The DecaCorp

1Introduction

Every economic decision in history has been made by one kind of intelligence. Our institutions are built around its properties. Firms, wages, management hierarchies, and equity grants are all devices for getting useful work out of minds that are self-interested, hard to observe, fixed in capacity, and impossible to copy. If a different kind of mind begins to do the work, there is no reason to expect those devices, or the industrial structure they produced, to survive unchanged.

This essay follows one consequence. I argue that artificial intelligence weakens the forces that have limited the size of firms until they no longer bind, and that the result is a small number of firms far larger than anything in the modern record. I call the limiting case the decacorp. The prefix means ten, and the ten counts percentage points of GDP. A reader who guessed ten trillion dollars of market value was thinking too small.

The argument runs in six steps.

1. The cost of compute keeps falling on its historical curve, and the return to compute rises with AI capability, so investment in compute keeps growing.
2. Enough compute yields systems that perform most economically valuable tasks.
3. Those systems are a different kind of labor: selfless, controllable, observable, replicable, and scalable.
4. Labor of that kind weakens the diseconomies of scale that bound the firm from above until they no longer bind, and leaves the economies of scale in place.
5. The firms that win are vertically integrated modelers, which own the loop from data to model to deployment to feedback. Intelligence is general, so the same few firms lead in every industry. Escalating fixed costs and data that rivals cannot copy give them rising returns as they grow, and capital follows.
6. They grow until they meet the state, and the question between them is force.

Steps one and two are premises. Section 2 states them and gives the reasons for holding them. Steps three and four are the core of the essay and occupy Sections 3 to 5. Section 6 asks how large the largest firms become. Section 7 takes up step five, Section 8 tests the argument against earlier giants, and Section 9 takes up step six. Section 10 collects objections.

2Premises

2.1Compute

The first premise has an old part and a new part.

The old part is Moore's law, or its modern equivalent. The number of floating-point operations per second that a dollar buys has doubled about every two years. Hobbhahn and Besiroglu (2022) measure a doubling time of 2.07 years for the GPUs used in machine learning between 2006 and 2021. I assume this continues. It has survived several changes in the underlying technology, and the industry plans around it.

The new part is about demand. Through most of the history of computing, what an extra unit of compute could earn was limited by the software available to run on it. With AI, what it earns is set by the work it can do in place of a person. A processor running a model that can draft an email earns very little. The same processor running a model that can do the job of an engineer earns some fraction of an engineer's salary. As capability rises, each unit of compute stands in for more valuable work, so the return to owning compute rises, and firms that invest until the return meets the cost invest more. More investment buys more compute, which under the second premise buys more capability. The premise is therefore a loop (Figure 1). Growth in spending follows from the return and does not have to be assumed separately.

The evidence so far fits. On reported figures, the combined annualized revenue of the two leading laboratories went from under $10 billion at the start of 2025 to roughly $100 billion by the middle of 2026.¹ Capital spending has followed the revenue. The four largest cloud companies spent roughly $250 billion in 2024 and now guide to about $730 billion for 2026. That is close to a tripling in two years, a doubling time of about 1.3 years.²

Installed compute is the product of the two parts, so it grows faster than either:

where $T_p$ is the doubling time of price-performance and $T_s$ that of spending. With $T_p$ near 2 years and $T_s$ near 1.3, $T$ is about 0.8 years, or nine to ten months. That is close to what is measured. Epoch AI estimates that the world's stock of AI compute has doubled every seven months since 2022, and that additions to it have doubled every nine months since 2024.³

The loop cannot run at this speed indefinitely. Spending cannot outgrow the revenue it earns for long, and at $730 billion the four companies' capital budgets are already more than two percent of GDP. The argument below needs much less than the current pace. It needs installed compute to keep doubling about once a year.

Set that against the combined capacity of human brains, which on this time scale is flat. The installed stock of AI compute is now roughly 30 million H100-equivalents, or about $3 \times 10^{22}$ FLOP/s. Common estimates put a brain between $10^{15}$ and $10^{16}$ FLOP/s, so eight billion of them come to between $8 \times 10^{24}$ and $8 \times 10^{25}$. Any conversion of a brain into FLOP/s is absurd. The conversion also matters very little, because one curve is exponential and the other is not. If the brain estimate is wrong by a factor $\kappa$, the date at which the curves cross moves by

so with annual doubling an error of ten times moves the crossing by 3.3 years. At annual doubling the machine stock passes the low estimate in 2034 and the high one in 2037. At the pace of the last two years it passes them in 2032 and 2035 (Figure 2). In 2024 I put the crossing at 2035, and the 2026 data leave it there. Sometime around then, most of the thinking capacity on the planet stops being human.

2.2Capability

I take from Sutton (2019) the lesson that general methods which scale with computation have beaten methods built on human insight wherever both were tried, and I assume this continues. “AGI” in this essay means only a system that is good at most economically valuable tasks. Nothing below requires reasoning, understanding, or consciousness in any deeper sense. My own confidence that more compute delivers this is 90 to 95 percent. Readers who put it much lower should read the rest as conditional. This is the premise that closes the loop in Figure 1.

The premise says something about the shape of capability as well as its level. Human intelligence is general. Spearman (1904) found that people who do well on one mental test tend to do well on the others, and a single factor, which he called $g$, typically accounts for 40 percent or more of the variation in scores across a varied set of tests (Deary, Penke, and Johnson, 2010). Models show the same pattern, and so far more strongly. Ilić (2023) finds that one factor accounts for 85 percent of the variation in benchmark scores across more than a thousand language models. Emberson and Edelman (2026) find that frontier models are ranked almost as consistently by benchmarks in different categories, such as mathematics, coding, and reasoning, as by benchmarks in the same one: the median rank correlation is 0.68 across categories and 0.79 within them.

The evidence is young and has known weaknesses. Benchmarks overlap, a newer model beats an older one at almost everything, and Krakauer (2026) finds the leading factor's share falling from 92 to 64 percent on one set of benchmarks once models specialized in reasoning arrived. I take the weaker reading, which is enough for what follows. A model that leads at one kind of work usually leads at most kinds. Section 7.2 draws the consequence.

3A Different Kind of Labor

For industrial organization, how capable these systems are matters less than how they differ from human workers as an input to production. Table 1 lists eight differences.

The machine wins seven rows. The eighth is a real concession and Section 10 returns to it. I will call labor with the first seven properties selfless, scalable intelligence, or SSI.

The word “selfless” describes what firms will demand and what developers are trying to build. It is a design goal and I treat it as achieved for the purposes of the argument. Each of the seven properties answers a specific problem in the economics of organization, and the next two sections match them up.

4Why Firms Are the Size They Are

4.1The planned economy we already have

We describe ours as a market economy, yet most production does not happen in markets. Roughly three quarters of GDP is produced inside medium and large enterprises, and inside an enterprise resources move by instruction, without prices. Simon (1991) imagined a visitor from Mars who sees firms as solid green areas and market transactions as thin red lines between them. The visitor would report an economy that is mostly green. Chandler (1977) called the same thing the visible hand. Apple earns on the order of $100 billion a year, which makes it a centrally planned economy the size of a mid-sized country,⁴ and it is widely admired for how well it is run.

So the useful question is where an economy sits on a spectrum defined by the average size of its planning units. At one end the unit is a single person and all coordination is by price. At the other the unit is the nation. Every real economy sits between the two, and its position needs two explanations: why the units are not smaller, and why they are not larger.

4.2The lower bound

If markets are so good at organizing production across the whole economy, why not inside the workshop? Why do firms exist at all? The literature since Coase (1937) gives four reasons.

1. (a)Transaction costs. Marketing, selling, and buying each day's work on a market costs too much. A stable employment relationship saves the cost of finding, evaluating, and negotiating with a contractor every morning.
2. (b)Measurement. Individual output is hard to measure, so it is hard to price. The firm pays for time instead, watches the work, and leans on the social pressure of colleagues (Alchian and Demsetz, 1972).
3. (c)Economies of scale. Many things are cheaper per unit when done in volume.
4. (d)Coordinated investment. Some production needs several expensive, specialized assets and people trained to use them together. Once built and trained, their best use is with each other and nowhere else, and common ownership protects each from being held up by the rest (Williamson, 1985; Grossman and Hart, 1986).

4.3The upper bound

The opposite question is harder. Careful planning beats improvisation in most human affairs. If there are good reasons for firms to be large, why are they not colossal? Why has the best-run firm in the world not absorbed every market and turned the global economy into a planned one? Antitrust law is one answer, though it has been used too rarely to be the main one. Four other forces do most of the work.

1. (e)Complexity. More than 800 firms contribute to a single Nvidia GPU, with supply chains reaching into nearly every industry. No management team can hold all the inputs and tasks of modern production in its head, even for something as simple as a pencil (Read, 1958). The knowledge needed is dispersed among people who cannot fully communicate it (Hayek, 1945).
2. (f)Information. As a firm grows, the people at the top find it harder to know what is happening below. Everyone has a reason to misrepresent their work. Each layer of hierarchy loses some of the signal (Williamson, 1967), and the number of possible links among $n$ people grows as $n(n-1)/2$. The fuzzier the output, as in knowledge work, the worse the problem.
3. (g)Incentives. When individual contributions are hard to measure, it is hard to align an employee's actions with the firm's interest. Why would anyone reveal what is really possible, or cut a cost, with no share in the gain? Equity helps in a small firm. In a firm of $n$ people each person's actions move the value of their stake by something like $1/n$ (Holmström, 1982; Jensen and Meckling, 1976).
4. (h)Politics. As the pie grows, an employee's time is better spent fighting and colluding over shares of it than enlarging it. Milgrom and Roberts (1988) call these influence activities, and the end state is managerial capture. This is a common explanation for why founder-led companies so often outperform: the founder's interests are the firm's.

Force (f) alone would be enough to stop growth. Doubling the number of people roughly quadruples the links among them (Figure 3), and every link is a place where information can be lost or shaded.

4.4A sketch

Write the net value of organizing $n$ units of work inside one firm as

where $B$ collects the gains from internal organization, forces (a) to (d), and $C$ collects the costs of administration, forces (e) to (h). $B$ rises with $n$. $C$ rises faster, since links, layers, and factions all multiply with headcount. The firm grows until

and stops. The observed distribution of firm sizes (Lucas, 1978; Axtell, 2001) reflects where this condition binds for firms of different quality. The point to hold on to is that every term in $C$ is a statement about human beings.

5What Machine Labor Does to the Bounds

Table 2 sets the eight forces against the properties of SSI from Table 1.

Read down the last column. Two of the forces that favored firms over markets are weakened. An AI contractor can be found, evaluated, and audited at almost no cost, so buying work on a market gets cheaper too. If that were the whole story the prediction would be ambiguous, and a world of tiny firms trading with each other would be as likely as a world of giants.

It is not the whole story, because forces (c) and (d) never depended on what labor is like. A training cluster, a robot factory, and a fleet's worth of operating data are worth more together and at scale, whoever does the work. On the other side of the ledger, every force that stopped firms from growing was a fact about people, and each one weakens as people leave production. Complexity yields to a planner whose capacity can be increased at will. Information loss yields to workers whose every action is logged and whose reports are not strategic. Politics needs someone with a private interest to pursue.

Force (g) is the big one. It is hard to overstate what perfect alignment of every worker means. A firm staffed by SSI has direct control over all of its inputs. A management that wants more innovation does not design a bonus scheme and hope. It turns up a setting.

In the terms of equation (3), $C$ falls from a steeply rising curve to something close to the linear cost of compute, while $B$ keeps its scale term. Condition (4) then has no interior solution (Figure 4).

6How Large

6.1The arithmetic of ten percent

The first version of this calculation was done on a napkin in October 2024. Table 3 repeats it with 2026 figures. United States GDP is now about $32 trillion. At two percent real growth it reaches about $42 trillion in 2040, in today's dollars.⁵ That is a conservative base, since an economy that contained a decacorp would presumably have grown faster. Suppose the leading AI and robotics firm accounts for ten percent of it, and take profit as the measure of the firm's contribution.⁶ At thirty times earnings,

and at a more sober twenty times, $84 trillion.

Compare this with the present. Apple, the most valuable company in the world, is worth about $5 trillion, and the seven largest technology companies together about $23 trillion.⁷ A single decacorp at $126 trillion would be worth more than five times all seven. To get there from $5 trillion by 2040, a company has to compound at 26 percent a year for fourteen years.

The comparison in profits is closer than it was. The largest operating profit today is Nvidia's, about $190 billion over the last four quarters, or 0.6 percent of GDP. A decacorp has seventeen times that footprint (Figure 5). In 2024 the largest was Apple's $100 billion, a third of one percent, and the multiple was thirty. The distance to ten percent has almost halved in two years.

6.2Why this is less wild than it sounds

Technology companies already grow as no companies did before them. They sell into global markets, distribute at close to zero marginal cost, enjoy large economies of scale, and are the first to adopt their own advances. The longer history points the same way: each fall in the cost of distribution and communication has enlarged the market one firm can serve.

AI adds something new, which is that it is unusually easy to adopt. Three things slowed every earlier general-purpose technology.

• Diffuse infrastructure. Cities had to be wired and fiber had to be laid. A model is trained in one place and can be used everywhere the same day.
• Acclimation. People had to learn spreadsheets, come to trust buying online, and reorganize their processes. AI meets users where they are and closes the gap in knowledge or interface itself.
• Network effects. A telephone was worthless until one's correspondents had one. AI gives its full value, and a competitive advantage, to the first adopter.

6.3A case: hair salons

There are tens of thousands of hair salons in the United States, 34,582 by one count, and nearly all are small. Urban form explains why each shop is small: buildings are limited in size and people want a salon within walking distance. It does not explain why one firm does not own them all. The theory of Section 4 does. There are minimal economies of scale and no coordinated investments. Customers have differentiated tastes. Above all, cutting hair is a skill that lives in the individual and owes nothing to the firm.

Replace the hairdresser with a robot and each term flips (Figure 6). Economies of scale become large, in manufacturing and in training. Coordinated investment runs through the entire robot supply chain. Differentiated tastes are easily met by a machine that can be customized without limit and that repeats a cut exactly. And the skill now belongs to the firm, since it is whatever the firm's hairdressing model can do.

In the cartoon version of the end state, a machine visits your home, gives you the cut you want in under five minutes at a tenth of the price, and every salon closes. The more likely version has no hairdressing robot at all. The same general-purpose machine that does the laundry, cooks dinner, and fixes the plumbing downloads the hairdressing skill.⁸ Then the event that happened to salons happens to plumbers, restaurants, and laundromats at the same moment, and a million small firms are replaced by one or two robotics companies.

6.4The firm with no obvious limit

Move forward a decade. The leading robotics company has economies of scale in its factories and in its training clusters, and knowledge that carries over from one domain to the next. Its robots are everywhere, and their live feeds are the richest record in existence of how goals are achieved in the physical world. Now ask of this company what Section 4 asked of firms in general. What constrains it? Which markets can it not win? The usual answer, that it would become unmanageable, was answered in Section 5. Nothing in the theory stops it from coordinating a large share of all production.

6.5The order of absorption

Digital goods and services go first, because they are the easiest to train on and deploy at scale. That covers any job that can be done entirely from a laptop: software, media, business operations and analysis, research, parts of biotechnology, and AI research and engineering itself. Intelligence is the only input to these goods, and AI puts the price of intelligence on a Moore's-law curve: a tenth, then a hundredth, then a thousandth. How much more is bought depends on elasticity, which is hard to predict because people find new uses for a cheaper input. Some demand is capped. Human attention is finite, and a company needs only one set of accounts. Research, simulation, design, and customer support have no obvious cap, and it takes only one of them growing a hundredfold for total demand to explode. In every case production moves from people to machines, and the main variable input to the machines is compute.

The physical world comes next. It is messier and far more expensive to deploy into, so absorption starts in constrained environments with large markets. Factories were first and roads are second. General-purpose machines follow. They will look slow and clumsy at first. Six months later they will have trained on billions of simulated hours and will be the best in the world at a given task, and every deployed unit will download that skill on the same day. We are not used to intelligence that scales.

7Which Firms: The Vertically Integrated Modeler

Claim 1 removes the ceiling on the size of the firm, and Claim 2 says that some firm will rise toward where the ceiling was. Neither says which. This section argues that the winners are firms of a particular shape, and that capital follows them.

7.1Scale is not enough

The natural first answer is scale. Training a frontier model is an enormous fixed cost, and a fixed cost spread over more sales is the textbook economy of scale. But a fixed cost of a set size does not keep a market concentrated. As the market grows, more firms can cover the cost, and the number of firms rises. With a fixed cost $F$ and a market of size $S$, the number of firms a simple market supports grows roughly as

Cement works this way. A bigger market has more plants and more producers.

Sutton (1991) showed what changes when the fixed cost is something firms choose. Where spending more up front buys a better product, and a better product takes customers from rivals, a bigger market justifies a bigger outlay. The leaders raise their spending in step with the market, entrants cannot match it, and the number of firms stops rising however large the market becomes. Jet engines, semiconductors, and pharmaceuticals work this way. So do frontier models, where the cost of a training run has risen with the revenue available to pay for it.

This explains why an industry of this kind has few firms at its frontier. It leaves every industry with a few of its own.

7.2Generality

AI adds a property those industries lack. A drug company's research is no help in making cars. Its fixed costs and its lead both stop at the edge of its industry, so each industry had its own escalation and its own few giants. A general model's training cost is spread over every task the model can do. For the purposes of equation (6), the whole economy becomes one market (Figure 7).

Generality also fixes the ranking. If capability were a bundle of unrelated skills, the escalation would run field by field, and the best lawyer, the best chemist, and the best driver could be the models of three different firms. The capability premise says otherwise (Section 2.2). The firm with the most capable model has the best of all three, or close to it, and a specialist that spends everything on one field is overtaken there by the next general model, which is the history Sutton (2019) describes. Rivals stand in much the same order in every market, and a lead in one is a lead in most. The concentration in Figure 7 then happens once, across the whole economy, with one set of winners.

Human intelligence is general too, and it never had this effect on industry. The reason is in Table 1. A gifted person can work in one field at a time, needs years to learn it, and cannot be copied, so the generality of a human mind is spent on a single career. A model's generality is available to every industry at once, in as many copies as there is compute to run them.

This explains why there will be few firms at the frontier, and why they will be the same few in every industry. It does not explain why one of them would pull away from the others, since rivals of equal size spread equal costs over equal sales. The general factor also leaves a remainder in every field, and that remainder is experience of the field.

7.3The data loop

The advantage that a rival of equal size does not get is data from deployment. The valuable data are the inputs of real tasks carried out over long horizons and, most of all, the record of how each attempt turned out. Nobody can scrape this from the internet or distill it from another model's outputs. It is produced only by doing the work, at scale, for real customers. A firm that deploys more learns more, a firm that learns more has the better product, and the better product is deployed more. Generality carries part of the gain across fields, since what a model learns from one kind of work raises its capability at others.

Whether the loop keeps paying depends on whether the returns to more data flatten. In well-covered tasks they do. Where reliability matters they flatten slowly, because what separates a system that is right 99 percent of the time from one that is right 99.99 percent is the rare case, and rare cases are seen only by whoever operates at scale. Driving, surgery, and the running of a factory are all of this kind. I assume the loop keeps paying across most of the physical economy. If it does not, the conclusion of this section weakens to a stable handful of similar giants (Section 10.2).

7.4The vertically integrated modeler

The loop has four steps: data, model, deployment to real users, and feedback. A firm earns the advantage only if it owns all four.

The third step is the demanding one. A firm that sells its model to others for deployment is a supplier. Its customers stand between it and the end user. They own that relationship, and they commonly forbid the supplier from training on what passes through. The feedback stops with them (Figure 8). To close the loop the modeler has to run the service. It operates the taxis where a supplier would license the driver, and it cuts the hair where a supplier would sell software to salons.

This is why VIMs absorb whole industries. Their route to a better model runs through the end customer, so every firm that stands between them and that customer is in the way, and Section 5 has taken away most of the organizational cost of replacing those firms. The general-purpose robot of Section 6.3 is the limiting case, with one firm and nothing else between the factory and the household.

A VIM does not need to own its supply chain. Chips, energy, and components can be bought. But nothing now stops it from integrating backward, and it will do so wherever it feels constrained or undersupplied.

7.5What the displaced firms lose

The firms in the way have defenses, and it is worth being exact about why these fail. The strategy literature holds that lasting advantage comes from resources that are valuable, rare, and hard to imitate (Barney, 1991). In practice these are a firm's people, the way they are organized, the routines they follow (Nelson and Winter, 1982), and the tacit knowledge that never gets written down (Polanyi, 1966). They are hard to imitate for one reason. They live in human heads and in the relationships among them.

To a VIM they are information. It does not need to hire an incumbent's people or spend decades learning its routines. Its model acquires the skill, and everything the incumbent's organization knew becomes a file (Table 4). The asymmetry runs one way. The VIM can reproduce what the incumbent knows. The incumbent cannot reproduce the VIM's loop, which is made of scale and deployment that the incumbent does not have.

7.6Capital follows

Everything a VIM needs besides its loop can be bought (Table 5). Energy is a commodity. Chips are a commodity with one dominant supplier, and so far that supplier has preferred selling them to using them.⁹ Research leads are real and short-lived, since the leading laboratories sit in a jurisdiction that does not enforce non-compete agreements and no secret is kept for long. Manufacturing at scale takes plant and time. Data is the exception. Capital can buy substitutes, by paying people to perform tasks while they are recorded or by simulating at high fidelity, but the real source is deployment.

Capital is therefore necessary, and it does not decide the race. A VIM spreads its fixed costs over a growing market, scales its organization at little cost, and improves its product with every customer it serves. Its return on capital rises with its size, and capital goes where returns are. Each increment of growth improves the terms on which the next is financed. A VIM that gets ahead becomes something like a black hole of capitalism, pulling harder the more it absorbs.

The two years since 2024 fit this account. Capital budgets that were around $50 billion a company in 2024 are now between $135 billion and $220 billion a company (note 2). In the first half of 2026 OpenAI and Anthropic raised rounds reported at $122 billion and $65 billion, at valuations approaching a trillion dollars each.¹⁰ The laboratories had a lead in capability and comparatively little capital, and the capital came to them.

How many VIMs will there be? The fixed-cost argument supports a handful, and there is room for several, each strongest where its own deployment is deepest. Generality works against a division of the economy among them, because the most capable model enters each new industry ahead (Section 7.2). Only a VIM whose loop spans many industries reaches the scale of Definition 1. I expect one or two.

In 2024 I named five candidates: Alphabet, Microsoft, Apple, Meta, and Tesla, chosen for their capital budgets, their data from billions of users or from machines deployed in the world, and their experience of manufacturing at scale. The test I would apply now is narrower. It asks which firms own all four steps of the loop, or can acquire the ones they lack before a rival closes it. States have capital too, and I expect them to lose a direct commercial race, because they lack real demand pulling their products forward and deployed systems returning data. States hold a different asset, which is the subject of Section 9.

8Precedents

No firm in a large economy has been a decacorp. Several came close enough to test parts of the argument (Table 6).

8.1The internal limits were real

The chartered companies are the cleanest test of Section 4. They had every reason to grow: monopoly rights, high fixed costs in ships and forts, and whole continents to trade with. What held them back was their own staff. The East India Company's servants in India traded on their own account, took presents from local rulers, and came home rich, while instructions from London took the better part of a year to arrive. Smith (1776) made the Company his leading example of why the directors of a joint-stock company, managing other people's money, cannot be expected to watch over it with care. Carlos and Nicholas (1988) describe the chartered companies as early multinationals whose central managerial problem was the control of agents at a distance. The Dutch company was wound up after decades of the same decay.

US Steel is the industrial version. It was formed in 1901 with about two thirds of American steel output and lost share for the next half century with no help from the courts (Chandler, 1977). These are forces (f), (g), and (h) operating as described, and they are the forces that SSI weakens.

8.2Every giant was confined

Each of these firms was limited to one trade or one territory: Asian commerce, oil, steel, telephony, cars. Its fixed costs and its know-how did not carry outside, so its ceiling was its industry's share of the economy. General Motors in 1955, with revenue near three percent of GNP, is close to the high-water mark for a private American company. Generality removes that ceiling (Section 7.2).

The nearest precedent to a VIM is AT&T, which held research, manufacturing, and operations in one firm, fed what it learned from running the network back into the next generation of equipment, and kept its position for seventy years (Temin, 1987).

8.3The state engages early, and at home it wins

None of these firms came near ten percent of a large economy before the state acted. Standard Oil was broken up in 1911. AT&T accepted regulation in 1913 as the price of its monopoly and was broken up in 1984. General Motors spent the 1950s under antitrust scrutiny, and in 1957 the Supreme Court ordered du Pont to give up its large holding in the company.

The East India Company's passage took eighty-five years and went by stages. It was bailed out in 1773, when it had become too large to be allowed to fail. A government board took control of its policy in 1784. It lost its monopolies in 1813 and 1833, and the Crown absorbed its territories in 1858 (Dalrymple, 2019). At no stage did the Company fight the British state. It lobbied, and many members of Parliament were its shareholders. Where a firm has rivaled its home state, the home state has prevailed every time.

8.4Abroad, the firm has often won

The record is different where a firm was large relative to a host that was not its home. The East India Company conquered Indian states, with an army about twice the size of Britain's own and with Britain eventually behind it. When Iran nationalized the Anglo-Iranian Oil Company in 1951, the answer two years later was a coup backed by the firm's home government and its American ally. Hosts have won as well, as the oil nationalizations of the 1970s showed.

A VIM that sells into every country will be a decacorp in many smaller economies well before it is one at home. Its first collisions are therefore likely to be with foreign governments, with its home state acting as patron.

8.5The Company's army

One episode bears directly on the next section. Company rule in India ended after the rebellion of 1857, which began as a mutiny in the Company's own army. Its soldiers were agents with interests of their own, and when those diverged far enough from the Company's, its military power turned against it. The armed part of a firm has always been subject to the same agency problem as the rest of it, with graver consequences. That is the constraint the third premise removes.

9Force

Capital is a legal claim. It buys processors because courts and police stand behind the sale. The one resource that is not purchased within that system, because it is what underwrites the system, is organized violence, and the modern state is the institution that claims a monopoly on its legitimate use (Weber, 1919). Tilly (1985) argues that European states themselves grew out of enterprises that sold protection.

For a firm of ordinary size this is background. For a decacorp it is not. Consider what such a firm controls: a tenth or more of national output, the compute on which much of the rest depends, the energy that feeds it, and a fleet of general-purpose machines in homes, factories, and streets. Measured by resources, this is an actor of the rank of a state. Its assets are also dual-use by construction. A machine that can do any physical task a person can do is not limited to civilian ones.

Figure 9 puts the firm and the state on one scale. A decacorp's value added is more than three times what the United States spends on defense, and more than half of everything the federal government collects in revenue.

The precedents of Section 8 say two things about what happens next. The state will engage long before the firm reaches ten percent, and at home the state has always prevailed. Three things are different this time.

First, a VIM's position does not rest on a charter or a concession that the state can withdraw. It rests on a loop that the firm built.

Second, the remedy that worked on Standard Oil and AT&T is poorly suited. Those firms could be cut along lines of geography or function into parts that still worked. A VIM's loop is its asset. Dividing it destroys the thing that made it valuable, and the state will want that thing for its own defense and administration.

Third, force has always needed the loyalty of many people. Soldiers who will not fire and officials who will not comply have been a check on states and firms alike, as the Company learned in 1857. Military capacity built from compute, energy, and manufacturing needs no such loyalty, and those are the inputs a VIM already commands.

I therefore expect a bargain or a fusion ahead of a break-up. The models are the regulated monopoly that AT&T accepted in 1913 and the closer entanglement of state and firm seen in the zaibatsu and the chaebol. Which of these it is, and on whose terms, is the open question.

That contest belongs to the social, military, and political parts of this series, which are still to be written.

10Objections

10.1Models are commodities

The best open model has stayed close to the best private one. That has been true because the leading models are little more than oracles trained on public data, and cheap enough to give away. As training runs grow and the data shifts toward self-play, synthetic generation, and end-user feedback, fewer organizations will have the resources and the data sources to keep up.

Distillation is the sharper form of the objection: train a small model on the outputs of a large one. This works when the input space is well defined and tightly constrained, as with text. When the inputs are live video and the outputs are actuator commands, generating even the inputs for distillation becomes prohibitive. Real-world processes of high value have too much complexity to sample cheaply.

10.2Profits are competed away

Suppose the objection above is right, or that the returns to deployment data flatten (Section 7.3). Then the handful of firms at the frontier have much the same product, and firms with the same product compete its price down toward cost. Generality sharpens this, because the same rivals meet in every market and none of them has a field to itself. The result is abundant, cheap intelligence, a large gain to everyone who buys it, and no decacorp. Escalating fixed costs would still keep the firms few and large.

A decacorp needs pricing power, and in this argument pricing power comes from the quality gap that the data loop opens between the leader and the next firm. Claim 3 stands or falls on that gap persisting. I think it persists in the physical economy, where rare cases matter and only the largest operator sees them. I am less sure of it for purely digital work.

10.3Falling prices shrink the measured share

When a technology makes something cheap, that thing usually becomes a smaller part of measured output, as food did. A firm could come to do much of the economy's real work and still be a small share of GDP.

There are three replies. The first is that the outcome depends on the elasticity of demand. Where demand is elastic, a lower price raises total spending, which is the paradox Jevons (1865) observed in coal, and Section 6.5 lists uses of intelligence with no obvious cap.

The second is the size of the pool. Google and Meta grew as large as they are on advertising, which is a percent or two of output. A VIM addresses the work now paid for in wages, which is more than half of it. It needs to keep only a fraction of that pool, and the fraction it keeps is set by its pricing power, which returns the question to Section 10.2.

The third is that the bottleneck moves. When one input becomes abundant, value shifts to whatever is still scarce, and a firm with no organizational limit follows it there, upstream included (Section 7.4). The process stops when the firm meets a constraint it cannot integrate around, one of physics and not of organization. A new competitive equilibrium forms at that point. Nothing in the argument says where that point is, and until it is reached the reasons that made today's largest firms large apply to a firm with far greater reach into the economy.

10.4AI may dissolve economies of scale

Suppose a fabricator at home can make anything in twenty minutes. That would cut forces (c) and (d), the two on which Claim 1 rests, and I have no decisive answer to it. I will say only that the fabricator's designs, feedstock, energy, and models come from somewhere, and that the economies of scale it removes downstream are likely to reappear upstream.

10.5The state will not allow it

It has not allowed it before (Section 8.3). Section 9 gives my reasons for expecting a bargain or a fusion this time, and those reasons could be wrong. A state that acts early enough, while the loop is still young and several firms are close, could hold the frontier to a regulated handful.

10.6The premise of selflessness

The last row of Table 1 stands. These systems are brittle and experimental, and “does what it is told” is an engineering goal that has not been met with certainty. Everything from Section 5 onward assumes that it is met. If it is not, this argument fails, along with a good deal else.

10.7Society may come apart first

It may. The transition described here closes most small businesses and reprices most labor within a couple of decades. Whether political institutions hold through that is the question the remaining parts of the series have to answer.

11Conclusion

The limits on the size of firms have been facts about people. People are hard to coordinate, hard to observe, and hard to motivate, and firms stopped growing where those difficulties outweighed the gains from scale. Take people out of production and the limits go with them, while the gains from scale stay.

The firms that grow into that space are the ones that own the loop from data to model to customer and back. Their fixed costs rise with the market and are spread across the whole economy. Intelligence is general, so the firms with the best models are ahead in every industry they enter. Their product improves with every customer served, and capital follows the returns. To close the loop they have to serve the end customer themselves, so they take the industries in between. One or two of them keep absorbing markets and capital until they meet the state.

Every earlier giant met the state early and lost to it at home. None of them held an advantage that the state could not revoke, could not divide, and needed for itself. The talk from which this essay grew drew the investor's conclusion, which was to own the firms best placed to close the loop and the firms that sell them compute. The larger questions concern what it is like to live in an economy of decacorps and who governs it.