In June 2026 something happened that many had ruled out. A new frontier model launched, was celebrated, and three days later was no longer reachable. Not slower, not pricier. Gone. A single government order was enough, and access was cut for everyone.
For most people it was a footnote. For those who had built their work on this one model, it was a preview of a question that now occupies practitioners, analysts, and regulators at the same time. What actually happens to what my AI has learned when the model behind it disappears?
Models are the perishable part
The case was not a one-off. It was the most visible example of a pattern. Early in 2026 a popular model was retired on such short notice that a petition gathered more than twenty thousand signatures. Access prices shifted within weeks. A large software corporation blocked third-party AI agents from reaching the data that sits inside its own systems. Models come and go. Providers change rules, prices, and access faster than a company can adjust its processes.
Most people still consider themselves free. In a survey of several hundred executives, nearly nine in ten said switching provider was doable within a few weeks. Of those who actually tried, only a good four in ten called it smooth. Between perceived and real switchability lies a gap, and that is exactly where the risk sits.
The value is not in the model
The real error of judgement is the assumption that the model is the valuable part. The model is the replaceable part. What remains and grows over time is the layer above it: the process knowledge, the context, the memory that an AI accumulates in operation. You swap a model like a supplier. You do not swap the knowledge an organisation has built over years.
Here it gets uncomfortable. "Portability" today almost everywhere means only one thing: moving within the provider's world. Taking the accumulated memory with you in raw form and running it elsewhere is usually not foreseen. The provider keeps control.
Right now a race for standards around this layer is taking shape. The first open protocols for portable agent memory are appearing, research is describing the verifiable transfer of memory between different systems, and a new term is going around: authority over what a system may keep, revise, and forget. This movement is right. But it answers only half the question. Interoperability is not ownership. A format that exports anywhere is worth little if control over it stays with the provider. Standards without ownership end up cementing the incumbents, not your independence.
Sovereignty is a question of control, not of location
That shifts what sovereignty actually means. It is not the question of where your data sits. It is the question of who controls the layer that turns your data into value. Being hosted in a compliant way protects you from a fine. It does not protect you from having the ground pulled out from under your AI.
The current debate circles heavily around sovereign models and the place of operation. That is understandable, but it falls short. A model in your own data centre is still a replaceable model. The lasting value sits one layer above it.
The consequence is simple. Own the expensive and hard-to-replace, rent the interchangeable.
The model is the supplier. The knowledge layer is property.
Concretely, that means setting up this layer so that it runs on your own infrastructure, exports in open formats, and can, if needed, be handed over in full. That keeps you able to act, even when a model is switched off, a provider changes the rules, or your partner of today is no longer around one day.
This is also the point where "artificial" intelligence becomes something more durable. The value arises not in the tool, but in the interplay of human and machine on a structure the company truly owns. Collective intelligence on a cognitive architecture that stays when the model goes.
The real strategic question
The model that disappeared in June will not be the last. The real strategic question is therefore not which model you choose today. It is whether what your AI learns about your company belongs to you when the model goes.
This layer can already be built this way today. We consider it the most honest form of independence, and it is exactly what we work on.