From Foundation Models to Sovereign Models

Most organisations using artificial intelligence today are tenants. They do not own the intelligence they rely on. They rent access to it, by the token, through a pipe, from a handful of providers who hold the weights, the infrastructure and the right to change the terms. This is rarely put so plainly, because the language of the industry is built to obscure it. We talk about platforms, partners and ecosystems. What we mean is that someone else owns the thing, and you are pay

To own intelligence you must own the weights. The weights are the model. Everything else, the prompts, the retrieval, the orchestration, the interface, is scaffolding around a core of numbers that encodes what the system knows and how it reasons. If you do not hold those numbers, in a form you can load, inspect, modify and run, you do not own the intelligence, however clever your scaffolding.

The dominant pattern in artificial intelligence today is the single giant generalist, one enormous model asked to do everything from poetry to protein folding. It is an extraordinary engineering achievement and, for a sovereign operator, the wrong shape. One model that does everything is one model you must trust with everything, one boundary that holds every secret, one failure that touches every function. We chose a different architecture deliberately.