Part one of a series. Before arguing about who should own a country’s AI, it helps to define what is actually being claimed.
This post is a starting point for discussion — not policy advice. The framing here is our own and is subject to revision as the series develops.
AI Sovereignty · Part 1
"AI sovereignty" gets used as if it means one thing. It does not. At its plainest, it is a country's ability to develop, deploy, and govern AI without depending on a foreign power that could withdraw access, change the terms, or watch what passes through. But it is not a single switch you own or do not. It is a stack, and a country can hold some layers and rent the rest — usually without noticing which is which until a layer is pulled.
| Layer | What it means | What you lose without it |
|---|---|---|
| Compute | Data centres and accelerators on home soil — and access to the chips to build them. | Your AI runs at the pleasure of whoever owns the hardware and controls its export. |
| Models | The ability to build, fine-tune, and run foundation models you control. | Your capability is a licence, and a licence can be revoked or repriced. |
| Data | Training and inference data kept under domestic jurisdiction. | Your population's data trains someone else's system, under someone else's law. |
| Energy | The power, and the grid, to actually run the compute. | The data centres are ornaments. Sovereignty stops at the substation. |
| Talent | Domestic researchers, engineers, and operators. | Every layer above is rented expertise that can leave or be hired away. |
| Rules | The legal authority to set terms, audit systems, and say no. | Sovereignty becomes whatever the vendor's home government permits. |
Almost no country holds all six. The honest question is not "are we sovereign?" but "which layers do we hold, which do we rent, and which one being pulled would stop us." Part 2 takes that apart for a mid-sized country that will never build its own frontier model — and should not try.