Sovereign AI is increasingly what governments say they want. What they mean by it is usually compute.
This is the wrong framing. Compute is only one element in a much longer chain: energy, compute, data, training, deployment, governance. You can’t just look at one element in isolation. You need to look at the whole chain and be clear-eyed about the choices you are making at each stage.
At each link, a government should assess the market. Where it relies on others, it has three options.
1. Take the dependency and manage it. Not all risks are material.
2. Hedge through diversification, partnership, or contractual leverage. With multiple overseas partners, the disruption of any one relationship is survivable.
3. Build the capability domestically. This takes time and usually requires explicit or implicit subsidy.
Each choice should be deliberate. The pattern of choices across the chain shapes a country’s competitive landscape and its international partnerships.
The map below is one attempt at this exercise for France. It borrows loosely from Wardley, though it isn’t strictly Wardley mapping. It plots each link by how commoditised it is and how close it sits to the end user. Colour carries the sovereignty reading: teal is genuinely French or European, red is a critical foreign dependency, yellow is contested. Most of the chain is yellow or red.
Three observations from this:
Frontier model training is, for almost every country, take. The dependency on a small number of labs and the hyperscale infrastructure they run on is real. The honest move is to take it consciously and hedge it deliberately, rather than announce sovereignty and hope.
Talent is the link most countries can plausibly make, and the one most under-invested in relative to compute. Domestic talent depth is slower to build and more durable than a data centre. Take Canada. Their AI advantage was not built in server halls. It was built in university labs in Montreal and Toronto, over twenty years, by researchers who trained more researchers.
Procurement and deployment inside government is the link every government must make. No external partner can build a state’s capacity to specify, buy, integrate, and operate AI systems at scale. It is also where most strategies are quietly thinnest.
The point of the exercise is not the answers. It is forcing the choice. A government that has not decided, link by link, where it is taking, where it is hedging, and where it is making, does not have a strategy. It has an announcement.
Grid photo by Sebastiano Piazzi at Unsplash
