The AI Monopoly: How a Handful of Companies Captured the Future

The big tech AI monopoly is the defining economic story of our decade, and yet the phrase itself is slightly wrong. There is no single company that owns artificial intelligence the way Standard Oil once owned refining. What we have instead is a tight cluster of firms that between them control the compute, the cloud, the data, the capital and the distribution that any serious AI system depends on. Call it what it is: an oligopoly with a very small membership. But the effect on the rest of us — the developers, the startups, the workers whose labour trained these models — is starting to look uncomfortably like the effect of a monopoly. A handful of companies captured the future, and now they charge everyone else rent to live in it.

I keep coming back to a pattern I have written about for years, because AI fits it so cleanly. New technology arrives wrapped in the language of liberation. It genuinely does democratise something for a while. And then, quietly, the essential layer underneath it gets captured — owned, gated, metered — by whoever moved fastest and had the deepest pockets. Understanding how technology gets captured is the whole game here. AI did not escape the pattern; it is the purest example I have seen in my lifetime.

Four chokepoints, one small club

To see how power concentrated, forget the chatbots and look at what sits beneath them. Modern AI runs on four scarce things, and each one has been cornered.

The first is compute. Training a frontier model requires tens of thousands of specialised chips running for months, and the supply of those chips flows through an astonishingly narrow channel — a single dominant designer, a single dominant fabricator, packaged into data centres only a few companies can afford to build. If you cannot get the chips, you cannot train the model.

The second is cloud. Almost nobody owns their own data centre at scale, so the compute you rent comes from the same three or four hyperscalers who already dominated cloud computing before AI existed. They rent you the machines to build your model, and increasingly own a stake in the model you build. The landlord is also your investor and, often, your competitor.

The third is data. These systems are trained on the accumulated output of humanity — our books, our code, our photographs, our forum arguments at two in the morning. The firms with the largest existing pools of user data start with an advantage no newcomer can replicate, because the data was gathered over decades of earlier dominance.

The fourth is distribution. It does not matter how good your model is if nobody can reach it. The companies that already own the phones, the browsers, the search boxes and the app stores can put their AI in front of billions of people by default, on day one, at no acquisition cost. A brilliant independent model with no distribution is a tree falling in an empty forest.

Notice that the same names recur across all four columns. When the firm that makes the chips, rents the cloud, holds the data and owns the distribution is drawn from a pool of maybe five companies, the question of who owns AI answers itself. It is not a conspiracy. It is a stack, and they own the stack.

A brilliant independent model with no distribution is a tree falling in an empty forest.

Why it snowballs instead of settling

Ordinary markets tend toward equilibrium — a leader emerges, competitors nip at the edges, prices fall, the advantage erodes. AI does the opposite. It compounds. And it compounds because of the specific way the technology is shaped.

Start with the obvious flywheel. More users generate more interaction data, which makes the model better, which attracts more users. That loop has driven platform monopolies before, but AI adds a crueller twist: the improvements require enormous capital, and only the incumbents have it. Each turn of the wheel raises the price of entry for anyone hoping to catch up. The gap does not stay static; it accelerates.

Then there is the control of the cloud, which quietly decides who is even allowed to compete. A startup with a better idea still has to rent its compute from one of the giants — who sets the price, sees the usage patterns, and can at any moment launch a competing product informed by what it watched its own tenants build. When the essential infrastructure is owned by your rival, you are not really competing. You are subletting.

Capital seals it. The sums required to stay at the frontier have grown so large that the field of players who can even attempt it shrinks with every generation of model. Money flows to the firms that already have money, in partnerships structured so that the cloud provider funds the lab and the lab spends the funding right back on the provider’s cloud. The dollars make a round trip and the concentration deepens. This is part of what people mean when they talk about technofeudalism — an economy where the richest players no longer merely sell in the market but own the territory the market runs on, and collect a toll from everyone who wants to build there.

An old playbook in new clothes

None of this is unprecedented; it only feels new because the product is dazzling. Strip away the neural networks and you find one of the oldest strategies in industrial history: capture the essential rail, then charge everyone rent to use it.

The railroads did it first, and most nakedly. In the nineteenth century, whoever controlled the tracks controlled the farmers, the factories and the towns that depended on moving goods. The railroad did not need to grow the wheat or forge the steel. It just needed to own the only route to market, and then it could set the price of everyone else’s livelihood. Farmers who had built their whole lives around a rail line discovered, too late, that the line owned them.

Standard Oil ran the same play in a subtler register. Its genius was never really about oil. It was about controlling the pipelines and the refining and, crucially, cutting secret deals with those same railroads so that competitors paid more to ship the identical product. The lesson: if you own the chokepoint, you do not have to win on quality or price out in the open. You win underneath, where the essential infrastructure lives, and the surface market bends to whatever you decide.

Swap the tracks and the pipelines for chips and clouds and you have described the AI industry with almost no translation. The compute layer is the new rail. The cloud is the new pipeline. The distribution defaults are the new secret shipping rate, quietly ensuring the incumbent’s product reaches the customer cheaper and faster than yours ever can. The technology is miraculous. The business model is a century old.

The technology is genuinely miraculous. The business model is a century old.

The honest nuance: oligopoly, not one ring to rule them all

Here I have to be fair, because the story is more interesting than a simple villain. This is not a monopoly in the strict sense, and pretending otherwise weakens the argument. There is real competition at the frontier. Several well-funded labs are racing each other hard, and the lead changes hands often enough that no single firm can rest. That rivalry has pushed capabilities forward at a startling pace and, for now, kept some prices lower than a true monopolist would allow.

More encouraging still, there is real life at the edges. Open-weight models have arrived that are good enough for a great many tasks, and they can be run by anyone with modest hardware, outside the control of any gatekeeper. Smaller specialised models increasingly beat giant general ones on specific jobs at a fraction of the cost. The moat is real, but it is not infinitely deep, and every efficiency breakthrough lets a little more water out of it.

So the accurate word is oligopoly — a market controlled by a few — not monopoly, a market owned by one. Why does the distinction matter? Because oligopolies are more stable and more comfortable than they look. A handful of dominant firms rarely need to collude explicitly. They simply read each other’s moves, avoid the ruinous price wars that would hurt them all, tolerate a fringe of small competitors as proof that the market is “open,” and quietly keep the essential infrastructure among themselves. Competition at the edges is real. It is also no threat to who owns the centre. The club can afford to let a thousand startups bloom precisely because they still have to rent the rail.

This was a choice, not a destiny

Here is the argument I care about most, and the one the “AI is just too powerful to distribute” crowd would rather you never examine. The concentration we have was not written into the technology. It was produced by a series of choices — some active, many passive — and different choices remain available. Nothing about matrix multiplication requires that five companies own the future.

Consider what those choices actually were. Through decades of slack antitrust enforcement, we let the firms that dominated the previous era buy or starve their way into dominating this one, rather than treating essential compute and cloud as the bottlenecks they plainly are. We chose weak data rights, letting models train on the collective output of millions of people who were never asked, never paid and never credited. We treated open models as a curiosity to tolerate rather than a public good to support. And we left the underlying infrastructure almost entirely in private hands, when much of the original science was funded by the public.

Every one of those was a decision, and every one has an alternative. Antitrust can be enforced against infrastructure chokepoints, not just consumer prices. Data rights can be written so that the people whose work trains a model have a say and a share. Open models can be funded and protected as deliberate policy. Compute can be treated the way we treat other essential utilities — with public options and rules that stop the landlord competing with the tenants on rigged terms. I have laid out elsewhere what a serious version looks like, because the mechanics of how to break up Big Tech apply to AI with barely any adaptation. The tools exist. What has been missing is the will to use them before the concrete sets.

What a less-captured future could look like

It is worth being concrete. A less-captured AI future is not one without powerful models or profitable companies. It is one where the essential rail is not privately owned by the firms that compete on top of it.

In that future, compute is available as a genuine public option, so a researcher or a startup can train and run models without begging a rival for capacity. Capable open models are a permanent, well-funded fixture, not a gift that could be withdrawn the moment it stops being strategically convenient. The people whose data trains these systems have enforceable rights over that use, so the value flows back to the many rather than pooling with the few. And regulators treat control of the AI stack the way earlier generations learned to treat the railroads and the refineries — as infrastructure too important to leave to whoever grabbed it.

In that world, AI is still transformative. It is just not a private tollbooth on the future. The gains from a technology built on all of our collective knowledge are shared more widely than the shareholder register of five companies. That is not utopian. It is the ordinary, hard-won work of refusing to let an essential thing be quietly enclosed — the same work that broke the railroads’ grip and split up Standard Oil, done again for the defining technology of our own moment.

The people who tell you the current concentration is inevitable are usually the people it benefits. It is not inevitable. It is a set of choices, most of them still open. Who takes, who pays, and who fights back are being decided right now, while the rules are still soft enough to write. That is the point of noticing the pattern early: not to marvel at how the future got captured, but to remember that it was captured by decisions — and decisions can be made differently.