The topic of AI sovereignty has come from me thinking about what defines choice or independence in the context of AI workloads. Restrictions on closed and open weight models keep tightening, and I wanted a definition of independence I could measure.
A firm that buys an enterprise memory product, holds a perpetual license, and runs it on its own hardware owns the product. Then the license turns out to permit running it only inside the vendor's runtime, or the tuned artifacts your data produced are licensed for use with the vendor's models. You own the software and rent the permission to use it.
Owned or rented tells you what happens when you stop paying. Portable or captive tells you whether it can move. The two are independent, and owned plus captive is the combination that fools people, because owning feels like safety and it is easy to miss whether what you bought can be moved.
So I judge AI sovereignty by portability rather than ownership. Portability may cost more than taking the vendor's default, and that premium gives you the option to leave. Portable means three things, and missing one is enough to fail:
- Artifacts are generated in a form other tools can read.
- You can acquire infrastructure to run it.
- The license permits running it elsewhere after the relationship ends.
The first two are technical facts you can test today. The third is a contract that renews, so you read it again at every signature, even when nothing technical has changed.
I have built small versions of this for myself. A memory layer with provenance in the schema and an executor that holds the credentials, both deployed in containers. I built them to be moved, and that is the bar I set whether the implementation is small or enterprise.
If your primary AI vendor changed its terms next quarter, what would you still have?
Written by Duane Grey
AI Strategy & Implementation
Independent AI consultant helping companies cut through hype and deploy systems that produce real results.