Sovereign AI has four meanings, and vendors rarely say which one they sell

The phrase turns up in nearly every enterprise AI announcement now, including the Cohere and Aleph Alpha combination signed this week. Vendors use it to mean at least four different things, and procurement documents routinely blur the difference.

Tier one: data residency

The loosest version. Your model runs on servers inside your jurisdiction. That is it.

Data residency protects against a specific and real risk, which is your data physically sitting under another country's legal process. It protects against nothing else. The vendor still controls the software, the updates and the account.

Tier two: operational control

Residency plus a guarantee that vendor staff cannot access the deployment. No remote support sessions into your environment, no telemetry leaving it, no administrative backdoor for troubleshooting.

This is the tier most credible sovereign AI products actually sell, and it is a meaningful step up. It also depends entirely on the vendor continuing to exist and continuing to honour the arrangement.

Tier three: supply chain sovereignty

Now it gets harder. The model weights, the provenance of the training data, and ideally the chips underneath are not subject to another country's export controls.

This is the tier where the marketing usually stops matching reality, because the accelerators underneath nearly every serious deployment are designed by a small group of US companies and sit under the same export regime. The same dependency shows up in national chip investment programmes aimed at closing exactly this gap.

Tier four: full capability sovereignty

A country can train frontier models itself, end to end, with domestic compute and domestic expertise. Realistically fewer than five can.

This is the only tier that survives a genuinely hostile relationship with a supplier. It is also the tier almost nobody is selling, because it is not a product.

How to read the announcements

Most sovereign AI on sale is tier one or tier two. That is not a scandal. It is worth knowing which one you are buying, because the gap between hosted in Frankfurt and not dependent on a US company's continued goodwill is enormous, and a single phrase covers both.

The demand underneath is real regardless. Public sector buyers in Europe and Canada, along with regulated finance, healthcare and defence, all have reasons to care about where a model runs that have nothing to do with benchmark scores, and the same instinct is driving decisions like Cloudflare blocking AI crawlers by default.

So: when the next announcement lands, the useful question is not whether it is sovereign. It is which tier, and what happens to your deployment if the supplier relationship goes bad.