AI Nationalism: Why Every Country Suddenly Wants Its Own Model

Not long ago, the phrase sovereign AI would have drawn blank stares in most cabinet rooms. Today it is on the lips of prime ministers, central bankers and defence chiefs on nearly every continent. The idea is simple to state and hard to deliver: a country should control its own artificial intelligence (the models, the compute, the chips and the data underneath them) rather than renting all of it from a handful of firms in a couple of foreign cities. What started as a niche worry among policy wonks has become a full-blown wave of AI nationalism, with governments racing to fund national models, build domestic data centres and lock down the hardware that makes any of it possible.

I find this moment fascinating, because it repeats a pattern I have watched play out with almost every powerful technology. Something world-shaping arrives, a few players quietly gather control of its most valuable layers, and everyone else wakes up and scrambles to claw back some ownership. Sovereign AI is that scramble, dressed in the language of statecraft. The interesting question is not whether countries will pursue it — they already are — but who holds power when the dust settles.

What “sovereign AI” actually means

Strip away the slogans and sovereign AI is a claim about control across a stack. At the bottom sit the chips, the specialised processors that train and run large models. Above them sits compute: the data centres, the energy, the networking that turn a warehouse of chips into usable capacity. Above that sit the models themselves, the trained systems and increasingly the ability to train new ones. Running through all of it is data: the text, images, records and behaviour that models learn from and act upon.

To be truly sovereign, the argument goes, a country needs meaningful control over each of these layers, or at least enough that it cannot be cut off, price-gouged or quietly shaped by an outside party. In practice almost no one has the full stack — even the largest economies depend on chips designed in one country, manufactured in another, using equipment from a third. So sovereign AI is really a spectrum of ambitions: from a national model trained in the local language, to domestic data centres, to a homegrown chip industry that takes decades and staggering sums to build.

What unites these ambitions is a refusal to be a pure consumer of someone else's intelligence — and that refusal explains why the topic has caught fire so fast.

Why it is suddenly everywhere

Several currents have converged, and each on its own would be enough to push a government to act. Together they have made sovereign AI feel urgent almost everywhere at once. Five stand out.

Economic ambition

The first driver is money. Leaders have decided — rightly or not — that AI will be the defining productivity engine of the coming decades, and that a country which merely imports it captures a sliver of the value while exporting the rest. If your firms, hospitals and factories all run on models owned elsewhere, the profits and the best jobs pool where those models live. Many governments have concluded that funding domestic capability is less an expense than an investment in not being left behind.

National security

The second driver is security, and it moves budgets fastest. Modern AI touches intelligence, logistics, cyber-defence and increasingly the battlefield itself. No defence ministry is comfortable running critical systems on infrastructure it cannot audit, cannot guarantee access to, and might see throttled the moment a political relationship sours. Recent trade shocks taught a hard lesson about dependency on distant supply chains, and AI inherited that anxiety. When a capability is deemed strategic, cost stops being the deciding factor.

Language, culture and fit

The third driver is quieter but, to me, one of the most legitimate. The dominant models were trained overwhelmingly on a handful of languages and worldviews. If you speak Malayalam or Amharic or Vietnamese, the off-the-shelf systems often stumble — mangling the grammar, missing the context, flattening the culture into something generic. A country that wants AI to serve its own people, in their own tongues and with their own norms, has a real reason to want models shaped locally. This is not vanity: a translation tool that fails in your courts or clinics is not a neutral inconvenience, it is a quiet form of exclusion.

Fear of depending on a few foreign firms

The fourth driver ties the others together. The frontier of AI is controlled by a strikingly small number of companies, most concentrated in one or two countries. I have written before about who owns AI and about the AI monopoly quietly forming at the top of the stack, and it is not reassuring. When your entire national capability depends on a few firms' pricing, terms of service and willingness to keep serving you, you are not a customer so much as a tenant. Governments have noticed that tenants can be evicted.

When your national capability depends on a few foreign firms' pricing and goodwill, you are not a customer so much as a tenant — and tenants can be evicted.

Not wanting to be a data-field

The fifth driver is closest to my own preoccupations. For many countries, especially outside the wealthy West, the danger is not just dependence but extraction. Their citizens generate enormous volumes of data — every search, transaction, medical record and voice note — and it flows out to train models owned elsewhere, whose profits and control flow back to a distant few. I have called this data colonialism, and the label fits: raw material taken cheaply from one place, refined into something valuable somewhere else, then sold back to the people it came from. Sovereign AI, at its best, is a country saying it would rather not be a mere field to be harvested.

The tensions nobody should ignore

So far this may sound like an unambiguous good: plucky nations reclaiming their technological future. I am sympathetic to the impulse, but the path is not clean. It is riddled with tensions, and some cut against the very citizens sovereign AI claims to protect.

The staggering cost

The first tension is brute economics. Building even a slice of the AI stack is extraordinarily expensive. Frontier models cost enormous sums to train; data centres devour capital and electricity; a domestic chip industry is the work of decades and fortunes, with no guarantee of catching a moving frontier. Money spent chasing sovereign capability is money not spent on schools, clinics or the grid. For a wealthy state this is a manageable bet; for a poorer one it can be a ruinous vanity project, cheered on by vendors delighted to sell the shovels.

Subsidising national champions

The second tension is about who catches the money. When a government decides AI is strategic and opens the treasury, that spending usually lands on a chosen few: a national-champion firm, a favoured cloud provider, a politically wired conglomerate. Public money flows in; private ownership stays put. The citizen pays through taxes and is told it is for the national good, while the upside accrues to shareholders well-connected enough to be picked. This is the mechanism I keep returning to in how technology gets captured: a genuine public need becomes the justification for a private windfall, and the people footing the bill rarely own any of what they funded.

“Sovereign” can just mean state-controlled

The third tension is the most serious, and it is baked into the word itself. Sovereignty is about who holds power — and “the nation” controlling AI is not the same as “the people” controlling it. In too many places, sovereign AI is quietly becoming a licence for the state to build capabilities its citizens would never knowingly approve: mass surveillance, censorship at scale, prediction and control of populations, all justified as protecting the homeland. A national model trained on national data, running on national servers, under national security law, can be a tool of liberation or of the interior ministry — the infrastructure does not care which.

A country controlling its AI is not the same as its people controlling it. The same sovereign stack can serve citizens or surveil them — the wiring is identical.

And even where the state is not the villain, sovereignty can shade into cronyism. Swapping a distant foreign landlord for a domestic one who knows the right ministers is not liberation; it is the same rent, collected closer to home. The citizen still does not own the model, still cannot audit it, still has no say in how it is used.

The question that actually matters

All of which brings me to the question everyone chasing sovereign AI should be forced to answer out loud: does this actually return power to citizens, or simply swap one set of landlords for another?

Because that is the real test. Foreign firms captured the stack; now domestic firms or the state would like their turn. From the point of view of an ordinary person — whose data trained the thing, whose taxes funded it, and whose life it will increasingly shape — the nationality of the owner is not the point. What matters is whether they have any ownership, any oversight, any recourse at all. A monopoly does not become benign because it flies a local flag.

This is the through-line I keep tugging at across every technology I look at: watch who takes control, who pays for it, and whether anyone can fight back. Applied to sovereign AI, the honest scoreboard is uncomfortable. A great deal of what is sold as sovereignty is really a change of landlord — the public financing the asset and a narrow few owning it. A domestic landlord may at least be reachable by domestic law, but that is a long way from the promise.

What public-interest AI sovereignty could look like

I do not want to end on cynicism, because there is a version of this worth fighting for — different in kind, not just in slogan, from what most governments are building. Public-interest AI sovereignty starts from a different question. Not “how do we build a national champion?” but “how do we make sure the people who supply the data and the money hold a share of the result?” A few principles follow, each testable.

• Genuinely public infrastructure. Compute and foundational models funded by the public should be treated like roads or water: shared resources any local researcher, startup or public body can use on fair terms, not a moat handed to one favoured firm.
• Open where it counts. Models and datasets built with public money should be as open as safety allows, so anyone can inspect, correct and build upon them, rather than sealed inside a black box the public paid for but cannot see into.
• Data dignity. If a nation's data is the raw material, its people deserve a real say in how it is used and a share of the value it creates — consent and benefit, not silent extraction rebranded as national pride.
• Hard limits on the state. Sovereign capability must come chained to enforceable rules against surveillance and abuse. Independent oversight, transparency, and the right to challenge these systems are not optional extras; without them, “sovereign” just means unaccountable.
• Language and culture as a public good. Building models that genuinely serve local languages and communities is one of the most legitimate reasons to do any of this — and exactly the work private frontier labs have the least incentive to do well.

None of this is utopian. Countries have built public institutions to steward shared resources before — libraries, broadcasters, universities, public health systems — precisely because leaving certain things purely to private hands, or to the state's unchecked discretion, served people badly. AI is now important enough to deserve the same seriousness.

So I welcome the sovereign AI wave, with my eyes open. The instinct behind it — that no country should be a helpless tenant or a harvested field — is sound. But the instinct is not the outcome. Whether sovereign AI returns power to citizens or merely relocates the landlord depends on choices being made right now, in budgets and statutes and procurement contracts most people will never read. The vocabulary of national interest can house a real democratic gain or a fresh act of capture. It is on us to keep asking the only question that separates the two: who takes control, who pays, and can the rest of us fight back?