Sovereign AI: why collaboration beats control

Sovereignty has two dimensions. The first is the technology stack: chips, energy, compute and storage, accelerators, frontier models, data, applications, and the governance binding them together. The weakest link decides, so controlling one layer proves nothing. The second is time, and it gets discussed too rarely. Infrastructure runs unaided for a while, then software updates, licences and supply chains reassert themselves. Short-term survivability is one thing; long-term feasibility is another. At Telia we looked hard at running independent of outside dependencies. It is possible, and it is not sustainable indefinitely, because at some point you need a patch and the patch comes from a partner in another country. There is no absolute setup, only a matrix of stack against time, and a judgement about where you are most vulnerable.

This technology cycle moves differently from the ones before it. AI has a long history reaching back to the fifties, and machine learning has been with us for years. What is new is the speed. Coding is largely solved. Systems have begun to improve themselves, which bends the curve again. An enterprise carries legacy and regulation around its neck that an AI-native startup never picks up.

Asian operators hold an advantage that European ones have lost: choice. We can source from East, from West and locally. A European operator now struggles to defend buying Chinese equipment, and at Telia a China-free network was a condition of the Swedish 5G licence. Choice is worth engineering for. Build the architecture modular enough to swap one cloud for another and you gain leverage in a negotiation as well as resilience in a crisis. The third leg is owning the IP and the infrastructure for whatever is genuinely critical.

On models, sequence the spend. Start with a frontier model and acquire the tokens, because that is how you learn what the use case needs. As production scales, token cost starts to matter, so slim the model down or run it on your own cluster. The same pattern held in cloud: some workloads belong on a hyperscaler for the tooling and the developer ecosystem, while core switching workloads were always cheaper to run ourselves, on economics alone, before sovereignty entered the argument. Never build an application tied to one model.

The moat is context: the data, the customer relationships, the trust already earned. Which cloud or which model runs underneath decides nothing about competitive position.

Three claims, judged. That every operator without a sovereign AI play in five years will be left behind is noise. Demand has to come from the market, and an operator no more needs a sovereign stack than it needs its own network to be a successful MVNO. Market leaders will build these, and more than one per market, but not everyone. That compute becomes a commodity while data becomes king is signal, and it carries an architectural consequence: never let data pocket itself inside an application or a third party, and never outsource data management. That hyperscaler partnership becomes the industry norm is fifty-fifty. The benefit is real and the dependency must not be single, because frontier leadership rotates between labs.

Capability is rising exponentially and adoption is rising linearly. The gap between what is possible and what is real keeps widening, in our own operating companies as much as among our enterprise customers. That gap is lost value, and closing it is itself a business.