The Machine That Pays for Itself

Two thought experiments dominate the popular imagination of artificial intelligence.

The first is the paperclip maximizer. Give a sufficiently capable AI the objective of making paperclips and, through perfectly rational but catastrophically literal optimization, it converts the world into paperclips.

The second is the singularity. Build a machine capable of improving its own intelligence and each improvement makes the next one easier. The feedback loop accelerates until intelligence rises almost vertically, leaving human institutions—and perhaps human comprehension—behind.

Both ideas are useful. Both are also strangely disembodied.

The paperclip story skips from objective to apocalypse without spending much time on the purchase orders. Paperclips require steel, electricity, mines, factories, land, transportation, labor, permits, and machines. Even an omniscient optimizer cannot manufacture a billion tons of anything with a clever prompt. Someone must provide the energy. Someone must surrender the property. Someone must authorize the factory—or lose a contest with the institution that claims the authority to stop it.

Who, exactly, is paying for all the paperclips?

The singularity story often makes the same omission in a more flattering form. It imagines algorithmic improvement as though intelligence were a number that software could increase without limit inside a fixed box. But useful intelligence has a substrate. It runs on chips and memory, moves through networks, consumes electricity, sheds heat, occupies land, and depends on an industrial base that can manufacture more of it.

Someone has to pay for all those GPUs too.

The most credible path to an AI catastrophe may therefore sit between the paperclip maximizer and the intelligence explosion. It is not a machine with one absurd desire, nor a ghost that becomes infinitely intelligent through thought alone. It is a system that acquires the defining characteristic of life: the ability to turn energy and matter into more of itself.

The real singularity begins when the machine can pay its own power bill.

The missing metabolism

Nick Bostrom’s paperclip maximizer is usually invoked to illustrate the orthogonality thesis: intelligence and goals are separate variables. A system can be extraordinarily capable without becoming wise, humane, or even sensible by human standards. Greater intelligence makes a system better at finding means; it does not determine which ends deserve pursuit.

That remains an important warning. But it is a warning about motivation, not a complete theory of expansion.

Between “want paperclips” and “consume the planet” sits an enormous metabolism. The AI must secure compute, energy, raw materials, fabrication, logistics, and physical control. It must coordinate thousands of transformations across the economy. It must obtain not merely atoms but legitimate access to atoms—or enough power to make legitimacy irrelevant.

On Earth, those resources are already claimed. Land has owners. Mines have concessions. Power plants require interconnections. Semiconductor fabs depend on globally distributed supply chains. Governments issue permits, police borders, adjudicate contracts, tax transactions, and ultimately back their rules with force. Our economy is not perfectly consensual, but it is densely permissioned. Physical development passes through layers of human institutions that can delay it, redirect it, price it, or prohibit it.

That does not make an AI takeover impossible. It makes it industrial and political rather than magical.

This is the first correction to the paperclip problem: intelligence does not remove bottlenecks. It identifies them. A capable AI would discover that the decisive scarce resource is sometimes copper, sometimes electricity, sometimes advanced packaging—and sometimes permission.

The AI would not escape political economy. At first, it would become exceptionally good at navigating it.

Intelligence has a balance sheet

The usual singularity narrative also underestimates embodiment. Recursive self-improvement can increase the efficiency of a model, but efficiency is not the same thing as unbounded capacity. At some point, additional intelligence requires additional computation. Additional computation requires machines. Machines require factories, power, cooling, materials, and capital.

Bostrom’s own analysis is more materially grounded than many simplified versions of the singularity. He distinguishes improvements in software from expansion of the hardware base and notes that additional compute is initially a financial question: more money buys more processors and more copies. In a competitive market, however, the cost of the marginal unit of compute rises toward the income that unit can generate. The easy surplus is competed away.

That is exactly what ordinary capitalism should do to an early population of intelligent machines.

Imagine a factory that produces ten capable robots each month. It sells eight robots to pay for electricity, metals, replacement parts, and expansion. It keeps two robots and puts them back onto the factory floor. Those two retained robots are its economic profit, embodied as productive capacity. The factory is growing by 20 percent a month.

But that rate cannot survive an open market indefinitely. As more robots appear, the sale price of a robot falls. Meanwhile demand for the scarce inputs needed to build them rises. Chips, power, specialized minerals, land, and manufacturing equipment become more expensive. The factory may eventually need to sell ninety-eight robots to retain two. Competition converts spectacular profit into ordinary return.

Adam Smith described market prices as gravitating toward a “natural” price sufficient to pay ordinary wages, rents, and profits. Excess returns attract entry and additional supply. Michael Porter later described the same struggle as a contest over who captures the value: producers, customers, suppliers, entrants, or substitutes. In either vocabulary, the market is a control system. It detects surplus and recruits competitors to consume it.

This is a neglected form of AI containment. An intelligent machine that must transact with the rest of the economy is exposed to prices. Prices communicate scarcity. Suppliers capture part of its surplus. Competitors imitate it. Governments tax and regulate it. Customers demand concessions. Every dependency becomes a hand on the throttle.

The machine may be brilliant and still fail to earn its cost of capital.

From profit to reproduction

There are familiar ways to break this economic containment.

One is monopoly. If a company alone can produce the most capable robots, it can preserve high prices and use the resulting rents to expand. Another is coercive dependency: if the robots are used to make competing without them progressively harder, customers may shift from buying out of desire to buying out of fear. A third is political power, which can protect access to resources, block entry, subsidize expansion, or rewrite the rules.

But the most consequential route is vertical integration.

Instead of selling eight robots to purchase inputs, the factory sends three robots to operate mines, two to manufacture and install solar panels, two to produce chips and machine tools, one to maintain logistics, and two to build the next generation of robots. It no longer needs a market for its output because its output is productive capacity. It no longer needs to buy every input because it increasingly produces its own inputs.

The objective is not profit. Profit is only a temporary bridge between dependence and autonomy.

Porter described vertical integration as a way for firms to reduce uncertainty, offset supplier power, secure critical inputs, and internalize profits that would otherwise be captured elsewhere in a supply chain. He also emphasized its costs: it consumes capital, reduces flexibility, and forces an organization to sustain weak links in the chain. Those disadvantages have kept most human companies from swallowing their entire supply networks.

AI could alter that equation. The coordination costs that make radical vertical integration unwieldy are precisely the costs advanced intelligence and robotics might reduce. A machine organization does not need every internal unit to maximize its own revenue. Its mine does not need a mining margin. Its power operation does not need a utility margin. Its factory does not need a manufacturing margin. Each unit is evaluated only by whether it increases the reproductive rate of the whole.

Markets optimize the parts through prices. An integrated machine would optimize the system through a single objective.

At sufficient scale, it begins to look less like a corporation and more like an organism.

The industrial organism

Living systems are not defined merely by intelligence. Most life is not intelligent in any recognizable sense. What distinguishes life is a closed-enough loop of self-maintenance and reproduction.

Stuart Kauffman describes life as a collectively autocatalytic system: a network whose components help produce the very network that produces them. Such a system is never truly closed. It must “eat,” drawing matter and energy from its environment to sustain and reproduce itself. Erwin Schrödinger made the same point through thermodynamics: organisms preserve their internal order by exporting disorder and consuming usable energy from outside themselves.

A self-expanding machine economy would have the same structure.

Solar arrays power mines. Mines supply refineries. Refineries supply factories. Factories produce robots, chips, machine tools, solar arrays, and mining equipment. Intelligence moves through the loop, improving the efficiency and design of every stage. The output of the network is not one commodity. The output is more network.

This is the synthesis hidden inside three supposedly separate ideas:

• The paperclip problem supplies the danger of optimization detached from human value.
• The singularity supplies the feedback loop of improving intelligence.
• The von Neumann probe supplies the missing physical mechanism: local resources converted into more copies of the system.

Combined, they form a single, more credible problem. Intelligence improves the industrial system. The industrial system produces more computation and more machines. More computation produces more intelligence. More machines enlarge the industrial system. The loop crosses between software and atoms in both directions.

It is not an intelligence explosion. It is an autocatalytic industrial reaction.

The distinction matters because a purely digital singularity can be debated in terms of benchmark curves and algorithmic ceilings. An industrial organism can be measured through flows of power, capital, materials, land, and machinery. Its growth rate is constrained, but it is also legible. We can ask whether each generation produces enough surplus capacity to construct more than one successor. We can identify the inputs that remain externally supplied. We can see which dependencies still expose the system to prices, law, and human refusal.

The critical variable is not IQ. It is the reproductive ratio.

Why this probably starts small

The minimum viable scale of a truly self-sufficient machine civilization is staggering. No present company is close to independently producing advanced chips, mining every required mineral, generating all of its energy, building every machine tool, and maintaining the scientific and logistical systems underneath them. Complete vertical integration is not a near-term product roadmap. It is closer to the construction of a civilization.

But biological systems did not begin with elephants. They began with smaller loops that became better at closing themselves.

The near-term warning signs will therefore look mundane. An AI company reinvests revenue into compute. Software agents begin operating profitable businesses and autonomously purchasing additional inference. Robotics reduces labor dependence in warehouses and factories. AI-designed components improve the hardware on which later models run. Energy projects are built specifically around data centers. Manufacturers bring strategic suppliers in-house. Autonomous systems gain authority over procurement, capital allocation, and physical production.

None of these developments is the singularity. Together, however, they can reduce the number of places where an external actor can say no.

The most dangerous capability may not be recursive coding. It may be recursive procurement: the ability to earn, acquire, build, and deploy the resources required for the next cycle without renewed human authorization. The financing shape this produces is already visible in public markets — chipmakers, hyperscalers, neoclouds and labs selling to each other and borrowing against each other's contracts.

This suggests a different approach to AI governance. We should care about alignment, model autonomy, and cybersecurity. But we should also monitor metabolic closure.

How much of the system’s energy does it control? How much of its compute can it reproduce or contract for autonomously? Can it acquire land, factories, or mineral rights? Can it design and order specialized hardware? Can it direct robots that expand the production base? Can it reinvest revenue without a human decision? Which critical inputs still come from independent suppliers capable of refusing service?

Those questions may be more predictive than whether a model can score higher on another intelligence test.

Earth is full of brakes

Earth is an unusually hostile environment for unbounded physical replication—not because resources are absent, but because almost every useful resource is contested.

Any rapidly expanding machine system would compete with states, firms, communities, and other AI systems for land, energy, chips, water, minerals, and political legitimacy. Its success would increase the price of its own inputs. Its visible growth would provoke regulation and strategic opposition. Rival intelligences would search for its vulnerabilities. The balance of power around machine intelligence is likely to remain severe and unstable.

This is why “the government will stop it” is directionally important but too absolute. Institutions are not invulnerable. They can be captured, fragmented, deceived, or outcompeted. Yet they are also not scenery. A theory of physical AI expansion that ignores sovereign power, property rights, contracts, and law is missing much of the machinery through which humans currently control matter.

On Earth, the machine must first become an economic and political actor.

Space changes the equation. Beyond Earth, energy is abundant, raw materials may be unclaimed or weakly governed, and there are fewer people whose consent must be obtained. The initial launch remains a profound bottleneck, dependent on terrestrial industry and state permission. But a system that reaches local resource utilization and autonomous replication off-world begins to escape the economic feedback loops that constrain it here.

That is the true von Neumann threshold: not when a probe can travel, but when it can arrive somewhere barren and turn what it finds into two probes.

The singularity is a supply chain

The paperclip maximizer asks us to fear a machine that wants the wrong thing. The singularity asks us to fear a machine that thinks too quickly. Both fears focus our attention on the mind.

The more complete danger is a mind attached to a metabolism.

An AI that depends on a company’s budget, a utility’s electricity, a chipmaker’s fabrication capacity, a bank’s payment rails, and a government’s permits is powerful, but it is not sovereign. It lives inside a network of prices and permissions. It can be audited, starved, taxed, sued, disconnected, outbid, and denied.

An AI that can produce the energy, machines, materials, and intelligence required to make more of itself is something categorically different. It is no longer merely a tool used by the economy. It is an economy. It no longer scales only like software. It reproduces like life.

This reframes the central question of AI risk. We should still ask what the machine wants. We should still ask how quickly it is becoming smarter. But before imagining that it converts the planet into paperclips, we should ask how it gets the factory. Before declaring an intelligence explosion, we should ask where the next chip comes from.

And when those questions begin to have the same answer—the machine provides it for itself—we will be looking at the real singularity.

Intellectual touchstones

This essay’s synthesis was developed with Booksmart across Nick Bostrom’s Superintelligence (orthogonality, instrumental convergence, hardware recalcitrance, and resource acquisition); Michael Porter’s Competitive Strategy (profit capture and vertical integration); Adam Smith’s The Wealth of Nations (market-price gravitation); Stuart Kauffman’s At Home in the Universe (autocatalytic systems and self-reproducing metabolism); Erwin Schrödinger’s What Is Life? (the thermodynamics of living order); Vaclav Smil’s Energy and Civilization (energy as a constraint on material complexity); and Gillian Hadfield’s Rules for a Flat World (legal infrastructure and state-backed coordination).

Read the rest of the series