Sovereign AI is mostly a mirage — and Asia keeps chasing it

I have now sat through a dozen government presentations about sovereign artificial intelligence, in four countries, and they tend to follow the same script. A minister announces a national large language model, trained on the local language and local values. A render of a data centre appears. A large number is attached. Everyone applauds. And almost none of it, I have come to believe, will deliver the sovereignty it promises.

I want to be careful here, because the instinct behind sovereign AI is not foolish. No country should want its public sector, its courts and its strategic industries running entirely on models it cannot inspect, controlled by foreign firms answerable to foreign governments. The anxiety is legitimate. It is the prevailing response to it that is mostly theatre.

Across Asia, governments are pouring public money into training their own flagship models, and calling the result sovereignty. In most cases what they are buying is an expensive symbol — a model that lags the global frontier, runs on imported chips, and addresses almost none of the dependencies that actually matter. We are confusing having a model with controlling the stack.

What sovereignty is not

A national language model, on its own, is the least sovereign part of the whole enterprise. The weights can be impressive and still leave you dependent on foreign hardware to train them, foreign cloud platforms to run them, and a foreign-dominated supply of the specialised chips without which none of it functions. You can own the model and control almost nothing.

Consider where the real chokepoints sit. The advanced processors come overwhelmingly from a handful of firms and are manufactured in a single concentrated geography. The cloud infrastructure that serves models at scale is dominated by three American companies. The frontier research talent clusters in a few cities, most of them not in your country. A flag on a model file does not touch any of these.

Training a national model and calling it sovereignty is like minting your own currency while importing the printing presses, the ink and the paper. You have produced a symbol, not independence.

The symbolism is not worthless. There is genuine value in models that understand local languages, scripts and contexts that global firms underinvest in, and in building domestic teams that know how to train and deploy them. But that is a capability argument, not a sovereignty one — and conflating the two leads to spending vast sums on the wrong layer of the problem.

What it would actually take

Real AI sovereignty, if a country genuinely wanted it, would be a far less glamorous and far more expensive project. It would mean securing compute capacity through long-term arrangements that survive geopolitical shocks. It would mean domestic data-centre infrastructure and the power to run it, which in much of Asia means solving a grid problem before an AI one.

It would mean investing in the unglamorous middle of the stack — the tooling, the inference infrastructure, the deployment platforms — rather than the headline-friendly model itself. It would mean keeping and attracting the small number of people who can actually build these systems, which is a visa and salary problem as much as a research one. None of this makes for a good ribbon-cutting.

Above all it would mean honesty about dependencies that cannot be eliminated, only managed. No medium-sized economy is going to manufacture its own frontier chips or replicate the global cloud. Sovereignty, realistically, is not autarky. It is having enough capability, enough alternatives and enough leverage that no single foreign actor can switch you off. That is a portfolio strategy, not a flag-planting exercise.

The better question

So what should governments do instead? Start by asking what they are actually afraid of. If the fear is that a foreign provider could cut off a critical service, the answer is redundancy and contractual leverage, not a homegrown model. If the fear is that global models ignore local languages and values, the answer is targeted investment in exactly those gaps, openly described as capability-building rather than dressed up as independence.

Pooling helps too. Smaller Asian economies chasing individual sovereign models are each too small to matter at the frontier. Shared compute, shared research investment and shared infrastructure across friendly states would buy more genuine leverage than a dozen underpowered national models competing for the same scarce chips and talent. Sovereignty, paradoxically, may be more achievable in coalition than alone.

I do not write this to mock the ambition. The desire not to be wholly dependent on foreign technology for something this consequential is one I share. But we owe that ambition more than a render and a big number. If sovereignty is the goal, then the model is the easy, visible, least important part — and the moment a government understands that, it can stop buying symbols and start buying the thing it actually wants.