Concentration of AI power: who controls the models

Behind the technical debates about AI sits a question of power: who actually controls the most capable systems, and what follows from the answer? Building frontier AI is expensive in ways that naturally push control toward a small number of well-resourced players. That concentration is neither inevitable nor harmless, and the counterweights to it are real but partial. This piece is a plain-language guide to why AI power concentrates, why that matters, and what pushes the other way — without cheerleading for any side.

Why concentration happens

The pull toward concentration is structural, not accidental. A few forces compound:

• Cost. Training the most capable systems requires enormous computing resources, large datasets, and scarce expertise. That price tag is out of reach for most organizations, which alone narrows the field.
• Scale advantages. The players who are already large tend to have the data, the infrastructure, and the money to stay ahead, and their lead can widen rather than shrink.
• Talent gravity. Scarce expertise clusters where the resources and the interesting problems are, concentrating capability further.
• Feedback loops. Successful systems attract users, users generate data and revenue, and those feed back into building even better systems.

Put together, these forces mean that without something pushing the other way, control over the most capable AI tends to gather in a few hands. Understanding that this is structural — a product of how the technology is built — is the first step to thinking clearly about it.

Why concentration matters

It is fair to ask why concentration is a problem rather than just an outcome of some players being better. The concerns are about dependence and leverage, not envy.

When a few entities control capabilities that many people and institutions rely on, those few hold significant power over the terms: what the systems will and will not do, who gets access, at what price, under what conditions. Decisions made by a small number of players ripple out to everyone downstream. Concentrated control also means concentrated single points of failure — if access changes, prices rise, or priorities shift, the dependent many have little recourse. And the values embedded in widely-used systems, chosen by their makers, shape how those systems treat everyone who uses them.

None of this requires bad intent to be worth caring about. Concentrated power is a structural risk regardless of whether the people holding it are well-meaning.

The counterweight: open approaches

The main structural pushback against concentration is the open-source and open-model movement. When capable models and the tools around them are openly available, control spreads. More people can build on the technology, inspect it, adapt it, and run it without permission from a gatekeeper. Open approaches lower the barrier that cost and scale otherwise raise, and they create alternatives to depending on a single provider.

This is not a complete solution, and it is honest to say so. Training the most capable systems from scratch still demands resources most people lack, so openness redistributes some power without erasing the underlying advantages of scale. There are also genuine debates about what should and should not be released openly, given that the same openness that empowers builders can empower misuse. But as a counterweight to concentration, open approaches are the most significant one, because they attack the access barrier directly rather than merely complaining about it.

Other counterweights

Openness is not the only force pushing against concentration:

• Competition. Multiple serious players checking each other is better than one, even if the field is not wide open. Rivalry constrains what any single player can dictate.
• Falling costs. What is expensive today often gets cheaper over time, which can gradually widen who can participate.
• Diffusion of expertise. Knowledge spreads. Techniques that were once rare become more widely understood, lowering one of the barriers.
• Public attention and governance. Scrutiny of concentrated power, and rules about how it may be used, can constrain it even when the concentration itself persists.

None of these individually breaks concentration, and it would be naive to claim they do. Together they make the picture less one-sided than the raw economics alone would suggest.

How to think about it without taking a side

This topic attracts strong, opposing stories. One says concentration is fine because the leading players are capable and responsible. Another says it is an emergency requiring drastic intervention. A more useful stance holds several things at once:

1. Concentration is a real structural tendency, driven by cost and scale, not a conspiracy.
2. It carries genuine risks around dependence, leverage, single points of failure, and whose values get embedded — regardless of intent.
3. Counterweights exist and matter, especially open approaches, even though none fully resolves the underlying economics.
4. The balance is contested and shifting, so confident predictions in either direction deserve skepticism.

Holding all four is more accurate than committing to either tidy story, and it leaves you better able to follow how the situation actually evolves.

Why this shapes everything downstream

Who controls the most capable models quietly shapes the rest of the AI landscape: what gets built, who can build it, what it is allowed to do, and who bears the risk if access or terms change. Debates that look purely technical or commercial often have this power question underneath them. Seeing it makes those debates easier to read — and makes clear why the push for open alternatives, the appetite for competition, and the interest in governance are all, at bottom, responses to the same structural pull toward concentration.

The takeaway

The concentration of AI power is a structural tendency, not an accident: the cost and scale required to build the most capable systems naturally gather control in a few hands. That matters because concentrated control means concentrated leverage over access, terms, and embedded values, and concentrated single points of failure — risks that hold regardless of anyone's good intentions. The most significant counterweight is the open movement, supported by competition, falling costs, spreading expertise, and public scrutiny, though none of these fully overrides the underlying economics. The clearest way to think about it is to hold the tendency, the risks, and the partial counterweights together, and to distrust anyone who tells you the question is already settled.