What Is Sovereign AI? A Guide for Everyone Else

Sovereign AI is a country's or an organisation's capacity to develop, deploy and govern artificial intelligence using its own infrastructure, its own data, its own models and its own talent. The definition comes from Ali Ustun of McKinsey, and the important half of it is what he adds next: it is not about owning the technology. It is about retaining control over the AI life cycle, from the physical compute to the algorithmic logic.

That distinction matters more than it looks, because almost every explainer of this term is written for a reader who already has options. Consultancies write for large enterprises. Think tanks write for Washington. Glossaries write for enterprise IT buyers. Very little is written for the ministry, the island jurisdiction or the mid-sized research agency that will never build a national model and knows it.

This is for that reader.

Sovereign AI is not data sovereignty

These two get used interchangeably and they are not the same thing. Melanie Krawina, also at McKinsey, draws the line cleanly: data sovereignty is about the data sets, where they are stored, where they are processed and which legal jurisdiction applies. Sovereign AI is about who controls the intelligence built on top of them.

Ustun puts the consequence more bluntly. You can have data sovereignty and still not have sovereign AI, because if you rely on a foreign model to process your data, it is not sovereign AI at all. The data never left, and the decision-making did.

If that sounds like a technicality, consider what it does to a procurement decision. A residency requirement satisfies the first and says nothing about the second. It is worth understanding why sovereign AI and data sovereignty are not the same thing before writing either into a contract.

Why governments started caring

Ustun gives three reasons, and they are worth taking in his order because they escalate.

The liability squeeze

Courts are increasingly holding deployers liable for failures such as bias or hallucination, while vendors work to cap their own liability. Sovereign AI, in this reading, is a liability firewall. It is the difference between being able to audit a decision and being able only to forward a complaint to a supplier.

Geopolitical resilience

The second reason is dependence on a small number of global providers, and the vulnerability that creates around what Ustun calls kill switches: service denials or geopolitical conflict. This is the argument that has moved fastest, because it is the easiest to picture.

Where the economic value lands

The third is economic. If a country relies entirely on foreign API access, the value of that data flow is captured elsewhere. Krawina puts a number on the upside for Europe, roughly 480 billion euro of annual GDP impact by the end of the decade, and her colleague Luca Bennici estimates that around 40 percent of the value of AI is underpinned by sovereign or sovereign-enough solutions.

Those figures are estimates from one firm, and they should be read as such. What they establish is that the question is now a budget question rather than a philosophical one.

The part nobody says out loud

Here is what the explainers tend to skip. Almost nobody gets full sovereignty, and the arithmetic is not close.

The compute asymmetry

Georgetown's Center for Security and Emerging Technology puts it plainly: the United States controls about 75 percent of global AI compute and China accounts for about 15 percent. Everyone else divides the remainder. No amount of strategy documents changes that denominator in the near term.

Partial sovereignty as the normal case

CSET's argument is that policymakers lack a language for partial sovereignty, and that the absence of that language is itself the problem. Discussion collapses into a false binary between fully closed indigenous stacks and fully open globally interoperable ones, when every real AI stack is hybrid. The useful move is to name which layers you control and which you do not, and then argue about the right ones.

Lawfare's reporting reaches the same place from the other direction. Reviewing what states have actually built, it concludes that pursuing sovereign AI is mostly about managing dependencies rather than eliminating them. The Australia Foundation Models Report, quoted in the same piece, says it in one sentence: sovereign capability does not necessarily mean the whole model is developed and managed domestically, it is about the ability to manage how the model is used.

Read that as permission. A jurisdiction that will never train a frontier model can still hold real sovereignty over how models are used inside its borders, and that is not a consolation prize.

What a small state can actually control

Ordered by what is cheapest and most controllable first, which is the reverse of how these programmes usually get announced.

Rules and procurement

This layer is available immediately and costs comparatively little. It is also where most leverage sits, because procurement is how a government's preferences become a vendor's roadmap. Data localisation regimes already exist in several major jurisdictions, including China, the European Union under GDPR, Russia, India and Saudi Arabia, alongside instruments such as the EU AI Act and Canada's Artificial Intelligence and Data Act. Buying rules can specify auditability, exit rights and where a model may run, long before anyone builds anything.

Data and its jurisdiction

The second layer is the one most programmes start with, and it is genuinely important, provided nobody mistakes it for the whole job. See the section above on why residency alone does not deliver sovereign AI.

The application layer

Most of the value a citizen experiences is created here, in the specific systems that touch specific services, and this layer is far more ownable than the model layer beneath it. It is also where domain knowledge beats scale.

Talent and the operating model

The slowest layer and the one that compounds. It is the reason the question of who can operate and switch these systems on and off ends up mattering more than who wrote them.

Small jurisdictions have one structural advantage in all of this, which is that they can move a rule through an entire system quickly enough to observe the result. Islands make this unusually visible. A working group of small island governments has been finalising a sovereign data framework for climate adaptation built to stay under public ownership even when the compute itself is leased, which is partial sovereignty stated as a design requirement rather than a compromise. In Indonesia, Bali Province's digital residency sandbox grants a ninety day residency, a policy pilot lane and access to municipal data, letting founders test against real regulators instead of slide decks.

Neither is finished work. Both are examples of a jurisdiction choosing the layers it can hold, which is the point. There are good reasons why islands make unusually good regulatory sandboxes, and most of them come down to a short distance between a rule and its consequences.

The traps

The sceptical case deserves the same space as the enthusiastic one, and the World Economic Forum makes it best.

The first trap is the sovereignty trap. Governments pushing ahead independently risk undermining the international cooperation meant to keep the technology transparent and equitable, which could make the whole system less safe. The example given is uncomfortable and specific: a state using a system trained on local values for security purposes may readily flag behaviour out of step with those values as a threat.

The second is cost. Training a large language model has become dramatically more expensive, and the funds required for hardware, energy and staff may soon exceed a billion dollars. For most of the world that closes the model layer as an option, which is exactly why the layered reading matters.

The third is trust, and it cuts against the assumption that domestic means accepted. Roughly one in three people in the United Kingdom, France, Australia and South Korea say they are willing to trust AI, falling to around one in five in Japan and Finland. A sovereign system is still a system somebody has to be willing to use.

The WEF's counterproposal is a Global AI Compact, on the argument that compute is closer to electricity than to a luxury good, essential infrastructure that should not be out of reach. It is worth noting that nearly 150 years after the first light bulb, about 760 million people still have no access to electricity. Universal access is not automatic just because something is deemed essential.

Where this conversation is happening

The EX Future Summit runs a Government track covering exactly this ground: sovereign compute, digital residency and the island treated as a policy sandbox. It sits alongside tracks on research, finance, health technology and AI ethics, because sovereignty arguments rarely stay in one department.

The format is unusual and deliberate. The summit runs from 18 to 20 November 2026 as a single continuous thirty hour broadcast moving between Las Palmas in the Canary Islands and Bali, twelve hours apart, so participants can join from any timezone. Online attendance is free for verified researchers, students and the EX community. For anyone whose next decision is a procurement rule rather than a data centre, the practical question is usually what regulatory compliance actually demands, and that is the conversation the room is built for.

FAQ

Is sovereign AI the same as a national large language model?

No. A national model is one implementation of one layer. Singapore initiated its SEA-LION model in response to a stated strategic need for sovereign capability in large language models, but that sits inside a stack that also includes energy, chips, infrastructure and applications. Owning the model and controlling the stack are different achievements.

Can a country have sovereign AI without owning data centres?

On the partial sovereignty reading, yes. The Australia Foundation Models Report position is that sovereign capability is about the ability to manage how a model is used, rather than developing and managing the whole model domestically. Control over deployment, audit and exit can be real even when the compute is leased.

Which countries have pursued sovereign AI so far?

Denmark unveiled a national AI supercomputer. France is investing across its full stack and has been switching public sector cloud and software providers as a national strategy. Singapore has taken a hybrid path, integrating multiple technology ecosystems while investing in regional capability. These are the cases the sources above document; the list is not exhaustive.

Does sovereign AI make AI safer?

Not automatically, and it can work against safety. Divergent national development risks weakening the cooperation that keeps systems transparent and accountable across borders. Sovereignty is a question of who decides, not a guarantee of what they will decide.