The brief sounds simple: give engineers modern AI tooling in an environment where no packet crosses the perimeter. No SaaS APIs, no telemetry, no 'phone home for licence validation'. What follows is how such a platform has to be designed when the internet is not available.
The constraints define the design
Air-gapped is not a deployment option you toggle at the end — it is the first architectural input. Every component is auditioned with one question: does it degrade gracefully when the internet does not exist? Much of the modern AI stack fails that audition on the first day, which is precisely why the components that pass are worth writing down.
The model supply chain, offline
Outside the gap, models arrive from a registry. Inside, they arrive like any other controlled artefact: through a transfer zone, with checksums, signatures, provenance records and a named human accountable for the import. We treat model weights exactly like container images — scanned, signed, versioned, and promoted through environments from a single internal registry.
Inference at the edge of the network
Self-hosted inference is a capacity-planning discipline. Without an elastic cloud behind you, quantisation strategy, batching and cache hit rates stop being optimisations and become the difference between a platform and a queue. Build the observability first: tokens per second, queue depth and per-team consumption become the dashboards that matter within a week.
Operating without phoning home
- Everything reconciles from an internal Git remote — the delivery model works identically with the cable cut.
- Licence-checking software is a liability class of its own; prefer components that do not need permission to boot.
- Updates travel the same signed transfer path as models — one ceremony, every artefact.
- Documentation lives inside the gap, versioned with the platform. An offline platform with online docs is offline twice.
The cultural effect is easy to underestimate: constraints produce clarity. With no vendor roadmap to wait for, the platform team owns every decision — and the architecture is stronger for it.