The phrase "sovereign AI" gets used loosely. We propose a four-axis test — residency, key custody, model lineage, and operational classification — that can be answered yes or no for any deployment.
There is a difference between rhetoric and architecture. The phrase "sovereign AI" has been adopted across the public discourse with increasing breadth and decreasing precision, until it now signals approximately nothing.
For our practice, sovereignty is a four-axis test. Workload residency: where do training data, model weights, and inference traces physically reside? Key custody: who holds the encryption keys protecting model artefacts at rest and in motion? Model lineage: is there a signed manifest documenting training data provenance, fine-tuning steps, and evaluation rubrics? Operational classification: under which national classification regime does the system operate, and is the operator cleared to that level?
A deployment that answers all four axes affirmatively is sovereign. A deployment that does not is something else — and saying so explicitly is the first act of governance.
"Sovereignty is not a feature added after a model is built. It is the decision taken before the first line of training."