AI & Data

Air-Gapped AI: How Private AI Systems Work Without Cloud Connectivity

How private AI deployment works in isolated networks: architecture, model updates, security controls and the real trade-offs of offline AI systems.

· 10 min read · DTN Editorial

Most AI services people use every day run in someone else’s data centre. For defence organisations, critical infrastructure operators and many industrial users, that model is often unacceptable: the data is too sensitive, the network is not always available, or the rules simply forbid it. Air-gapped AI is the answer to a simple question — can useful AI run entirely inside an isolated environment, without any connection to the cloud?

It can, but not for free. This guide explains what air-gapped AI is, how private AI deployment works in practice, where offline AI systems are strong, and where they are limited. It is a general, vendor-neutral overview.

What is air-gapped AI?

Air-gapped AI is an AI system that runs on computing infrastructure isolated from external networks. The model files, the software that runs them, the data they process and the outputs they produce all remain inside a controlled boundary. There is no live link to the internet, to a model provider’s API or to a vendor’s telemetry service.

The term “air gap” originally meant physical separation: no cable or wireless path between the protected system and the outside world. Today it is also used more loosely for logically isolated environments with tightly controlled, often one-way, transfer points. When evaluating a solution, it is worth asking exactly which of these is meant.

Air-gapped AI sits at the strict end of a spectrum usually described as private AI: AI that an organisation hosts and controls itself, rather than consuming as a public cloud service.

Why organisations choose air-gapped AI

  • Data sensitivity: classified, personal, commercially confidential or export-controlled information may not be sent to external services.
  • Sovereignty and jurisdiction: organisations may need assurance about where data is processed and which legal regimes apply.
  • Disconnected operations: ships, remote sites, deployed headquarters and some industrial facilities cannot rely on continuous connectivity.
  • Reduced attack surface: removing network paths removes whole classes of remote attack, although not all risks.
  • Predictability: a fixed, locally hosted model version behaves consistently and is not changed silently by a provider.

Cloud vs on-premise vs air-gapped AI

The three deployment models are best seen as points on a scale of control and convenience:

Cloud, on-premise and air-gapped AI deployment comparedThree columns: public cloud AI with data leaving the organisation; on-premise AI with data in the organisation's own facility but network connected; and air-gapped AI with no network connection, updates via controlled media transfer.CLOUDON-PREMISEAIR-GAPPEDProviderdata centreOwn facilitynetworkedIsolatedenclaveInternetrequiredInternalnetworkNo externalnetworkUpdatescontinuousUpdatesmanagedUpdates viacontrolled mediaMore data control, less exposureEasier updates, monitoring and scale
Figure. Three deployment models. Moving right increases control over data and reduces exposure, but makes updates, monitoring and scaling harder. Simplified comparison.
  • Cloud AI: the provider hosts models and infrastructure. Fast to adopt and easy to scale, but data leaves the organisation and depends on connectivity.
  • On-premise AI: models run on the organisation’s own servers, connected to its internal network and possibly, under control, to the outside. Data stays in-house; updates and monitoring remain relatively easy.
  • Air-gapped AI: no external network path at all. Maximum control over data, at the cost of harder updates, monitoring and scaling.

How an offline AI system is built

A typical offline AI system has the same building blocks as any AI service, rearranged so that every dependency lives inside the boundary:

Conceptual private AI architecture inside a controlled enclaveApproved data sources enter through a controlled import gateway into a private enclave containing data validation with provenance tagging, locally hosted AI models, and a shared situational picture. A human reviewer approves outputs, and an audit log records all steps. No connection to public cloud services.APPROVED SOURCESReportsSensor feedsLogisticsControlled import gatewayscanning · labelling · one-wayPRIVATE ENCLAVE · NO PUBLIC CLOUDValidation & provenancesource · time · confidenceLocal AI modelssummarise · correlate · flagShared situational picturewith uncertainty shownHuman review & approvalAudit loginputs · model version · decisions
Figure. Conceptual private AI architecture: data enters through a controlled gateway, models run locally inside the enclave, and people review outputs. Illustrative concept, not a depiction of a deployed system.
  1. Controlled import: data and software enter through a defined gateway where they are scanned, labelled and approved.
  2. Data layer: local storage with access control, classification labels and provenance metadata recording where each item came from.
  3. Model serving: locally hosted models — for example language models for summarisation, or classifiers for structured data — with pinned versions.
  4. Retrieval: many private AI deployments use retrieval-augmented generation, where the model answers using documents retrieved from a local index, so responses can cite internal sources.
  5. Applications: user interfaces, dashboards or integrations with existing systems.
  6. Logging and audit: records of inputs, model versions, outputs and user actions, retained inside the boundary.

Every external dependency must be removed or replaced: licence servers, package repositories, time services, identity providers and telemetry. Overlooking one of these is a common reason pilot projects stall.

Models and hardware

The growth of openly licensed models has made private AI deployment far more practical than a few years ago. Organisations can choose model size to fit hardware: larger models generally need more GPU memory and power; smaller or compressed (quantised) models can run on a single server or a rugged edge computer at lower capability.

Hardware choices are driven by the environment: data-centre servers in a fixed facility, or ruggedised computers with power, cooling and size limits in vehicles and deployable shelters. Licensing terms for each model and component should be checked for the intended use, including any restrictions on defence applications.

Updating AI without connectivity

Air-gapped systems still need updates: security patches, improved models, new reference data. Without a network, updates become a deliberate process:

  1. Prepare the update package in a separate, controlled environment.
  2. Verify integrity with cryptographic hashes or signatures from a trusted source.
  3. Scan for malware and review the bill of materials for software and model components.
  4. Approve through change management.
  5. Transfer on authorised media or through a one-way transfer device.
  6. Test inside the boundary before release, and keep the previous version for rollback.

This is slower than cloud updates. Organisations should decide in advance how often models will be refreshed and who is accountable for each step.

Security: the air gap is not enough

Isolation is a strong control, but it is not a complete security model. Removable media, insiders, compromised hardware or software supply chains, and poisoned training data can all reach an isolated system. AI adds specific risks: manipulated model files, prompt injection through imported documents, and leakage of sensitive data through model outputs.

The joint Guidelines for secure AI system development led by the UK NCSC cover secure design, development, deployment and operation of AI systems. Principles from NIST’s Zero Trust Architecture — authenticate and authorise every access, assume breach — remain relevant inside the boundary too.

Limitations and trade-offs of offline AI systems

  • Capability: locally hostable models may be less capable than the largest cloud models.
  • Freshness: models and reference data age between updates.
  • Operations: the organisation carries the cost of hardware, power, specialists and lifecycle management.
  • Monitoring: vendors cannot see problems remotely; local monitoring must be designed in.
  • Accuracy: isolation does not make outputs more correct. Language models can still produce confident but wrong answers.

Governance and human oversight

Who may use the system, for what, and with which data? How are outputs checked? The NIST AI Risk Management Framework offers a structure — govern, map, measure, manage — that applies equally to offline deployments. In the defence context, NATO’s Principles of Responsible Use emphasise accountability, explainability, traceability, reliability and governability. In the EU, note that the AI Act excludes systems used exclusively for military, defence or national security purposes from its scope, but dual-use and civil deployments may fall within it.

A practical rule: AI outputs in sensitive environments should be decision support, reviewed by an accountable person, with sources and confidence visible. For an example of how this applies to shared situational awareness, see our guide to the common operational picture. For private AI in everyday enterprise work such as document search, see military software development in the AI era.

Evaluation checklist for private AI deployment

  • Is the isolation physical, logical, or both? Where are the transfer points?
  • Are all dependencies (licences, identity, time, packages) available offline?
  • Which models are used, under which licences, and what are their known limitations?
  • How are updates verified, approved, transferred and rolled back?
  • What is logged, and who can review the logs?
  • How are outputs evaluated for accuracy on representative data?
  • What hardware, power and cooling are required in the target environment?

For how private AI fits into wider defence information architectures, read C4ISR systems: architecture, data integration and the role of AI and AI orchestration in battle management systems. DTN’s Private AI for Defence Situational Awareness concept page describes a proposed proof of concept built on these principles.

Frequently asked questions

What does air-gapped AI mean?

Air-gapped AI is an artificial intelligence system that runs on infrastructure physically or logically isolated from external networks, including the internet and public cloud services. Models, data and processing all stay inside the isolated environment.

Can large language models run offline?

Yes. Openly licensed models of various sizes can run on local servers or workstations without any internet connection, provided the hardware has enough memory and compute. Smaller models are less capable, so the right choice depends on the task and the hardware available.

Is an air-gapped AI system automatically secure?

No. Isolation reduces exposure to remote attacks, but insiders, removable media, supply-chain compromise of models or software, and misconfiguration remain risks. Air gaps are one control within a wider security architecture.

How are air-gapped AI models updated?

Typically through a controlled transfer process: new model files, software and data are prepared outside, verified with checksums or signatures, scanned, approved, then moved in on authorised media or through a one-way transfer device, and tested before release.

What is the difference between private AI and air-gapped AI?

Private AI is a broader term for AI that runs under an organisation's own control, for example on-premise or in a dedicated environment. Air-gapped AI is the strictest form of private AI deployment, with no network path to the outside.

Sources and further reading

Private AI for situational awareness

Explore a proof-of-concept discussion

See how DTN frames private, human-reviewed AI for secure information integration — demonstrated with synthetic data only.

Related insights