Most organizations that believe they have data sovereignty actually have a contract. Somewhere in a master services agreement is a clause naming a region, and everyone treats that clause as a wall. It isn't. It's a promise, and promises about where sensitive data lives are only as durable as the legal system, corporate structure, and political conditions surrounding the party who made them.
This rarely matters until it matters a great deal. A cloud region physically located in your country can still be operated by a company incorporated elsewhere, subject to disclosure demands under a foreign legal regime, and managed by a control plane that reaches in from another jurisdiction. Your data hasn't moved, but your control over it was never really there. The region was geography. Sovereignty is authority, and the two were allowed to drift apart.
For consumer applications, this gap is tolerable. For a power grid, a defense system, a national health record, or a central financial rail, it is not. As AI moves into that kind of infrastructure, the workloads depending on third-party control are no longer marketing databases. They are the systems a country runs on, and the stakes of the promise-versus-boundary distinction rise with them.
The fix is not a stronger clause. It's moving enforcement from the contract layer to the infrastructure layer. When residency is a property of the architecture, meaning where compute physically sits, which control plane governs it, and who holds the cryptographic keys, then sovereignty becomes something you can demonstrate rather than something you're assured of. The test is simple. Can you prove, right now, where a given workload runs and who can compel access to it? If the answer lives in a legal document rather than a system you control, you have a promise. The organizations that will weather the next decade of AI in infrastructure are the ones already turning their promises into boundaries.