Last updated · Fusionative
Short answer
Air-gapped AI is AI infrastructure, models and the platform around them running on a network physically or logically isolated from the internet. Everything, from container images and model weights to licenses, updates and monitoring, must work without reaching an outside service, and moves in only through controlled transfer.
What truly air-gapped requires
- No call-home dependency: licensing, telemetry and the control plane work offline.
- A private container registry and package mirrors inside the boundary.
- Model weights imported, scanned and verified before use.
- Updates delivered as signed bundles through approved transfer media.
- Local identity, certificates, logging and monitoring.
Deployment steps
- Prepare an offline bundle on a connected staging system: platform images, Helm charts, OS packages, drivers and models.
- Scan and sign the bundle, then transfer it through the approved process.
- Load images into the internal registry and install the platform from it.
- Import models into the internal catalog and approve them for tenants.
- Repeat the bundle process for each update; keep a record of what entered the boundary.
Where FusioNative fits
FusioNative's Deployment Engine installs from air-gapped bundles into a local Harbor registry with offline license files; the LLM gateway has no cloud egress by default; and FusionVault can export and import backups across the air gap. Tenancy, quotas, metering and security work as in connected installs.