Who Actually Owns AI? Follow the Money Behind the Machine

Almost every conversation about artificial intelligence is a conversation about power — but the wrong kind. We argue about how smart the models are, whether they can reason, when they will take our jobs. Far fewer people stop to ask the question that actually decides the outcome: who owns AI? Not who builds it or who talks about it loudest, but who controls the machinery underneath — the compute, the data, and the capital. Because ownership, not raw capability, is what determines who captures the gains from this technology and who ends up paying for them. If you want to know where AI is heading, don’t follow the demos. Follow the money.

This is not a new story. It is the oldest story in the history of technology, wearing a new set of clothes. Every era has a machine that promises to liberate us, and every era ends with that machine owned by a handful of people who charge the rest of us for access. The steam engine, the railway, the electrical grid, the search box — the same move, a new machine, every time. AI is the latest and most concentrated version of it. So let’s do the unglamorous work of tracing who actually holds the reins, and why that matters more than any benchmark score.

The three things that confer control

Strip away the mystique and a modern AI system rests on three pillars. You need enormous amounts of compute to train and run the models. You need vast quantities of data to train them on. And you need staggering amounts of capital to pay for both. Whoever controls those three inputs controls AI — regardless of who writes the clever research papers or ships the friendly chat interface. Capability is downstream of ownership. A brilliant model you cannot afford to train is just a diagram.

What makes this moment unusual is how tightly all three pillars have concentrated into the same small circle of firms and their investors. In earlier waves, the inputs were at least somewhat distributed — many factories, many landowners, many banks. With AI, the compute, the data and the money increasingly loop back to the same names. That overlap is the whole game.

Compute: the new landlords of the cloud

Training a frontier model requires tens of thousands of specialised chips running for months inside purpose-built data centres, drawing enough electricity to power a small city. There is no garage version of this. The hardware is expensive, supply is constrained, and the facilities to house it are owned by a very short list of companies — the same large cloud providers that already sit underneath most of the internet.

This is where the phrase cloud capital earns its keep. Think about what a landlord does: they own an asset you cannot easily build for yourself and let you use it in exchange for rent. Cloud compute works the same way. A tiny number of firms own the data centres, the chips, and the networks. Everyone else — startups, researchers, even rival AI labs — rents access. And because there is nowhere else to go, the terms are set by the owner, not the tenant. That is not a market of equals trading. It is rent extraction, dressed up in the language of platforms and services.

Whoever owns the compute owns the ceiling. Every model anyone else builds is a tenant in someone else’s data centre.

The rent has a compounding quality, too. The more you own the underlying infrastructure, the cheaper it is for you to train your own models, and the more you can charge others to train theirs. Profits from renting compute fund the next generation of even bigger data centres, which raises the price of admission again and pushes the frontier further out of reach for anyone without landlord-scale resources. It is a flywheel that turns temporary advantage into structural ownership. This is the mechanism I keep coming back to when I look at how technology gets captured: control of the essential input quietly becomes control of everything built on top of it.

Data: the raw material nobody agreed to hand over

The second pillar is data. Models learn from text, images, code and video scraped from the open web, from books, from the accumulated output of millions of people who never imagined their work would become training fuel. The firms with the largest and richest data holdings — search indexes, social graphs, email, cloud documents, years of user behaviour — start with an enormous advantage that no newcomer can replicate, because you cannot go back in time and collect two decades of the world’s digital activity.

There is a quiet injustice buried in this. The value of a model is built from the collective labour of everyone whose writing, art and code went into it, yet the ownership of the resulting system sits entirely with the firm that did the scraping. The people who supplied the raw material get nothing, while the company that assembled it captures the whole return. This is the same take-without-asking pattern that runs through every capture story, and it is why the fight over who owns what AI makes is not a side issue for lawyers — it is a fight about whether the people who created the inputs have any claim on the outputs at all.

Data ownership also feeds a second flywheel. The more users a system has, the more interaction data it collects; the more data it collects, the better it gets; the better it gets, the more users it attracts. Incumbents who already run the platforms where billions of people spend their time can bootstrap this loop instantly. A challenger starting today cannot. So even the data pillar, which in principle belongs to all of us, hardens into an asset owned by a few.

Capital: who funds the labs decides what they become

The third pillar is money, and it is the one that ties the other two together. Building a frontier lab now costs sums that only a handful of players on earth can write. Those sums come from a small pool of large technology companies and deep-pocketed investors — often the very same firms that own the compute. When a cloud giant invests billions in an AI lab and much of that money flows straight back to the giant as payment for compute, the “investment” is also a way of locking in a customer and a dependency. The capital and the compute are two ends of the same rope.

This matters because whoever funds the lab shapes what the lab is allowed to become. A research group that set out to build AI for the public good discovers, once it needs billions to keep training, that it must answer to investors who expect a return. Mission bends toward monetisation. The pressure is not a conspiracy; it is simply what capital wants. Money that expects to be paid back reshapes any organisation around the goal of paying it back. And so the labs that could have been stewards of a shared technology drift toward becoming vehicles for extracting rent from it.

Ownership decides whose interests the machine serves. Capability just decides how well it serves them.

Put the three pillars together and you see the shape of the thing. The same firms own the compute, hold the deepest data, and supply — or receive — the capital. The circle closes. This is why I find the framing of technofeudalism so useful: we are drifting from a world of markets, where firms compete to sell you things, toward a world of digital estates, where a few owners control the essential platforms and everyone else pays for access. AI is the most powerful estate yet, because it sits on top of all the others.

Why ownership beats capability

Here is the claim I want to make plainly. When people ask whether AI is dangerous, they usually mean the technology itself — will it deceive us, will it escape our control, will it be too capable. Those are real questions. But the more immediate danger is quieter and more mundane: a technology of extraordinary usefulness, owned by a very small number of people, who get to decide the price of admission and the rules of use for everyone else.

Ownership is what turns capability into leverage. If you own the model that a million businesses now depend on, you set the price, and you can raise it once they can no longer function without you. You decide what the model will and won’t do, which uses are permitted, which voices are amplified or suppressed. You capture the productivity gains — the work that used to be done by paid humans is now done by your machine, and the savings flow to you, not to them. Capability without ownership is a public good. Capability plus ownership is a private tollbooth on a road everyone has to travel.

This is the through-line of every capture worth studying. A new machine arrives with the promise of liberation. For a brief window it genuinely is liberating. Then the owners of the essential input consolidate their grip, the tollbooths go up, and the gains that were supposed to be shared get funnelled to whoever holds the deed. Who takes, who pays, who fights back — those are the questions that ultimately matter, and the answer is written in the ownership structure, not the technology.

What broader ownership could look like

None of this is inevitable. The concentration is a choice — a series of choices about who is allowed to own what — and choices can be made differently. There are at least three routes toward a world where the gains from AI are more widely held. I want to lay them out honestly, caveats included.

Open models

The first is opening up the models themselves — publishing the weights so that anyone can run, study and adapt a capable system without renting it from a landlord. When a genuinely strong model is released openly, it resets expectations overnight; it shows that the frontier is not a private secret and that competition is still possible. The arrival of credible open-source AI did exactly this, puncturing the idea that only a couple of firms could ever operate at the top. The honest caveat: open weights still need expensive compute to train and run, so “open” does not automatically mean “accessible to all,” and open models can be misused. But distributing the model is still a real transfer of power away from the few, and it changes the negotiating position of everyone downstream.

Public compute

The second route attacks the compute pillar directly. If access to large-scale computing is the chokepoint, then treating it partly as public infrastructure — publicly funded compute for researchers, universities, startups and public institutions — loosens the landlord’s grip. We have done this before with roads, electricity and the early internet: infrastructure too important to leave entirely to private tollbooths. The caveat is that public compute is expensive and hard to govern well. It is not a magic solution. But even a modest public option changes the market, because it gives tenants an alternative to the private estate — and an alternative is exactly what rent extraction depends on you not having.

Data rights

The third route addresses the data pillar by giving people real rights over the material their lives generate — the right to know when your work is used to train a system, the right to refuse, and the right to share in the value created from it. If the training corpus is built from collective human output, then the collective should have some claim on the result. The caveat is that data rights are genuinely hard to design: individual data points are nearly worthless on their own, value emerges only in aggregate, and clumsy rules can entrench incumbents who can afford the compliance while crushing the small players who cannot. Done badly, it backfires. Done well, it re-establishes a basic principle: the people who supply the raw material deserve a stake in what gets built from it.

None of these three is sufficient alone, and none is easy. But notice what they have in common: each one takes a pillar of ownership — compute, data, capital — and tries to widen the circle of who holds it. That is the real fight. Not whether AI becomes more capable — it will — but whether its ownership stays locked inside the same small circle or gets prised open.

Follow the money, then decide

So the next time you read a breathless headline about what AI can now do, train yourself to ask the second question immediately after the first. Not just how powerful is it, but who owns the thing that makes it powerful — the data centres, the training data, the capital. The capability will keep improving no matter what we do. The ownership is the part still up for grabs, and it will decide whether this technology becomes a shared inheritance or another private estate we all pay rent to enter.

The machine is new. The move is ancient. Someone is going to own AI. The only open question — the one worth fighting over — is whether that someone is a handful of firms, or the rest of us.