Sovereign AI Economics: Who Controls the Future?
For most of the last decade, the conversation about artificial intelligence has been a conversation about capability. Can it pass the bar exam? Can it write code, fold proteins, hold a conversation that fools you for a paragraph or two? Those questions are mostly settled now, and they were always the easy ones. The hard question — the one that will decide what kind of century this becomes — is not what AI can do. It is who controls what AI does, and who keeps the value it creates.
That is an economics question before it is a technology question. And economics, unlike capability, does not trend toward abundance on its own. It trends toward whoever owns the bottleneck.
The Three Bottlenecks
Right now, the production of frontier AI rests on three scarce inputs, and each one is concentrating rather than dispersing.
The first is compute. Training and running a leading model requires enormous clusters of specialized chips, and the supply chain for those chips narrows to a handful of choke points: a few designers, essentially one company making the most advanced lithography machines, one foundry producing the leading-edge silicon at scale. You can have all the ambition in the world, but if you cannot get the hardware — or the energy and cooling to run it — you are not in the game. This is the most physical and least negotiable of the three.
The second is data. Models are built from the accumulated output of human civilization: text, images, code, behavior. The companies that already sit on the largest reservoirs of proprietary and behavioral data start each round of the race a lap ahead, and they are now moving to lock that data behind licensing walls precisely because they understand it is an asset, not exhaust.
The third is the model itself — the trained weights, the distribution, the user relationship. A frontier model costs a fortune to create and almost nothing to copy in operation, which is the classic shape of a natural monopoly. Whoever reaches a capability frontier first can fund the next frontier from the profits of the last. Advantage compounds.
Stack these together and you get a picture that should sober anyone who assumed AI would be a great democratizer. The inputs to intelligence are concentrating in fewer hands than the inputs to almost any previous general-purpose technology. Electricity and the internet diffused widely partly because their core was hard to monopolize. Frontier AI may be different.
Where the Value Lands
Concentration of inputs would matter less if the gains flowed broadly anyway. Sometimes they do — cheap tools lift everyone who uses them. But value tends to settle at the layer that is hardest to replace, and in this stack that is rarely the user and often not even the application builder.
Consider the structure. A nurse, a paralegal, or a small-business owner using an AI assistant becomes more productive — real value created. But much of that value is captured upstream: by the application that packages the model, by the model provider that the application pays per token, and ultimately by the compute owner the provider pays to run it. The person doing the work may keep a sliver. The chip at the bottom of the stack is rented by everyone above it.
This is the quiet redistribution already underway. We talk about AI "augmenting" workers, and it does. But augmentation that runs on infrastructure you rent, using a model you do not own, trained on data that may well include your own past work, is a peculiar kind of empowerment. You become more capable and less sovereign at the same time. The leverage is real; it just is not yours.
Multiply that across an economy and you can see why "who captures the value" is the defining distributional question of the era. If the answer is "the owners of the three bottlenecks," then AI does not flatten the economic pyramid. It pours a new, taller floor on top of it.
Sovereignty at Three Scales
The response taking shape, sensibly, is a push for sovereignty — economic control over AI rather than mere access to it. It looks different at each scale.
For nations, sovereignty means not being a permanent tenant in someone else's intelligence economy. Governments are funding domestic compute, backing open models, securing chip supply, and treating AI infrastructure the way an earlier generation treated railroads, ports, and power grids — as something too foundational to outsource entirely. A country that buys all its intelligence the way it buys all its oil has handed a lever to whoever holds the other end.
For companies, it means deciding how much of the stack to own. Renting a frontier model is fast and cheap to start; it also means your core capability, your margins, and your data flows run through a supplier who could change terms, raise prices, or become your competitor. The strategic move is to own the layer that is yours to own — the proprietary data, the fine-tuned model, the workflow — rather than becoming a thin wrapper on someone else's brain.
For individuals, sovereignty is the hardest and most important. It is the difference between AI that works for you — local, owned, aligned to your interests — and AI you merely log into, that learns from you while you rent it by the month. The same fork I keep returning to in my fiction shows up here in dollars-and-cents form: an intelligence that serves the many, or one that serves whoever owns it. The technology does not pick. The ownership structure does.
The Honest Counterpoint
None of this is destiny, and the pessimistic read can be overstated. Open-weight models keep arriving, often startlingly close to the frontier, and they erode the moat every time. Inference costs have fallen by orders of magnitude in just a few years, which pushes value back toward users. History offers some comfort too: railroads, telecom, and cloud computing all looked like permanent monopolies at moments, then got commoditized, regulated, or routed around. Concentration is a tendency, not a law.
But the comforting precedents took decades, and AI compounds faster than railroads did. The honest position is not despair — it is urgency. The window in which the structure of this economy is still soft, still shapeable by policy and competition and deliberate choices about openness, is open now and will not stay open forever.
So the question in the title is not rhetorical, and it is not really about the machines. AI will be as concentrated or as distributed, as extractive or as empowering, as the economic arrangements we build around it — arrangements we are writing this decade, mostly by default. The future is not something AI does to us. It is something we are deciding who gets to own. We should decide it on purpose.