The Rise of Sovereign AI

For most of the modern era, you could read a nation's ambitions in what it dug out of the ground. The Industrial Age ran on coal and steel; the twentieth century ran on oil. We mapped power by tracing pipelines, shipping lanes, and the chokepoints where a single strait could hold a continent hostage. The mental model was physical: control the resource, control the future.

That model is quietly breaking. The defining contest of the twenty-first century will not be fought primarily over oil fields or rare-earth deposits — though those still matter — but over something stranger: cognitive infrastructure. The capacity to train, run, and direct artificial intelligence at scale is becoming the strategic substrate of national power, the way roads, railways, and the electrical grid once were. And just as no serious state in 1920 was content to import all its electricity, no serious state today is content to rent all its intelligence. This is the rise of sovereign AI, and it is reshaping the balance of power faster than most governments are prepared to admit.

The new strategic stack

An advanced AI capability is not one resource but a stack, and every layer has become contested.

At the bottom are the chips — the specialized processors that make large-scale training possible. Design of the most capable ones is concentrated in a handful of firms, their manufacture in an even smaller handful of plants, several clustered in places of acute geopolitical sensitivity. Above the silicon sits energy: frontier models consume electricity at industrial scale, which is why data centers have become a line item in national energy planning. Above energy sits data — the raw material, walled off by language, regulation, and national interest. At the top sit the models and the talent who build them, a population small enough to fit in a few buildings yet decisive enough to move markets.

Sovereignty here means operating up and down that stack without being switched off by someone else — the difference between owning your grid and leasing your lights from a landlord who can pull the plug whenever it suits him. The danger is asymmetry. A country can buy access to a powerful model and feel, for a while, that it has joined the future. But access is not control. The model can be throttled, repriced, or restricted under an export regime in another capital; the values baked into it were chosen elsewhere; the data flowing through it may not stay home. For a government, that is not procurement but a question of whether the nervous system of its economy answers to it or to a foreign provider.

Blocs, not just nations

What makes this moment distinct is that the response is forming along bloc lines, not merely national ones. The United States and China are the obvious poles, each treating advanced AI as strategic competition and using the tools of statecraft — export controls, industrial subsidy, investment screening — to widen its lead or narrow the other's. Restrictions on advanced chips are, in effect, an attempt to ration a rival's future.

But the more interesting movement is among everyone else. The European Union has leaned into regulation as a form of sovereignty, betting that whoever writes the rules exports influence the way an earlier era exported legal and financial standards. Gulf states are spending oil wealth to buy into the compute layer directly, converting a depleting resource into a durable one. India is pushing to build domestic capability rather than remain a back office for someone else's models. Smaller nations are pooling compute into consortia, precisely because none can fund a frontier effort alone.

Underneath the policy language is a single recurring fear: technological vassalage — not only falling behind, but becoming structurally dependent on a patron, your industries optimized around their tools, your leverage gone. The twentieth century knew this as resource dependence. The twenty-first is learning it as cognitive dependence, harder to see because it doesn't arrive in tankers.

The case against the panic

It would be too easy to end there, with a clean story of a new great game. Honesty requires three counterpoints.

First, autarky is a fantasy. No country, not even the largest, controls the entire stack — chips, the tools to make them, energy, talent, and data are distributed across borders by deep economic logic, and forcing them home is ruinously expensive and often impossible. In its maximalist form, "sovereign AI" becomes a justification for waste: duplicated infrastructure, subsidized champions that never reach the frontier, walls that mostly inconvenience a country's own builders.

Second, fragmentation carries a cost the resource framing obscures. Intelligence is not oil: oil burned in one country does nothing for another, but a breakthrough shared can lift everyone. If the world splinters into incompatible blocs — divergent standards, restricted research, models that cannot speak across the divide — we may all end up with less capable, less safe systems than open exchange would have produced. The pursuit of sovereignty can quietly trade away the gains of cooperation.

Third, and most uncomfortably, sovereignty over a technology is not wisdom in using it. A state can own its compute and still aim it at surveillance or repression. The question of who controls the cognitive infrastructure is real, but it sits alongside an older one — to what end — that no amount of national ownership answers on its own.

What we are actually deciding

I've spent enough time inside technology companies to distrust both the breathless and the dismissive accounts. The shift from buying intelligence to building it looks like ordinary procurement up close, and a redrawn map from a distance — the kind of change the people living through it rarely recognize.

In Utopia, I kept circling a question this moment makes literal: when intelligence itself becomes infrastructure, what does it mean to be free of it — or beholden to it? A people who do not control the systems that increasingly think alongside them have outsourced a piece of their own agency, and may not notice until they try to reclaim it and find they no longer remember how.

That is the real stake beneath the geopolitics. On its surface, the race for sovereign AI is about chips and grids and export licenses. Beneath it is a question every nation — and every citizen — will answer this century: do you want to be a maker of the intelligence that runs your world, a renter of it, or a user who never thought to ask who held the keys?

The strategic substrate of the next age is being poured right now. The states that understand it as infrastructure, not a gadget to be purchased, will be the ones still steering when the concrete sets.