Sovereign AI is procurement, not patriotism

Digital sovereignty gets argued as politics and decided as purchasing. Nobody signs a cloud contract because of a flag; they sign it because of price, capability and how much work it would be to leave. Which is good news — it means the sovereign case has to be made in the language of procurement, and in that language it is much stronger than the patriotic version.

The three questions that actually decide it

Strip away the geopolitics and a sovereignty decision comes down to three unglamorous questions about each system you depend on.

Who can switch it off? Not “who would”, but who has the technical and legal ability. A supplier acting on its own terms of service, a court in another jurisdiction, a government using export controls or sanctions. If the answer includes anyone outside your contract chain, that is an operational risk, not a political opinion.

What does leaving cost, in weeks? Not licence fees — weeks. Data export in a usable schema, re-integration, retraining, the parallel run. If nobody in the company has ever tested it, the honest answer is “we do not know”, which is itself the finding.

Where does the processing physically happen? Storage location is the question people ask. Processing location is the one that matters, because that is where the data is readable and where compulsion works.

Why the European rules changed the arithmetic

For years the sovereign option carried a capability penalty and no compensating benefit. That has shifted, and not because of speeches.

The EU Data Act made switching between cloud providers a regulated process rather than a commercial favour, including the phasing out of the egress charges that quietly made leaving unaffordable. The AI Act put documentation, logging and human oversight obligations on certain uses — obligations that are much easier to satisfy when you can actually see inside your own stack. And data protection law has never stopped asking where processing happens and under whose jurisdiction.

Timelines and details in this area have moved more than once — verify the current text before making a decision on it.

The point is not that regulation forces sovereignty. It is that regulation has removed some of the friction that made staying the automatic answer.

Sovereignty is a spectrum, and the middle is where the work is

The debate is usually framed as a binary: hyperscaler or your own basement. Almost every real answer sits between those, and the useful exercise is mapping your stack layer by layer.

• Data — never rent. This is the layer that compounds and the one you cannot recreate.
• Models — open weights where the capability is core to what you sell; hosted APIs where it genuinely is not.
• Compute — European provider, colocation or your own hardware, chosen on jurisdiction and cost rather than on ideology.
• Orchestration — the system of action that runs your operations. If it is your process, own it.
• Everything else — rent it, on short contracts, with the exit actually tested once a year.

The uncomfortable part

Sovereignty costs something. It is usually not money — hardware and open-weight inference have both got dramatically cheaper — but attention. Somebody has to own versioning, updates, backups and the boring discipline of checking that things still work.

That is real, and it is exactly the work you are paying a vendor to hide from you. The question is whether hiding it is worth handing over the ability to change your own business. For a company whose operations are its product, the answer is usually no.

For everything else, rent away. Just know which is which — and write the answer down before the renewal, not after.