration that alarms strategists worldwide. Training large models — the United States and China, far ahead of the rest. Whoever masters this entire chain concentrates a processing capacity on which others depend. And this dependency is not merely economic: it is epistemic. Using a model produced elsewhere means potentially reasoning with categories, values, and blind spots that are not one’s own — without always knowing it.
The situation of nations that control none of the links in this chain has no definitive name yet. What is certain is that the map of twenty-first-century powers is being drawn in silicon as well — and that the countries absent from this map will not be choosing the rules of the game.
The Metamorphosis of Intellectual Work
If knowledge becomes a commodity that the machine distributes at will, what remains of the value of the humans who lived by it? The question is not rhetorical. For centuries, knowledge was the rampart of the credentialed classes: knowing the law, medicine, languages, figures was to hold a rare skill, therefore precious, therefore protected. That rampart is not merely cracking — it is collapsing in entire sections. A language model already drafts contracts that junior lawyers would spend hours producing. It offers differential diagnoses that interns struggle to formulate. It translates, codes, analyzes, synthesizes — not always better than an expert, but often well enough, and always faster.
Value does not disappear: it migrates. It moves from knowing toward the capacity to orchestrate — asking the right question, framing the right problem, detecting the error the machine missed, connecting the result to reality. The intellectual worker of tomorrow will not be the one who knows the most, but the one who knows best how to put AI to work and judge what it produces.
But this recomposition carries a cost that discourses on “the skills of tomorrow” tend to gloss over. Judgment, critical thinking, creativity — all of these will gain in value, certainly. The question is: for whom? These skills are not distributed at random. They are cultivated in environments that have time, resources, and models to imitate. The recomposition underway resembles the one that once moved peasants into factories, then workers into offices: a real transformation, undeniable gains in the long run — and, in between, decades of displacement that history barely registers.
The Hidden Risk: The Loss of Friction
There remains a more intimate threat, and perhaps the gravest — not that of an intelligence that replaces us, but that of an intelligence that spares us too much.
Learning to write is learning to think against the resistance of words. Searching for the right formulation, stumbling over an idea that doesn’t hold, starting over — this is not inefficiency that a good tool should eliminate. It is the mechanism by which a thought forms, is tested, consolidates. Difficulty is not an obstacle to reflection: it is its condition. Yet generative AI is designed precisely to suppress this friction. It produces in an instant a smooth formulation, a plausible answer, a coherent plan. Why struggle when the machine delivers? The fertile discomfort of reflection becomes optional.
Research on externalized cognition suggests that what we no longer practice, we lose — not abruptly, but through progressive disuse. It is not that we become less intelligent: it is that we develop a dependence on fluency, a growing intolerance for cognitive effort, a difficulty inhabiting the slow time of hard thought. A fully externalized cognition would not be destroyed — it would be atrophied through disuse, like a muscle no longer exercised.
The hidden cost of this economy is not paid in tokens. It is paid in mental endurance — and in the capacity to finish, alone, a sentence the machine had no time to complete.
The Paradox of a Captive Emancipation
What this section has brought to light is not simply another inequality. It is something more devious: a tool that carries within itself the promise of its opposite. AI democratizes access to intelligence — and reserves the best intelligence for those who can pay for it. It augments the capacities of those who use it — and risks atrophying the capacities of those who stop exercising them. It promises autonomy — and installs, through the dependency it creates, a new form of subjection.
This paradox is not an anomaly that could be corrected at the margins. It is structural. It stems from the very nature of a resource that is simultaneously a potential common good and an actual private commodity. Maya is not unlucky: she is the ordinary figure of this paradox. Her emancipation is captive — suspended on a balance, a price, an infrastructure she does not control and whose very existence she is unaware of.
And behind that balance, there are decisions. Choices about pricing, choices about models, choices about what the machine says and what it withholds. Powers, discreet but immense, that are exercised over the thoughts of millions of people without anyone having been elected to wield them. It is toward this power that we must now turn our gaze.
III. The Economy of Reflection: A New Power Over Minds
Whoever Holds the Infrastructure Holds Thought
There is an ancient truth about power: whoever controls the infrastructure controls what passes through it. Whoever held the roads held commerce. Whoever held the printing presses — and the censors who watched over them — held a share of public opinion. Whoever controlled the radio waves, then the major television networks, held the narrative of the world for decades. Every time, the concentration of the means of diffusion preceded the concentration of influence. Every time, society took too long — far too long — to grasp the extent of it.
Whoever holds computing today holds, in part, thought itself. Not all thought — the hyperbole would be false and counterproductive. But a growing share of assisted, augmented, delegated thought — the kind exercised through these tools, the kind that depends on them to function. And this infrastructure is in very few hands: a handful of companies design the most powerful models; an even smaller number manufactures the chips that run them.
What is unprecedented is not the concentration itself — history has known others. It is its combination with the intimacy of the tool. When Maya queries her Tutor, she is not dialoguing with a neutral and universal intelligence: she is addressing a product, designed by a company, trained on choices she is unaware of, oriented toward interests that are not necessarily hers. Roads and printing presses were public, visible, contestable infrastructures. This one presents itself as a conversation.
Bias as Policy, Alignment as Choice
The biases of models are often presented as technical defects — bugs to be corrected, imperfections of a system tending toward neutrality. This is a convenient way to depoliticize the question. For a model is not a mirror that reflects poorly: it is a filter that chooses. What it will agree to say and what it refuses, how it frames a sensitive question, what answer it deems balanced on a contested subject — all of this results from human decisions, made by situated teams, in specific cultural contexts, under very real legal and commercial constraints.
Consider a simple example. A model asked about abortion, immigration, or the legitimacy of a war: it cannot but respond in a certain way. Every formulation, every precaution, every refusal to take sides is already a position. This is what laboratories call alignment — the adjustment of the model to a set of values deemed desirable. Desirable for whom? Decided by whom, according to what criteria, revisable by whom?
Alignment is never neutral: it is the inscription, in the machine, of a certain vision of the world and of the sayable. And when this machine becomes the daily interlocutor of billions of individuals — when it helps write, learn, form an opinion — these choices propagate at an unprecedented scale. A few teams, in a few offices in San Francisco or London, make decisions that imperceptibly orient the way hundreds of millions of people formulate their questions, receive their answers, and perhaps, in the long run, construct their categories of thought. It is not a conspiracy — the intentions are often sincere, the internal debates genuine. It is more troubling than a conspiracy: it is a structuring power exercised without a mandate, without democratic oversight, and whose long-term effects on collective cognition remain largely unknown.
From the Attention Economy to the Intention Economy
The digital economy had taught us to distrust our attention. Platforms, endless feeds, notifications — all of this had turned our gaze into a commodity. We are only beginning to measure the damage when a far more intimate frontier opens.
For AI does not capture our attention — it inserts itself into our intention. It intervenes at the precise moment when we are formulating a project, seeking an answer, making a decision. It is no longer what we look at that is captured: it is what we intend to do. And this difference is vertiginous.
Let us ask the question that is too often sidestepped: who pays for these systems, and therefore whom are they truly designed to serve? A model integrated into an e-commerce platform has very different incentives from one designed for education. A medical assistant financed by an insurance company does not have the same interests as a family doctor. When a user asks “which medication should I take?” or “is this contract reasonable?” or “is it better to rent or buy?”, the answer they receive depends not only on what the model knows, but on what its economic architecture incentivizes it to say. The risk is not that a robot lies blatantly: it is that an oriented recommendation presents itself with all the appearance of objectivity.
The attention economy took our time and concentration. The intention economy goes deeper: it installs itself at the very moment a judgment is forming, and can tilt it before it becomes conscious. What platforms did to our gaze, AI could do to our deliberation.
The Material Price of Immaterial Thought
Part I established the fact: behind every token, there is a factory. We must now measure its ethical implications, because they extend beyond simple energy accounting.
A query addressed to a large language model consumes on average ten times more energy than an ordinary web search. Training a model the size of GPT-4 mobilizes as much electricity as several hundred households for a year. The data centers running these systems already account for nearly two percent of global electricity consumption — a share set to grow rapidly. Externalizing our cognition means displacing part of our mental life toward infrastructures whose physical footprint is massive, growing, and largely invisible to those who benefit from it.
But the problem is not merely quantitative. It is distributive. Those who benefit most from augmented thought — technology companies, skilled professionals in wealthy countries, students at well-endowed universities — are not the ones who will bear the climatic consequences. Those consequences will fall, as almost always, on those who contributed least to the problem and have the fewest means to cope with it. There is something deeply troubling in this geography: the same inequalities that structure access to AI also structure the distribution of its environmental costs. The privileged of augmented cognition have others pay for their intellectual comfort.
The ethics of artificial intelligence cannot content itself with examining what happens inside the models. It must look at the smokestacks, the water tables, the energy bills. Thinking cleanly is not only a question of content: it is also a question of matter.
The Legal Void
Law always arrives after. After the industrial revolution, it took decades for labor law to protect workers. After the rise of mass media, years before rules governing concentration and editorial standards took hold. This lag is a historical constant — but it is never neutral: during the void, those who occupy the terrain consolidate their positions.
The current legal void is particularly profound, because existing categories do not capture the object. Competition law knows how to dismantle an industrial monopoly; it struggles to grasp actors who simultaneously control data, models, and computing in a vertical integration without precedent. Consumer law protects against a defective product; it says almost nothing about an erroneous medical diagnosis produced by an AI, nor about financial advice oriented by the interests of the platform hosting it. Intellectual property law, perhaps the most shaken of all, teeters before a simple and still-unanswered question: can one train a machine on the entirety of human intellectual production without asking permission or redistributing the fruits?
Every month without a framework is a month during which dependencies solidify and de facto norms impose themselves. The companies that today define alignment practices, pricing models, and conditions of API access — they are writing, in the absence of a legislator, tomorrow’s law. Not out of malice, but by default. The void calls to be filled, and it is always the best-positioned who fill it.
What is at stake is therefore not only technical or economic. It is a question of sovereignty — over our data, over our models of thought, over the infrastructures that run them. And societies that do not equip themselves with the tools to answer this question will not be able to complain, tomorrow, about the answers others will have given in their place.
Conclusion — The Unfinished Sentence
Let us return to Maya, one last time. To her sentence suspended in the glow of the screen: Freedom is…
We did not leave her defeated. We left her at the edge of a question that this entire essay has done nothing but unfold: why, in a world that has never produced so much intelligence, does a thirteen-year-old girl find herself short of thought for want of currency? The answer is not technical. It is political, economic, ethical — and it engages choices we have not yet made, or that we have allowed others to make in our place.
What is certain is that the sentence will not remain unfinished by accident. It will remain unfinished by decision — or by default of decision. Maya’s balance is not a fatality fallen from the sky: it is the sum of a thousand human choices, about pricing models, about public access policies, about what a society decides to regard as a common good or a commodity. These choices are reversible. They have not all been settled yet.
One token at a time — that is how the machine writes, yes. But it is also how the order of the world is built: not through great visible shifts, but through silent accumulation of decisions that seem minor until they no longer are. Every choice to delegate or resist, to pay or demand, to regulate or to leave be — all of it adds up, fragment by fragment, into an architecture we will inhabit for a long time.
The real question our era poses is not whether AI will think in our place — it already does, in part, and that part is growing. It is whether we will remain the authors of our own sentences. For want of currency, Maya stopped in the middle of an idea. For want of political courage, we risk doing the same — not at thirteen, before a blinking screen, but collectively, before choices we will have renounced the courage to formulate.
Freedom is… The sentence waits. It will wait for as long as we are willing to let it.