Sovereign AI Infrastructure Is a Jurisdiction Problem, Not a Hardware Problem

Discussions of sovereign AI usually begin with silicon. Where are the accelerators, who owns the datacentre, what is the power envelope. These are real constraints, and they are also the easiest ones to solve with money.

The constraint that money does not solve is jurisdiction.

Reachability beats residency

Data residency answers where bytes sit. It does not answer who can lawfully compel their production. A model hosted domestically but operated by an entity subject to foreign disclosure obligations is not sovereign in any operational sense — it is merely local.

Three questions worth asking

1. 01Which legal entity holds the operational keys, and under whose orders can it be compelled?
2. 02Are inference logs, prompts and embeddings treated with the same custody discipline as the source data?
3. 03Can the system be run, audited and updated without an external dependency that can be withdrawn?

Designing for withdrawal

Assume every external dependency will eventually be withdrawn, priced out or politically constrained. Systems that survive that assumption are built with substitutable model layers, portable weights where licensing permits, and an operating posture that does not degrade when a single supplier disappears.

That is the standard CARIO builds to, because it is the standard our clients are eventually held to.