Air-gapped AI
Air-gapped AI is an AI system that operates entirely inside a network with no external connectivity, meaning model inference, data access, and licence validation all function without any call leaving the perimeter.
The term is often used loosely. A system that keeps data on-premise but calls a hosted model API is not air-gapped; neither is one that phones home to validate a licence, check for updates, or report telemetry. A single outbound dependency in a critical path is enough to break the property, and it usually surfaces at the worst moment — during an outage or a security review.
Running genuinely offline changes several design assumptions. Models must be hosted locally, which constrains model choice to weights that can be deployed inside the environment. Licensing has to be verifiable without a callback, which generally means cryptographically signed licence files rather than server checks. Updates become an operational process rather than an automatic one.
Organisations adopt the pattern for different reasons: classification requirements, data residency obligations, contractual restrictions on where customer data may be processed, or simply an unwillingness to send sensitive material to a third-party endpoint.
What distinguishes it
- No outbound network calls in any critical path, including licensing and telemetry
- Model inference runs on infrastructure inside the perimeter
- Licence validation works offline, typically via signed files rather than server checks
- Updates are deliberate operational events rather than automatic background fetches
In DataQube
DataQube is designed to run fully air-gapped: the agent runtime and model inference sit inside the customer's environment, and licensing is validated from signed licence files with no check-in. The same architecture runs on-premise, in private cloud, and disconnected — the deployment boundary changes, the behaviour does not.
Common questions
Is on-premise the same as air-gapped?
No. On-premise describes where software runs; air-gapped describes whether it can reach the outside world. On-premise software that calls a hosted model API or validates licences over the network is not air-gapped, even though it runs in your data centre.
Can you use commercial language models in an air-gapped deployment?
Only where the weights can be deployed inside the environment. Models available exclusively through a hosted API cannot function without outbound connectivity, so air-gapped deployments are limited to models that can be self-hosted.
What usually breaks the air gap in practice?
Licence check-ins, telemetry, automatic update checks, and hosted model endpoints. These are easy to overlook because each seems minor in isolation, and each is typically discovered only when the network is genuinely disconnected.
Related terms
AI data agent
An AI data agent is software that answers analytical questions by planning and executing its own steps — querying approved databases, running code, and citing what it used — rather than only generating text.
Concurrent-session licensing
Concurrent-session licensing charges according to how many users are active at the same time, rather than how many named accounts exist, so occasional users do not each consume a permanent seat.
Data provenance
Data provenance is the recorded lineage of a result: the sources, queries, transformations, and decisions that produced it, retained in enough detail that someone else can verify the figure independently.