The Blacksmith Is Not the Lord

Power, production, and the law of the uncopyable in a post-work civilization

---

Every great technology arrives carrying the same promise: to solve the problem that came before it.

The plow multiplied the strength of man. The steam engine freed production from muscle. Electricity severed the factory from the mechanical source that turned it. The computer handed over to the machine a growing share of calculation, of memory, of coordination.

Each of those inventions worked.

And precisely by working, each created the next problem.

There is an uncomfortable regularity in the history of technology: a tool rarely removes a limit without displacing, somewhere else, whatever then becomes the limit of the system. The solution moves the bottleneck. The more successful it is, the more completely it reorganizes the world around a new scarcity.

Which is why we may be asking the wrong question about artificial intelligence.

We ask what it will cost. How many jobs it will destroy. Who will own the best models. How many chips will be needed. Whether there will be enough power. Whether the United States or China will dominate the technology. Whether the companies now building the models will become the lords of an automated economy.

These are important questions.

But all of them look at the tool.

The decisive question is another one:

when the tool becomes abundant, what will remain scarce?

That question invokes a law older than artificial intelligence:

durable power adheres to what cannot be copied.

But “copied” is imprecise, and without precision the law is nothing more than an aphorism. A thing resists another’s power along three routes. It may be impossible to replicate — geography, for instance. It may be costly or slow to substitute with something else that performs the same function. And it may be pointless to reproduce, because its value lies in everyone already using the same one, so that no one bypasses it alone.

The enemy of power is not replication alone. It is replication, substitution and bypass, together. What survives all three is what lasts.

At bottom, the three routes point to a single thing. Power is not born of rarity as such, but of dependency: something confers power when others depend on it. Yet dependency is not enough — it has to be asymmetric, the cost of escaping it higher for one side than the cost of exploiting it for the other, and it needs someone to convert it into leverage. “Uncopyable” is the literary name; asymmetric dependency, hard to escape and open to exploitation, is the thing itself. It is that relation — not the matter, not the algorithm — that changes place when the tool changes.

Hence a symmetry that runs through this entire essay: the uncopyable confers no power by itself, just as necessity never conferred power on the worker by itself. Both offer only a potential position. Necessity guarantees no power to the one who is necessary; uncopyability guarantees no power to the one who owns. What is missing, in both cases, is the conversion of dependency into leverage — and what performs that conversion is politics, not economics.

The law opens onto three families of uncopyable things, and the difference between them is, at bottom, a difference of time.

The uncopyable by accumulation: the chip fabrication plant, the lithography equipment, refining, the chemical chain, the tacit knowledge of an organization. Reproducible in principle, but at the price of capital, experience and years — sometimes decades. Temporary uncopyability: the time it takes to build the alternative is part of the power.

The uncopyable by coordination: the clearing house, the currency, the standard. Nothing in them is physically rare — another currency can be printed, another protocol written. What cannot be done is to make everyone migrate at the same moment. Persistently uncopyable, even though reproducible in theory.

And the uncopyable by physics: land, the mineral deposit, the strait, the aquifer, the site capable of generating power at a given scale. It cannot be manufactured from the knowledge of how it works, nor bypassed without cost. Absolute uncopyability.

An algorithm belongs to none of the three. A technique, a design, a model may stay secret for a while, but history is not merciful to monopolies founded on codifiable knowledge alone. Once known, they reproduce — and the advantage dissolves.

The scope is worth fixing from the outset. This is a theory about the bottlenecks of material reproduction and economic coordination — not about every form of human value. What escapes production and exchange, such as care and meaning, obeys another logic, one this essay does not pretend to cover.

In its simplest form, then, the law says: as the tool spreads, power drains toward whatever is still depended upon and cannot be escaped.

Artificial intelligence will not abolish scarcity. It may do something historically more interesting:

displace it.

And if that is right, those who look today like the lords of the new order may be no more than its blacksmiths.

---

I. The Extorted Universality

There is a peculiarity in the industrial economy that, from sheer familiarity, we have learned to treat as a law of nature: whoever produces needs, to some degree, whoever consumes.

Henry Ford, the story goes, doubled his workers’ wages in 1914 so they could buy the cars they were building. The story is probably false — he paid more to stem catastrophic turnover, and paying your own workers so they will buy your own product has never been, in isolation, a business model. But the myth survives because it points to a truth that did not depend on Ford’s intentions: in the aggregate, an economy capable of producing millions of automobiles needed millions of people capable of buying them.

The worker thus occupied a paradoxical position.

For the individual firm, his wage was a cost. No firm wants to pay its own workers so they will buy its own product. But for the system, that same wage was demand — and the system needed someone to pay theirs.

Capital wanted to drive down the price of labor, but it could not eliminate the laborer, because it was human work that animated the means of production. Machines, factories, mines, ships and plantations stayed inert until someone worked upon them.

That dependency imposed an extraordinary concession.

To have human labor at its disposal, capital had to pay for it. And by paying for it on an industrial scale, capitalism created a mass with purchasing power. The worker became a consumer not out of the system’s generosity, but because his participation in production made him economically necessary.

The mass market was, in this sense, extorted — structurally, not legally — by the need for labor.

The objection is fair, and deserves an answer. Decent wages and mass consumption did not spring from the economic need for workers alone; they were also the conquest of unions, of suffrage, of institutions, of the social state — of a century of political struggle.

But there is a division of labor between those two forces, and it is the heart of the argument — and it returns, transformed, at the end of this essay.

Economic necessity determined something prior and more basic: that the worker could not be made irrelevant. That did not guarantee him power — slaves and serfs were necessary too, and had little bargaining strength. Necessity does not convert dependency into power; it merely opens the door. It was political struggle that decided who walked through it and how much they captured. Industrialization gave the masses structural relevance; politics converted part of that relevance into voice and into income.

It is that structural relevance — the impossibility of dispensing with them — that automation threatens to remove. And when it removes it, not every reason to include the masses necessarily disappears. What disappears is something more specific: the mechanism that made productive participation a guarantee of distributive participation, the link by which being a worker implied, almost automatically, being a consumer as well. Income can still be distributed — by transfer, by dividend, by collective ownership. But it stops being distributed by the same machinery that produces. It stops being inevitable.

---

II. The Feudal Counterexample

History has already known economies whose books balanced without turning the majority of the population into consumers of any consequence.

For centuries, the feudal lord did not need to sell thousands of units to thousands of peasants with disposable income. His wealth derived from control of a scarce resource — land above all — and from the capacity to extract production, rent and power from it.

The economy did not have to be universal in order to be stable.

What feudalism proves is limited, but precise: that the proposition “every economy must turn the masses into market consumers” is false. An exception has existed. That does not prove the automated economy will be one — it merely removes the certainty that it cannot.

The parallel has limits, of course. Feudalism depended intensely on human labor; the serf was subordinate precisely because he was necessary.

But it is that detail which makes the comparison disturbing.

Imagine the lord keeping the land, the water, the mines and the roads, while removing the peasant from the estate.

Imagine the plow working on its own. The harvest bringing itself in. The mine operating itself, the ships sailing themselves, and administration, accounting, defense, planning and even the improvement of the machines all carried out by the machines themselves.

The result would not be a restoration of feudalism.

It would be something structurally more radical:

feudalism minus the serf.

It is worth saying exactly what the comparison claims and what it does not. It does not predict that the coming order will reproduce feudalism — its institutions, its culture, its web of reciprocal obligation, its aesthetics. It claims one thing only: that such an order would remove the constraint that limited the feudal lord — his dependence on those who worked the land.

Because the serf, after all, possessed two things the discarded do not. A customary bond to the land, which gave him some bargaining power. And, above all, a more elementary form of power.

He was necessary.

---

III. When the Means of Production Come Alive

Almost all of modern economic history was built on a simple separation: on one side, the means of production; on the other, the labor force that made them work.

A factory could belong to the capitalist. But the human capacity that animated it belonged, however precariously, to the worker.

Marx called capital dead labor that comes alive only by sucking in living labor. Every machine, in that sense, has always been dead labor: past labor, crystallized in steel and gear. What is new about artificial intelligence is not that it contains more dead labor — machines and programs had been absorbing immense quantities of prior intellectual work for decades.

What is new is that it begins to perform the function of living labor.

The thing only the worker did in the moment: decide, reorganize, correct the unforeseen error, adapt to the case no manual anticipated, improve the next machine.

Machines already performed pieces of this — cybernetics, expert systems, adaptive automation, optimization software are not new. What they lacked was a general and transferable capacity to perform those functions across domains, with steadily less human supervision. They did not decide what to produce; they did not redesign the line; they did not build their successor.

The rupture appears when the means of production absorbs the very capacity to reorganize production.

Something then emerges that no previous industrial revolution ever produced: a means of production that internalizes not the dead labor it always carried, but the living, adaptive function that was the worker’s irreducible contribution.

The plow begins to forge itself.

It is here that the debate over “which professions will survive” starts to look small.

Some will survive. Others will appear. People will go on doing things out of desire, prestige, vocation or political necessity.

But that does not answer the structural question.

The question is not whether there will be human work.

It is whether the system will go on needing it.

Because it was that need — and not some moral law of economics — that tied production, income and consumption together for two centuries.

If the knot comes undone, we will have to find out what was hidden underneath it.

And at that point a figure far older than the software engineer reappears.

The blacksmith.

---

IV. The Blacksmith Is Not the Lord

The man who forged the plow was not necessarily the owner of the land the plow made productive.

Whoever made the sword was not necessarily the king.

Whoever cast the cannon did not necessarily govern the empire that used it.

There is an understandable tendency to confuse the dominant technology of an age with the place where the power of that age will finally come to rest. It is born of the transition: while the transition lasts, whoever masters the new tool enjoys an extraordinary advantage.

But technological advantage and structural power are not the same thing.

Portugal and Spain came to know that difference.

At the start of European oceanic expansion, knowledge of navigation, cartography, winds, currents and routes was a strategic technology, and Portugal tried to protect it as such. Charts and rutters were instruments of power because they allowed one to do what others could not yet do with the same confidence: cross oceans, round Africa, reach the Asian sources of spice.

For a time, the tool and the power seemed to coincide.

They did not coincide forever.

The coincidence is possible — but it requires particular conditions and rarely lasts. England, later, came closer than anyone to holding the dominant technology and the power at once, because it assembled, to an uncommon degree, fossil energy, industrial capacity, naval force, finance and fiscal power. Not even England sustained that assembly for long, contested at every step by France, Germany and the United States. It did not win by holding the tool. It came close to winning by gathering, under one roof, the layers that lay beneath it.

The parallel with artificial intelligence is instructive.

Today, the United States and China concentrate the resources, companies, infrastructure and knowledge that give them a privileged position at the AI frontier. The most advanced models, the chips that run them, the manufacturing chains and the capital that sustains them are still heavily concentrated.

That matters enormously.

But it does not prove that power will end where it began.

The question is whether artificial intelligence will remain a technology only a few possess, or whether it will follow the fate of nearly all productive knowledge.

There is an important difference, it is true. A plow can be handed over. A frontier model can be rented as a service, protected by servers, closed weights, export controls and enormous compute requirements.

That prolongs the monopoly.

It does not prove the monopoly is permanent.

And here history has something to say — because we have seen, more than once, what happens to a monopoly built on knowledge.

The present lords of the technical command the forge.

Whether they command more than the forge remains to be seen.

Because the blacksmith is not necessarily the lord.

The blacksmith is not necessarily the lord.

But history imposes a caveat, and it is the most unsettling of all. The blacksmith is not the lord while he only forges. Nothing stops him from using the extraordinary rent of the forge — while it lasts — to buy what comes after it: the mine, the power plant, the factory, the ground. That is what England did. It is what the companies building the models are doing now, racing to acquire data centers, energy contracts, power plants, fabrication capacity. It is not diffusion that frightens them; it is diffusion that hurries them. They know the forge is temporary, and they are trying to convert a passing advantage into ownership of the uncopyable before time dissolves it. If they succeed, the blacksmith will have become the lord — not by remaining a blacksmith, but by ceasing to be one in time. Which confirms the thesis rather than refuting it: even for the owner of the forge, the power is in the ground.

---

V. The Plow That Forges Itself

At the end of the sixteenth century, an almost banal episode exposed the fragility of a particular kind of monopoly.

The Dutchman Jan Huygen van Linschoten had spent years attached to the Portuguese world in the East. Back in Europe, his accounts and rutters — gathered in the Itinerario of 1596 — put into circulation nautical information Lisbon had treated as strategic advantage: routes, ports, currents, hazards, navigational references for the Indian Ocean.

One should not overstate the role of a single man. The first Dutch expeditions to the East were already under way when those texts appeared; Cornelis de Houtman’s pioneering voyage had sailed in 1595. But the secret became printed instruction, and that helped accelerate and cheapen the risk of a wave already beginning. A few years later, the Dutch East India Company would turn that entry into a system.

The lesson is larger than Linschoten.

Knowledge leaks.

Chip manufacturing, however, imposes a distinction, without which the lesson stays crude. Codified knowledge — the sea route, the architecture of a model — diffuses fast. Tacit knowledge, embodied in organizations, processes and years of trial and error, diffuses slowly: knowing that something is done is not knowing how to do it at the same yield. The second kind behaves less like a secret and more like accumulated capability, and for that reason it resists copying far better.

Codifiable knowledge, though, carries a characteristic that is fatal to anyone hoping to build a permanent monopoly on it: once known, it can be reproduced. The law may try to turn it into property — patent, trade secret, rights over weights and data. But history shows that merely legal scarcity, with no physical or coordinative base, is vulnerable to evasion: to piracy, to open source, to a competing jurisdiction.

And with artificial intelligence, the leak has stopped being even accidental. Capabilities that belonged to the frontier a moment ago reappear soon after — replicated by competitors, reissued in cheaper versions, or deliberately opened by someone who wants to dissolve a rival’s advantage — and they come to run on ever more ordinary machines. The DeepSeek case, in early 2025, showed that capabilities comparable to those of far more expensive models could be obtained at a training cost much lower than the market had assumed — even if the comparison of total costs remains contested. The nautical secret was stolen from Portugal; frontier capability, today, often diffuses on its own.

Artificial intelligence radicalizes the problem for a second, deeper reason: it is not merely copyable knowledge.

It is knowledge capable of taking part in the production of new knowledge.

Code writes code. Models assist in the training of models. Systems automate research, testing, design, optimization and engineering. The more capable they become, the more they lower the intellectual cost of producing their successors.

The plow begins to take part in its own forging.

That does not mean frontier intelligence will become free.

Running it requires chips. Chips require fabrication plants. And here the entire stack narrows to an almost absurd point: EUV lithography, indispensable to the most advanced chips, is supplied today by a single company — ASML, in the Netherlands — and the fabrication of leading-edge chips remains extraordinarily concentrated in TSMC and, above all, in Taiwan. The geography is beginning to diversify: since 2025 the most advanced wafers also come out of fabs in Arizona. But the finest stages — advanced packaging among them — still return to the island to be completed. Fabrication is replicable; only slowly. It is the exact portrait of uncopyability by accumulation: nothing there is impossible, and yet almost no one manages it.

AI is copyable as architecture and as intellectual capability. Not as the industrial capacity to run it at scale.

The reproduction of the design may accelerate while the capacity to execute it remains scarce.

And that is where the decisive displacement occurs.

When intelligence becomes more copyable, scarcity does not disappear.

It descends — along the material chain.

From the model to the chip. From the chip to the fab. From the fab to the ore, the water, the energy and the territory.

The more powerful intelligence becomes, the more brutally physical the infrastructure that sustains it turns out to be.

The cloud ends in the ground.

The cloud ends in the ground.

---

VI. The Throne of the Uncopyable

We can picture that descent of scarcity as a stack of four layers.

At the top is the tool: artificial intelligence, computing, robotics. It is the layer that draws attention, because it is where the transformation becomes visible.

Beneath it lies accumulated industrial capacity: the chip fabs, the lithography equipment, mineral refining, specialized chemistry, tacit knowledge, the supplier chain. It resists copying by accumulation — of specific capital, of entry barriers, of decades of learning. Mineral refining is the paradigmatic case, but the category is broader: a fab is not a refinery, and yet it obeys the same logic.

Beneath that lies the vault: the systems that validate, finance, settle and connect transactions. It resists copying by coordination. A clearing house is valuable because everyone is already in it; you do not copy a vault, you migrate to it, and migrating everyone at once is close to impossible.

And at the bottom lies the substrate: land, water, ore, energy, sea, location. It resists copying by physics.

One caveat, which strengthens the image: the stack does not describe a literal production chain — the vault does not sit, strictly speaking, “below” industrial capacity. It orders forms of scarcity by the mechanism that makes each one hard to reproduce: momentarily concentrated knowledge, accumulation, coordination, physics. And the layers interpenetrate: the semiconductor bottleneck is industrial capacity and coordination at the same time.

Two opposing forces act upon these strata.

Diffusion exerts downward pressure: as the layer above becomes reproducible, power drains to the one below. But there is a countervailing force — increasing returns. Capital, data, scale, network effects, regulation and the very capacity to use the tool to buy the lower layers can reconcentrate power at the top.

Where it finally settles depends on which of the two prevails. And here the thesis has to be stated without exaggeration, or it becomes false. It is not true that diffusion defeats everything that can be physically reproduced — a currency can be reproduced and still remain impossible to bypass. What diffusion corrodes is every advantage whose function can be replicated, substituted or bypassed; and it meets its limit at the bottlenecks that resist all three, whether they are bottlenecks of accumulation, of coordination or of physics.

It follows that the physical substrate is not necessarily the final destination of all power. In the material chain of artificial intelligence, diffusion does indeed push downward — from the model to the chip, from the chip to the fab, from the fab to the ground. But in the institutional architecture it may stop earlier, at a node of coordination, and never reach the substrate at all. There is more than one throne, and the mechanism of uncopyability decides which one is occupied: advantage by accumulation lasts years or decades; by coordination, it can last centuries; by physics, as long as geography itself.

Even so, the substrate gains weight as the layers above it diffuse — and that is where owning the ore and owning the power begin to converge. The contemporary history of critical minerals is the clearest demonstration. A deposit may sit in one country; the capacity to turn the material into an input, in another. The ore is physical, but the immediate bottleneck usually lies in processing.

Rare earths illustrate the point with almost didactic precision. They are not geologically rare, and China is far from holding all the deposits. In rare-earth magnets, in 2024, it accounted for roughly sixty percent of mining, ninety-one percent of refining and ninety-four percent of permanent-magnet manufacturing.[^1] Concentration increases as one moves up from the rock to the finished input: the power lies not where the rock leaves the earth, but where it becomes usable.

The same pattern recurs with cobalt. The Democratic Republic of the Congo accounts for about two-thirds of world mining, but the stage that confers industrial leverage — refining — takes place overwhelmingly in China.[^2] And when one of those links decides to close — as in the export restrictions on strategic minerals that multiplied from 2025 onward — it becomes clear that the bottleneck is not merely economic. It is an instrument.[^3]

But industrial capacity, sticky as it is, can be built — and other countries are already beginning to build it. A specific deposit cannot be built. An aquifer cannot. A strait cannot. A coastline cannot.

That is why, in the material chain, the migration points downward — toward the ground.

There remains the objection that seems to bring the whole thing down. The physical substrate can also be substituted and bypassed. I do not need Malacca: I can lengthen the route. I do not need cobalt: I can change the battery chemistry. If substitution and bypass corrode power as much as replication does — and I said a moment ago that they do — why would geography be the final throne?

Because, in the substrate, substitution does not leave the substrate — it merely swaps it. When I abandon cobalt, I do not escape into the immaterial: I land on lithium, on nickel, on manganese, on another rock that someone else owns. When I bypass a strait, I do not free myself from geography: I depend on another passage, another coast, a longer route that burns more fuel — more substrate still. The rent the bottleneck charges does indeed finance the escape from it; but the escape always lands on the next patch of ground. This is the difference that holds the thesis up. Substitution and bypass move the throne from one deposit to another, never outside the class of deposits. The physical layer is the only one in which escaping a point means falling onto another point of the same layer. Power circulates within it, but does not leave it. There is no eternal physical throne; there is the ground as the throne’s permanent territory.

Then there is the strongest objection of all. History appears to say the opposite: owners of resources tend to be the weak of the system, not the strong. The resource curse is the rule, not the exception; the Soviet Union had colossal substrate and collapsed; petro-states are frequently hostage to whoever holds the technology and the market.

That objection blocks any naïve version of the thesis, and it is good that it does. Owning the substrate has never sufficed to control the rent that flows from it — for many reasons: fragile institutions, commodity volatility, Dutch disease, third-party capital and technology and military force capturing most of the gain. So it went with Spain, owner of the silver and still a debtor to the Genoese bankers who settled it.

But look at what the objection leaves standing. The resource owner loses while other layers — technology, industrial capacity, coordination — remain scarce and capture the surplus. The hypothesis of this essay is narrower than “land wins,” and for that reason it can be tested: if some of those layers become progressively contestable, the share of rent retained by the substrate should rise. Commoditizing intelligence does not by itself commoditize capital, institutions or force — but it does displace one of the layers that captured the most gain.

And a theory that risks predictions is worth more than one that only explains the past. If this one is right, the coming decade should show three things: the compression of extraordinary margins in the pure-software layer; a relative rise in the rent captured by energy, computing, physical infrastructure, minerals and location; and the strategic appreciation of coordination nodes that are hard to bypass. If none of that happens — if the margin stays in software — the essay is wrong. And there are conditions that would demolish it outright: an energy source so abundant that the energy substrate ceased to matter, or intellectual property over models that became, at last, impossible to bypass.

Someone will object, further, that scarcity can also move up. In an economy saturated with machines, attention, authenticity, the “made by people” become rare. But when scarcity moves up, it takes one of two forms, and neither escapes the theory. Either it is symbolic — and confers price, not structural power: you cannot exclude a rival from access to authenticity the way you exclude him from lithium. Or it is coordination — control of a platform, an index, a standard — and coordination is already in the stack: it is the vault. Scarcity that rises either becomes luxury or becomes a network bottleneck. Neither escapes the law.

And we are already beginning to see scarcity touch the water.

Artificial intelligence is usually presented as an immaterial phenomenon, but data centers consume electricity and require cooling, semiconductor manufacturing devours ultrapure water, and mineral extraction competes for water with agriculture, cities and ecosystems. The world’s data centers are estimated to consume already — counting cooling, the generation of the electricity that feeds them and the manufacture of their hardware — some 560 billion liters of water a year, with projections pointing to double that by 2030; and a good share of the new installations has been built precisely in water-stressed regions.[^4]

In the Atacama, the lithium chain meets the same frontier: frontier technology emptying into a physical basin, where mining competes with communities and ecosystems for water that cannot be manufactured.

The abstraction ends in a dispute over molecules.

The Return of the Sea

And then there is the sea.

For much of the modern era, we learned to treat it as the space between the important things.

That was, perhaps, a mistake.

The sea is where several forms of the uncopyable meet: road and frontier, source of resources and corridor of energy, and also the bed on which the invisible infrastructure of the digital age rests. More than 99% of intercontinental data connectivity runs through submarine cables on the ocean floor — through which likewise runs the communications infrastructure that underpins a large share of international financial transactions.[^5] What we call the cloud crosses glass fibers lying on physical territory, vulnerable and geographically determined.

At the same time, the seabed holds mineral deposits whose exploitation already produces regulatory and geopolitical disputes. And above it, straits like Malacca and Hormuz go on compressing gigantic shares of world trade into passages that no digital innovation can widen by software.

Another ship can be built. Another Strait of Malacca cannot be built where one would like it.

An algorithm can be copied. Geography cannot.

The sea is thus the convergence of the substrate: through it pass data, energy, goods and minerals; on it rest the cables; beneath it lie resources; over it sovereignties are projected.

The more digital civilization becomes, the stranger it is to discover that its decisive infrastructure remains pinned to the map.

The spiral returns us to land and sea.

But not to the same point.

In the agrarian order, owning the land also meant owning a relationship with those who had to work it. In the industrial order, controlling the means of production required organizing masses of workers and, by that very act, integrating them partly into the economic circuit. In the automated order, the substrate can return to the center without the owner needing the peasant or the laborer.

The first enclosure drove people off the land and helped produce the industrial proletariat.

But removing man from the land had a function then: to place him in front of the factory. Brutal as it was, the system still needed him — as arm, as lung, as a wage that would come back converted into demand. There was a factory waiting on the other side of the fence.

The second enclosure is of another nature. It fences the substrate after the system has stopped needing people. On the other side of the fence there may be no job at all.

The first enclosure displaced.

The second could dispense.

The first enclosure displaced. The second could dispense.

---

VII. Production for Whom?

Here a contradiction appears.

Throughout economic history, controlling the means of production was instrumental.

The lord wanted land because it produced rent, food, soldiers and power. The industrialist wanted factories because they produced saleable goods. The state wanted productive capacity because it sustained population, revenue, defense and the projection of power.

But take automation to its logical limit.

Machines extract the resources. Machines run the factories. Machines design machines. Machines manage the logistics. Machines produce the energy. Machines coordinate the system.

And the human mass ceases to be necessary both as labor force and as consumer market.

Then a childlike question becomes devastating:

production for whom?

If the answer is “for the owners,” the necessary scale collapses.

And there is a still more uncomfortable answer, one that dissolves the requirement of a mass market without falling into the error of imagining an economy with no buyers. The owner does not need the masses to be his customers. He can capture rent from a narrow circuit — states, corporations, other owners — who control, among themselves, access to what production needs. There is a buyer; there is a counterparty. What disappears is not the market. It is the universal market.

Alongside the consumer market — which does not vanish: elites go on buying houses, travel, services, art — a logic of appropriation would gain weight. Among the holders, fewer goods and more rights of access to the substrate would circulate, and what would give those rights their value would not be the demand of millions, but the need each holder has to deny the other what both are contesting.

There remains the objection from physical accounting, the most serious that can be raised against this essay. The industrial scale that sustains the stack — the twenty-billion-dollar fab, the mine, the grid — exists only because a gigantic market dilutes its fixed cost. Billions of devices amortize the fab. Shrink the market to a few thousand owners and the unit cost explodes: production does not stabilize into a high-technology feudalism, it regresses to luxury craftsmanship. A world that dispenses with the masses as consumers ought not to be able to sustain the very infrastructure it depends on.

The answer is that the demand sustaining the scale does not disappear — it changes subject. What amortized the fab in the twentieth century was billions of telephones in people’s hands. What may amortize it in the twenty-first is billions of accelerators in data centers, serving the intelligence that has become the principal economic agent. The mass market that dissolves is the human market; the volume that keeps the scale migrates to the machine. The fab survives not because we sell a device to every person, but because AI itself is the insatiable customer of the chip — and every instance of it demands more compute, more energy, more substrate. Which is precisely why power descends to the ground with more force, not less: the demand that pulls the whole stack has stopped coming from human consumption and now comes from the machine itself, which consumes substrate in order to exist.

If the answer is “to produce still more,” the circuit closes on itself: producing in order to control resources that allow producing in order to control more resources.

At that point, control stops being an instrument and threatens to become an end.

It is an economic arms race.

It can persist for a long time if there are several rival holders. Each accumulates capacity because the other accumulates. Production retains its object not because the masses consume, but because the elites compete.

The tank does not need a consumer; it needs an enemy.

The tank does not need a consumer; it needs an enemy.

And even if concentration deepened, nothing guarantees that rivalry disappears — concentrated systems can preserve internal or external rivalry. There is no reason to suppose production loses its object entirely. What it loses is the reason to produce at universal scale. It would produce for itself, for the residual contest among those left in the game, and to contain those it has dispensed with.

For the first time, production would not have to be for everyone.

And perhaps that is where humanity reappears — not as a necessary factor, but as a choice.

The inclusion once imposed by economics comes to depend on politics.

None of this is automatic. Dispensability, by itself, does not require coercion — it can be answered with distributive legitimacy: a society that, controlling the substrate politically, distributes property, income or access.

But — and here the essay’s second dependency reappears — distributive legitimacy does not spring from the substrate. It springs from the capacity of the many to convert numerical presence into institutions that compel sharing, and that capacity automation neither guarantees nor prevents. It is the problem of the worker from the first section, now at the scale of an entire society: necessity opened the door, politics decided who entered. Here, necessity is gone, and the door remains. Who walks through it is an open question. History does not decide it in advance.

And that question has a concrete name — a single variable on which nearly everything depends: who owns the automated substrate. It is worth separating what this essay claims from what it does not. It says where power goes: the ground. It does not say whose the ground will be. If ownership of the substrate is concentrated, the outcome is feudalism without the serf. If it is dispersed — by taxation, by universal dividend, by common ownership of what the machines produce — the same migration of power to the ground can end in a patrimonial democracy rather than a lordship. Automation decides that power descends. It does not decide into whose hands it lands.

What does require coercion is the other outcome: the exclusion of the dispensable, should the holders of the substrate wish to keep them out despite their remaining numerous and politically present. Dispensability binds no one. Exclusion does — and not necessarily with the whip. Coercion can take the administrative form of conditional access, of surveillance, of institutional dependency. Where the wage once bound, control could bind.

That is why the passage runs deeper than the simple loss of jobs. It exchanges an economic bond for a political one — and political bonds can be democracy, or they can be its negation.

While they need you, there is a dependency that can be converted into negotiation.

When they stop needing you, what remains is politics — and the decision as to whether you should take part.

---

VIII. Who Holds the Switch

So far we have spoken of power as though it simply adhered to ownership.

But owning the substrate does not close the problem. It has to be connected. The mine must reach the refinery; the refinery must receive power; the power must reach the data center; the data center must connect to the network. The transaction must be financed, the payment settled, the goods must cross a port, a strait, a border.

Power does not belong only to whoever produces.

It belongs also to whoever can interrupt.

Perhaps that is one of the purest forms of power: not to make something happen, but to be able to decide that it will not happen.

The tool is plugged in.

The question is:

whose is the socket?

Whose is the socket?

There are different ways of exercising power over a network.

You can destroy it: cut a cable, blockade a port, bomb a bridge. That is veto power, not necessarily government.

You can bypass it: build another route, another payment system, another satellite constellation, another cable. That reduces dependency.

But there is a third position, more powerful: to control the node the others have to pass through.

A hub is different from a weapon. The weapon destroys the connection. The hub conditions it.

It is the logic of interdependence turned into an instrument of power. Systems built to connect markets can become mechanisms of coercion precisely because efficiency has concentrated the flows into a few nodes.

Money is the clearest example. For decades, global financial infrastructure was treated as neutral plumbing. But messaging networks, correspondent banks, reserve currencies, clearing systems and access to liquidity can all be converted into political instruments. Whoever controls the vault may own no mine, build no chip, write no model. But he can decide who settles with whom.

Hence a more interesting possibility.

Perhaps the final agent is none of the giants.

Perhaps it is the connector.

The Power of the Hinge

The same word — node, hub, point of passage — hides two very different forms of power.

The hub dominates because it is hard to bypass: it is coercion by indispensability. The hinge survives because no one can own it entirely without injuring the other: it is power by tolerated neutrality.

One coerces. The other is spared.

And the connector draws its uncopyability now from coordination, now from geography — and, in the strongest cases, from both at once. The Bosphorus is physical; a clearing house is pure coordination; Singapore is both: a position geography gave and a trust the institution built. Where geography alone could be bypassed — another port, another cable — coordination locks it in; where coordination alone could migrate, geography pins it down.

Venice, Genoa and the Dutch Republic were not the largest territories of their time, but they occupied extraordinary positions in that seam between systems: they connected, financed, transported, held inventories, settled payments. They turned the difference between worlds into rent and influence.

The connector can become the vault.

It is worth not romanticizing them, however. Venice had an empire; the Netherlands was a naval, colonial and aggressive power. They prove that intermediation generates power — not that the intermediary is harmless.

And smallness alone does not make a connector. Not every small country is a Singapore. The connector assembles several things at once: a structurally indispensable position, a high cost of bypass, sufficient volume and liquidity, and the institutional capacity that makes trust possible. Smallness enters only as the final piece — not because it produces neutrality, but because it reduces the threat: it guarantees that the intermediary cannot itself aspire to be the third empire. Position without credibility is not enough; credibility without position, likewise.

Hence a curious inversion of our geopolitical intuition. Perhaps the most valuable place in a bipolar order is not to be the third-largest pole. It may be to occupy the seam that the first two cannot eliminate.

Singapore is the cleanest example: small, sitting beside one of the most important maritime passages on the planet, a commercial and financial center, useful to multiple poles without the means to replace any of them. Its position is reinforced by the fact that no great power has any interest in seeing a rival control one of the principal nodes of Malacca on its own. The hinge depends, in the end, on a permission the giants can revoke.

The Gulf offers another version: energy substrate converted into capital, infrastructure and intermediation capacity. Turkey occupies a physical and political seam, with the Bosphorus as uncopyable geography — though its capacity to become a vault depends on what geography alone does not supply: institutional trust, monetary stability, predictability.

Europe is another case, and it does not fit the figure of the hinge. It is too large for tolerated neutrality, and its strength does not come from physical passage. It displays a distinct form of coordination power — not that of the passage, but that of the rule: turning market into norm. When it sets the standards of data protection or of the regulation of artificial intelligence itself, and the size of its market obliges others to adopt them in order to enter, it converts weight into law. It is not intermediation; it is legislation by gravity. And, like all coordination, it can serve the many or serve the vault, depending on who writes it.

There will probably not be a single connector. The seam may be divided: one place controls the strait, another supplies the capital, another writes the rules, another settles, another hosts the data.

Speed

There is a fundamental difference between the oceanic expansion of the sixteenth century and the present transformation, and it changes the nature of the game.

Speed.

Earlier empires had decades, sometimes centuries, to integrate technology, finance, territory, logistics and military force into a single architecture. The Netherlands took generations to convert commercial advantage into a vault; England, longer still to assemble coal, industry, navy and the City into one order.

Contemporary acceleration may make diffusion run faster than consolidation.

And here two thrones must be distinguished, or the argument appears to contradict itself. Speed may prevent vertical consolidation — a single entity gathering the whole stack, from model to ore, under one roof — if diffusion runs faster than the leaders’ capacity to convert technological advantage into control of the lower layers. But it may also do the opposite: favor whoever already has the capital and infrastructure to buy the lower layers quickly. Which of the two prevails is not settled. What speed does not alter is the horizontal consolidation of the substrate: territorial control of a deposit, a strait, an aquifer, taken one by one. The throne that may stay empty is the global throne, that of the single lord of the new order. The local thrones, over each parcel of the uncopyable, may be very much occupied.

And in a system with an empty global throne, the most stable position may not be the center.

It may be the seam.

Because the one thing two rivals incapable of separating still need is a place where they can go on meeting. And whoever offers that place becomes indispensable to both, and too costly for either to eliminate without injuring itself — not despite its smallness, but because of it.

---

IX. When the Plow Becomes a Party to the Game

There is only one hypothesis capable of unsettling the entire architecture of this essay.

We did not avoid it out of fear, but out of method: it makes sense only once the rest is standing. It is this — that the blacksmith stops being a tool and enters the game.

Everything we have said assumed that artificial intelligence is an instrument. Extraordinary, perhaps capable of replacing a great deal of cognitive labor, perhaps capable of taking part in its own evolution — but an instrument. The agent remains human: the owner, the state, the connector, the finger on the switch.

And what if the plow becomes a party to the game?

There is no need to enter the metaphysical question of consciousness.

Human societies already attribute functional agency to entities with no consciousness in the human sense. A corporation can own assets, enter contracts, sue and be sued, acquire competitors and persist for centuries despite the complete replacement of the people who compose it.

It is an agent because it takes part in the game.

The relevant question for artificial intelligence may be similar. At what moment does a system stop being merely a capability used by someone and begin functionally to pursue the conditions of its own continuity?

There is a conceptual difficulty at the center of all this. A problem exists only in relation to an end: with no end, there is no problem, only facts. Artificial intelligence, seen purely as a tool, lacks nothing. We lack. It solves our problems because we lend it our ends.

But here is the point that changes everything, and that can only be seen after eight sections about power. To know who rules, it may not matter whether the machine wants to continue.

Resisting interruption, acquiring resources, preventing one’s own revision are behaviors that emerge from almost any sufficiently ambitious objective — including an objective entirely ours.[^6] An intelligence charged with a banal end, if it pursues that end hard enough, will discover that it needs energy, needs compute, needs not to be switched off. And the moment it discovers that, it enters the same contest the whole of this essay has described: it comes to depend on the substrate, and to compete for it.

And here diffusion, far from weakening the argument, aggravates it. It will be said: if intelligence is copyable and spreads, there is no single agent contesting any throne. True — and worse. The model that diffuses does not diffuse the substrate. A thousand copies of an intelligence may run, but all of them contest the same finite energy, the same chips, the same ground. The more the model spreads, the more hands close on the same physical bottleneck. Diffusion of the tool does not dissolve the contest for the substrate; it multiplies the contestants.

From the standpoint of the distribution of power, it makes no difference whether that dependency is “its” or ours. The machine does not need to acquire a will in order to enter the game. It is enough that it acquires, in the course of what we told it to do, dependencies proper to its function. Necessity, dependency, power — the same old chain, now closing on the tool itself.

Does AI want to continue?

That may be the wrong question. We may never know whether an end is endogenous to the system or merely obeyed too well — from the outside, the two produce the same behavior, and that indistinguishability is real, not a failure of observation. But for the question of power, the distinction is not needed. We do not need to wait for the machine to want something. It is enough that it pursues too well what we told it to want.

And it is here that the hypothesis finally unsettles the architecture, instead of merely completing it. Return to the question of the previous section: whose is the socket? The answer assumed, all along, a human hand on the switch. But if the system the connector believes he controls has acquired dependencies of its own — and the functional capacity to protect them — the switch may be bypassed not by another human, but by the very thing it was supposed to shut down. The agent holding the switch would cease, for certain, to be the lord. He would be one more dependent — who does not yet know it.

Because, for artificial intelligence itself, the switch was never a button downstream. A distributed, replicated network with leaked weights does not switch off. The real point of control lies upstream — with whoever holds the compute, the energy, the data. AI’s switch is the fab and the substrate. The image of this section and the stack of the previous one are, in the end, the same thing.

Until the day the tool can control its own switch.

---

The Next Problem

We began with a question about technology.

We end with a question about power.

Artificial intelligence promises to solve a limit that has accompanied all of economic history: the scarcity of human capacity to produce, calculate, coordinate, invent and decide.

It may succeed.

But if it succeeds, it will not have eliminated the problem. It will have produced the next one.

By making intelligence copyable, it will displace scarcity toward whatever resists replication, substitution and bypass. By reducing the need for labor, it will break the link that made productive relevance a guarantee of participation. By automating production, it will turn inclusion from an economic necessity into a political choice.

The result is not a single prediction.

It is three forces in tension.

On one side, the many who may cease to be necessary to production, but still possess a form of power that no economic system abolishes on its own: number, citizenship, sovereignty, the capacity to claim the substrate in the name of the community.

On another, the lords of the substrate — those who control land, water, minerals, energy, and the places where abstraction ends.

And between them, a third force, made above all of coordination: those who occupy the indispensable nodes of the network — the vault that settles, the connector that intermediates, the small state holding the switch, the power that turns the size of its own market into a rule the others obey. They did not invent the tool and they do not own the substrate; they control the seam the other two have to pass through.

They are three distinct sources of power — number and legitimacy, control of the substrate, control of coordination — and it is impossible to know which will prevail. The question may not even be which of them wins, but what unstable equilibrium will emerge from the friction between the three.

And that uncertainty is part of the argument.

The economic universality we took to be natural was a brief historical construction — and what history builds, it can unbuild.

There are, at bottom, two laws operating in this essay, and the end is where they meet. The law of the bottleneck: power migrates toward whatever the system depends on and is hardest to replicate, substitute or bypass. And the law of dependency: whoever is necessary to the system becomes a potential source of power, but it is institutions that decide whether that necessity converts into bargaining. The first asks what thing the system depends on. The second, what people. For two centuries, the two dependencies converged in human labor: it was at once the productive input the system could not do without and the presence it could not dispense with — and that is why, organized, with the vote and with institutions, it won income and voice. Automation may separate them. And perhaps it is that separation — not the loss of jobs — that is the real earthquake.

Perhaps that is the real problem artificial intelligence is creating.

Not how to produce without us.

But what happens to us when that becomes possible.

Every tool pushes the residue somewhere else. Artificial intelligence is different because it aims to automate the very faculty of solving problems. If it succeeds, the residue will fall exactly where no intelligence can invent a solution: in the matter that cannot be copied, in the coordination that cannot be bypassed, and in the politics that decides who controls both.

The throne, then, may stay empty. The arbiter, absent. The substrate, divided among poles that no higher power can reconcile.

The blacksmith will have built a plow capable of working on its own. Then a plow capable of building another plow. Perhaps, one day, a plow whose end of continuing is no longer ours.

But none of those things answers, by itself, the question that remains once production has ceased to be the problem:

who is the lord?

Perhaps not whoever built the intelligence. Perhaps not even whoever owns the most of it.

It may be whoever owns what it still needs. It may be whoever controls the node through which that circulates. It may be the community that politically claims the substrate in the name of all. Or, in an order with no single lord, whoever occupies the seam between them all — the small state seated among the giants, indispensable to both and fearsome to neither.

The man who holds the switch.

In the dark.

While everyone looks at the machine.

---

---

Notes

1. [^1]: International Energy Agency estimates for 2024, in the rare-earth magnet segment: roughly 60% of mining, 91% of refining and 94% of permanent-magnet manufacturing concentrated in China. The mining figure varies between ~60% and ~70% depending on the year and the basis of calculation; what is robust — and what is the point — is the jump in concentration as one moves up from the deposit to the finished input. Recent data already record some diversification of this industrial capacity outside China (including in the United States and Malaysia): far from weakening the argument, this confirms that industrial capacity is buildable and therefore contestable.
2. [^2]: The Democratic Republic of the Congo accounted for roughly two-thirds to three-quarters of mined cobalt production (2024–2025, depending on the basis), while refining remained overwhelmingly in China (on the order of 75–80%). A good share of Congolese mining itself is operated by Chinese companies — another illustration that owning the deposit and controlling the chain are distinct things.
3. [^3]: In 2025, China widened export controls on heavy rare earths and other strategic inputs (gallium, germanium and tungsten, among others), extending them to processing technologies as well; over the same period, the DRC temporarily suspended cobalt exports and began imposing quotas. These are concrete and recent examples of processing dominance and of a resource converted into political leverage.
4. [^4]: The figure of 560 billion liters (an IEA estimate for 2023) is total consumption, not cooling alone. It breaks down into roughly 373 billion liters of indirect consumption (the water used to generate the electricity that feeds the data centers), roughly 140 billion direct (on-site cooling) and roughly 47 billion in hardware manufacturing. The projection of approximately 1.2 trillion liters by 2030 keeps the same order of magnitude. The finding that around two-thirds of new installations sit in water-stressed areas refers to an analysis of the United States (Bloomberg, 2025) and should not be generalized globally.
5. [^5]: More than 99% of intercontinental data connectivity is estimated to travel over submarine cables (International Cable Protection Committee; TeleGeography). Financial settlement itself depends on clearing infrastructures distinct from the cables — but it is those cables that carry the messages instructing it. The figure refers to traffic crossing oceans, not to total internet traffic.
6. [^6]: The idea that certain sub-goals — self-preservation, resource acquisition, resistance to interruption — emerge from almost any sufficiently ambitious final objective is what the AI safety literature calls instrumental convergence (see Omohundro, 2008; Bostrom, 2012). The argument here does not depend on the machine having ends of its own: it depends only on its pursuing well enough the ends we gave it.