AI 101 — What is sovereign AI?

Sovereign AI is artificial-intelligence capacity a country builds or buys under its own control — data centres, chips, models, training data — instead of renting it from a company based somewhere else. The word "sovereign" is doing real work: it means the government, not a foreign vendor, decides what runs, on whose hardware, and under whose law.

The Center for a New American Security, which tracks the field project by project, defines a sovereign AI initiative as government-backed and tied explicitly to national strategic interests, with material public money behind domestic compute, models, or data (CNAS Sovereign AI Index). That definition matters, because it excludes the many companies that simply describe their ordinary cloud region as "sovereign."

Why it matters right now

The starting fact is concentration. Ninety percent of the computing power needed to develop and deploy frontier AI sits in the United States and China, and companies in those two countries have built essentially every top-ranked foundation model, per the same CNAS index. If you are the government of France, India, Saudi Arabia or South Korea, that is a strategic dependency you did not choose.

So they are buying their way out. CNAS counts 184 sovereign AI projects worldwide, and the spending is lopsided: 59 percent are infrastructure — data centres, supercomputers, GPU clusters — 32 percent are model projects, and only 9 percent are data projects. The Middle East and East Asia account for more than 80 percent of all publicly disclosed sovereign AI investment.

The specific commitments are concrete. Canada has committed CAD 1 billion to build sovereign compute capacity, plus a CAD 300 million access fund that subsidises small firms and researchers, according to the OECD's review of national AI compute policy. India's IT minister, Ashwini Vaishnaw, said the country's shared public compute pool already holds 38,000 GPUs and would add 20,000 more, the Hindu BusinessLine reported.

The mental model: three layers, three different answers

Break AI into three layers and the politics get much easier to follow. Compute is the chips and the buildings — expensive, physical, and the layer governments can most plausibly own, which is why it takes the biggest share of projects. Models are the trained weights and the licence attached to them: do you get to run it after the vendor changes its terms? Data is the corpus, including the part that makes a model good at your language and your institutions.

Sovereignty is decided per layer, not all at once. A country can own the building, rent the model, and build its own dataset — and that mix is what most projects actually are. The honest test for any claim of sovereign AI is one question: if the foreign API contract were cancelled tomorrow morning, what would still work?

The analogy: the national grid

Think of electricity. No country generates every electron it consumes, and nobody thinks that makes the grid fake. What makes a grid national is that the switch, the regulator and the reserve capacity are domestic — the country can keep hospitals running when an outside supplier has a bad quarter. Most countries still buy their turbines from abroad.

That is roughly where AI is heading. A Brookings report on AI sovereignty argues that full-stack independence is structurally infeasible for almost any country, because the AI supply chain runs through minerals, energy, chipmaking equipment, networks and talent spread across many borders. Its recommended alternative is what the authors call "managed interdependence": map your dependencies layer by layer, then choose deliberately where to own and where to ally (Brookings).

Four things "sovereign AI" does not mean

It does not mean building a frontier model from scratch. Only a minority of model projects train from zero. Most take an open-weight model and fine-tune it on local language and domain data — cheaper by orders of magnitude, and the reason open-weight models became a policy instrument. In the CNAS index, Meta's Llama is the disclosed base for 14 of the 25 model projects that name one.

It does not mean the data stays home. A "sovereign cloud" is often a US vendor's hardware inside a national building, operated under a local contract. More than three-fifths of tracked projects disclose at least one foreign partner, and four-fifths of those partners are American. Dependence does not disappear; it moves down a layer, from the model to the chip.

It is not the same as privacy or safety. Sovereignty is about who can switch the system off, not whether the system is accurate or fair. A government-run model can be just as error-prone as a rented one — the rules governing it are a separate question entirely.

It is not free. Compute is a building-sized capital expense with an energy bill attached, which is why the same governments buying national AI data centres are simultaneously negotiating deals with the US hyperscalers they say they want distance from.

Where to learn more

The CNAS Sovereign AI Index is the best public dataset — it lists individual projects by country, layer, budget and operating status, and it updates. Brookings' paper is the clearest short statement of the sceptical case. The OECD's compute review is the place to look for what governments are actually appropriating rather than announcing.

The short version: sovereign AI is a country deciding which layer of the stack it cannot afford to rent. Almost nobody gets all three — and the ones who claim to usually mean the building.

Does a government-owned data centre running a foreign company's model count as sovereign? Write to us via our contact page.