Sovereign AI · Architecture

Sovereign AI architecture

A sovereign AI architecture is a way of building AI so that the models, the data and the compute all stay under the control of the operator that runs them: open weights models, self hosted in the operator's own country, on hardware it owns or controls, with no third party AI service anywhere in the serving path. It is built in seven layers, from the systems of record at the bottom to the screens where a named person decides at the top. CodeNinja Atoms has published 28 worked designs in this architecture, across 11 sectors in the United States (14), Saudi Arabia (7), Pakistan (6), Germany (1).

The seven layers of a sovereign AI architecture

Read bottom up. Each layer runs on infrastructure the operator controls; the right column is read from the 28 published designs.

Sovereign AI architecture patterns

The same seven layers deploy in 3 ways. Each design is placed by rule from its own compute figure, so the pattern is the hardware the operator has to own.

What makes an AI architecture sovereign

• Open weights only. Every model in every one of the 28 designs is open weights, with its licence named. No closed model serves a request.
• In country. Inference, the object model and the data stay in the operator's country, on hardware it owns or controls.
• Read, never replace. The architecture reads the systems the operation already runs and writes back only what a named person approved.

Worked designs in a sovereign AI architecture

Newest first. Each is a complete design for one real operation, free to reuse under CC BY 4.0 with a DOI.

Built from the published designs only: designs.jsonl in the dataset and each paper's At a glance compute figure. Each design with its compute and three year cost, side by side, at sovereign AI reference architectures. Designed on Praxis; object models load into Hyper Ontology. Updated 2026-11-04.

Want one of these designs for your own operation? Talk to us.