Every month, another ministry or bank asks us the same question: can we have this without sending anything outside? The answer is yes — but the playbook looks nothing like a cloud deployment. This note describes what we have learned shipping air-gapped AI platforms.
Residency is a starting point, not the goal
Data-residency laws tell you where bytes must live. Real sovereignty is broader: who can update the models, who can read the logs, what happens when the vendor disappears. We design every offline deployment so the client can operate it alone for years — documentation, retraining recipes and admin tooling included. If your vendor cannot hand you the keys, you rented a dependency, not a platform.
Air-gapped does not mean out of date
The hardest engineering problem offline is not inference — it is the update path. We ship signed, versioned bundles that move across the boundary on approved media, verified inside before activation. Model weights, prompt catalogs and application code all follow the same disciplined route. Updates arrive on a schedule the security office controls, not whenever a cloud provider pushes.
Efficiency is a security feature
When compute cannot burst to a cloud region, every wasted token is capacity you do not have. Offline work forced us to make agents cheaper: tighter prompts, smaller specialist models, aggressive caching of intermediate results. The surprise is that these disciplines improved our cloud deployments too. Constraints are a good teacher.