Governments want to safeguard their tech. But not all AI should be sovereign

This year, many enterprises and governments moved from being dazzled by AI to being dependent on it. In the process, they have come to an alarming realization: they are just one vendor’s changed contract terms or foreign court order away from losing control of their critical infrastructure.

Many are now acting to secure the technology. Sovereign AI was a headline topic at the Annual Meeting in Davos last January, and governments around the world have since launched national AI strategies, funds, and data-centre build-outs of their own. As they commit capital and enterprises follow, one question keeps surfacing: how much sovereignty over AI do organizations actually need?

Our answer: less than the current all-or-nothing framing implies. A major risk is over-applying sovereignty; declaring every workload sovereign-critical works about as well as stamping every file top secret. Protection spread across everything protects very little, and it makes all of it slower and more expensive. Sovereignty is a precious resource. The winners of this build-out will be the ones who allocate it precisely to the workloads where control changes their real exposure.

The confusion starts with the word itself. Sovereignty has no single definition, and governments pursue it for different ends – some for security, some for ownership and value capture, some to build local ecosystems insulated from supply-chain disruption. One word carries all of these agendas, and the most expensive interpretation – full ownership of the AI stack, from chips and data centres to models and applications – tends to win by default.

Even when the agenda is clear, sovereignty bundles four separate kinds of control: where data and compute physically sit (territorial), who runs and secures the systems (operational), who owns the stack and its intellectual property (technological), and whose courts govern access (legal). These come apart in practice. A hospital network may need patient records kept inside national borders under domestic law, while running diagnostic models built and owned abroad. Collapse those four questions into one yes or no decision and you get over-classification. Once the choice is all or nothing, cautious institutions choose all.

There is no one-size-fits-all model. France’s Bleu, a private-sector venture, keeps operational control in French hands while a global hyperscaler supplies the underlying technology under agreed governance – sovereignty at the layers French requirements demand, global technology everywhere else. Others build the full stack: French providers such as OVHcloud and Outscale operate their own infrastructure end to end, often on open-source foundations, for customers in defence, government, and regulated industries where independence from foreign technology is the requirement. Both are legitimate answers; the right one depends on the workload.

Either way, sovereignty comes at a cost. Sovereign AI offerings cost 10 to 30% more than global alternatives, and migrations to sovereign environments typically take three to four years. Most enterprises have added sovereign AI to their 2026 roadmaps anyway, but few have a detailed strategy, budget or a way of sorting workloads by sensitivity.

The discipline we recommend is to classify each workload by its underlying data, following local and regional data-classification protocols, and apply sovereignty accordingly. The default is not-sovereign, with escalation only when the data’s sensitivity clears a defined bar. At national scale the stakes are higher. For defence, national security and heavily regulated industries, sovereign infrastructure is often a genuine requirement – the price of independence from foreign providers and the kill-switch risk that comes with them.

Where the stakes are lower, that same capital should be weighed against investment in talent, applications and adoption. For most workloads, a global platform is the more efficient choice, which is why it should be the starting point. Enterprises already behave this way elsewhere: about 70% say they would switch AI vendors for better pricing and performance. Save the strict requirements for the narrow set of workloads where a vendor’s changed terms or a foreign court order could do real damage, and protect those seriously.

McKinsey estimates that sovereignty could shape up to $600 billion in AI spending by 2030. All of it is meant to buy resilience, and resilience comes from the flexibility to move a workload or change providers. An institution that can do so under pressure is more agile than one that insists on owning the whole stack. That flexibility is won in the fine print: portability requirements, exit terms and interoperable standards negotiated with the same care as price. Infrastructure that cannot be maintained, staffed or upgraded simply replaces one dependency with another.

Heading towards Davos 2027, the governments and enterprises that woke up to their dependence this year must now decide how much of that dependence to reduce.

Putting off the decision until sovereignty has a clean definition would be a mistake, since an AI vendor changing its terms or a foreign court demanding access won’t wait for one. The best-protected institutions will be those that decide which specific forms of control are worth paying for, and let the rest go.