Who Controls AI? How a Few Companies Became Its Real Regulators
Synthocracy Institute · Explainer · by Martin Novak · July 2026
Short answer: Not governments, mostly — and not any single company. Control over AI is concentrated in a small number of private firms that own the layers everyone else depends on: the chips, the cloud computing, the frontier models, and the app stores and platforms that reach users. Together they act as the real, largely unelected regulators of what AI can do and who can use it.
Key takeaways
- The debate over “AI governance” fixates on states, but the day-to-day rules are set by private infrastructure owners.
- The AI stack has a few chokepoints: chips (Nvidia holds an estimated 70–95% of the AI chip market), cloud compute (five hyperscalers are spending over $600 billion in 2026), frontier models, and distribution.
- Antitrust regulators in the US, EU, and UK have flagged the concentration — but frontier labs and cloud providers have largely avoided direct action so far.
- Even government control tends to run through these companies: when a state wanted a frontier model switched off, it had to order a private lab to do it.
- The accountability problem is structural: you can’t vote out a model provider.
Who controls AI?
Ask “who controls AI” and most people picture governments, or maybe one famous CEO. The more accurate answer is a short list of companies that own the layers everyone else builds on. You don’t need to own “AI” to control it. You only need to own a layer no one can route around.
There are a handful of these layers, and each is a point of control.
The chokepoints of AI power
Chips. Advanced AI runs on specialised processors, and one company dominates: Nvidia holds an estimated 70–95% of the AI chip market, with most of the field’s infrastructure built around its hardware. Upstream, a tiny number of firms — the Dutch equipment maker ASML, the Taiwanese foundry TSMC — are effectively irreplaceable.
Compute. Training and running frontier models requires enormous data centres. In 2026, five hyperscalers — Amazon, Microsoft, Alphabet, Meta, and Oracle — are collectively spending on the order of $700 billion, roughly three-quarters of it on AI infrastructure. Cloud contracts can lock customers in; smaller buyers get longer waits, higher prices, and worse terms.
Models. The most capable foundation models come from a few labs, most now tied to a cloud giant that funds them and hosts them. Microsoft’s stake in OpenAI and Amazon’s in Anthropic are the best-known; regulators have examined whether such deals and contract terms make it hard to switch providers.
Distribution. Reaching users runs through more gates: app stores (Apple, Google), operating systems, browsers, and messaging platforms. When a company restricts rival AI’s access to its platform — as regulators alleged over access to WhatsApp — it is exercising a form of rule-making over the whole market.
Own enough of these, and you don’t compete in the AI market. You set its terms.
How did private companies become AI’s regulators?
Because a rule doesn’t have to be a law to function like one. When a platform decides which AI tools may plug in, which uses are allowed, whose content is visible, and who gets compute first, it is writing rules that bind millions — without a vote, a hearing, or an appeal. Scholars of platform governance have long called these firms the “new governors”: private bodies setting public rules.
In AI, that power is unusually concentrated because it sits on physical scarcity — chips, energy, data centres — not just code. And it compounds: the same firm can own the chips, the cloud, the model, and the app, so that a decision made for commercial reasons at one layer quietly shapes what is possible at all the others. These are, in effect, private constitutions of visibility and access.
Isn’t the government in charge?
Governments are trying to be — but two things complicate the picture.
First, the enforcement gap. Antitrust authorities have clearly identified the problem: in a joint 2024 statement, the US DOJ and FTC, the UK’s CMA, and the European Commission warned about concentrated control of key inputs, incumbents entrenching their power, and deals that steer the market. Since then, Nvidia has faced investigation and had its Arm acquisition blocked; the EU’s competition chief has put Big Tech’s “entire” AI operations under scrutiny. Yet the frontier labs and cloud providers themselves have largely avoided direct antitrust action, even as their market power grows.
Second — and this is the deeper point — even government power over AI often runs through these private companies. When a state recently wanted the world’s most capable model taken offline, it could not flip the switch itself. It had to order a private lab to do it, and the model went dark across that company’s own platforms and every major cloud at once. The public authority was real, but it was exercised through private infrastructure. That is not the state controlling AI. That is the state depending on the same chokepoints as everyone else.
→ We examined that episode here: The Switch Is the Story
Why does this matter?
This is the corporate face of synthocracy. The debate about who governs AI usually looks at parliaments and regulators. But the infrastructure that actually decides what AI can do, who can access it, and what it will and won’t be allowed to say sits with a handful of private firms — and they are accountable to shareholders, not citizens.
The concern is not that these companies are villains. Many do serious safety work, and concentration can bring real benefits: capability, coordination, and the resources to manage risk. Reasonable people argue that some concentration is efficient, even safer, and that punishing firms for size alone would harm innovation. That debate is genuine and worth having.
But the structural fact remains, and it is simple:
You can vote out a government. You cannot vote out a model provider.
When the power to shape a technology this consequential concentrates in private hands with no equivalent of an election, an appeal, or a public record, the question is not whether those hands are good or bad. It is whether anyone outside them can see, contest, or reverse what they decide.
What would accountable platform power look like?
Not a binary between nationalising AI and leaving it alone. More useful are the ordinary tools of accountability, applied to a new kind of power:
- Transparency about how access, ranking, and moderation decisions are made.
- Interoperability and real switching — so lock-in doesn’t turn a vendor into a sovereign.
- Contestability — a way to challenge a platform’s decision and reach a human with authority.
- A right to another path — so being cut off by one provider isn’t the same as being cut off from the service.
- Admissibility of access decisions — a reviewable record of why a system, or a user, was granted or denied access.
None of this requires treating these companies as enemies. It requires treating their decisions as the exercises of power they already are.
→ For our research on this, see Research. For the method underneath, see Admissibility. To test any single system, use The Ten Questions.
FAQ
Who controls AI? No single entity. Control is concentrated across a few private firms that own the key layers — chips, cloud compute, frontier models, and distribution. Together they set the terms for what AI can do and who can use it.
Does the government control AI? Only partly. Antitrust regulators in the US, EU, and UK have flagged AI market concentration, but frontier labs and cloud providers have largely avoided direct antitrust action. Government power over AI often has to be exercised through the same private companies.
Why is AI power so concentrated? Because AI depends on physically scarce inputs — advanced chips, energy, and data centres — controlled by very few firms, and because the same companies often own several layers at once (chips, cloud, model, and app).
Are tech companies effectively regulators of AI? In practice, yes. By deciding which tools may access their platforms, which uses are allowed, and who gets compute, they set rules that bind millions of users — without a vote or a formal appeal.
Why does concentrated AI power matter for democracy? Because decisions that shape a foundational technology are made by firms accountable to shareholders, not citizens — and unlike a government, a model or cloud provider cannot be voted out.
Martin Novak is the founder of the Synthocracy Institute, an independent research institute studying how decision-making power shifts through AI systems. Warsaw, operating internationally. Sources include the July 2024 DOJ/FTC/CMA/European Commission joint statement on AI competition, US and EU antitrust reporting, and public data on 2026 AI infrastructure spending.
