AI Sovereignty Is a Procurement Question, Not a Philosophy Seminar
European boards have started asking me about “AI sovereignty,” usually in a tone that suggests they expect a geopolitics lecture. I give them a procurement checklist instead, because that is what sovereignty actually is at the level of a single company: the ability to keep operating when access terms change.
And access terms do change. The frontier models are built by a handful of companies under one government’s jurisdiction, and that government has already shown it treats advanced computing as a policy lever — export controls on chips, release restrictions on frontier systems. You do not need a prediction about which rule changes next. You need to notice that your dependency is a policy exposure, and price it like one.
The test I use with executive teams is one sentence: if your primary model vendor became unavailable in ninety days — embargo, ban, price shock, shutdown — what would break, and for how long? The uncomfortable silence that usually follows is the real audit result.
Getting to a good answer is not exotic. It is five capabilities, all buildable in a quarter:
- An abstraction layer, so the model behind your products is a configuration choice, not an architecture.
- Portable scaffolding — MCP connectors, harnesses, prompts — that works across vendors.
- An eval suite, which is your switching insurance: with one, re-qualifying a substitute model takes days; without one, it takes a leap of faith.
- A tested open-weight fallback for the workloads that must survive anything. The open-weight frontier is now close enough that this is a real option, not a gesture.
- Clarity about which assets are durably yours: your data, your knowledge graphs, your evals, your redesigned workflows. Models are rentals. Those are property.
Notice what is not on the list: abandoning US frontier models. That would be trading capability for a purity that helps nobody. The point is the opposite — you can engage the frontier labs confidently precisely because you have made yourself hard to hold hostage. Dependence you chose, with an exit you rehearsed, is a strategy. Dependence you drifted into is just risk with good UX.
Treat models as resources: valuable, priced, and replaceable. Save your loyalty for the assets that can’t be revoked by an election.
The full argument, including the partnership and vendor-narrative sides, is Section 13.6 of the book.