What Is Air-Gapped AI? A Plain-English Definition
Air-gapped AI is an AI system that runs on hardware with no physical connection to any other network, so data enters or leaves only by hand. The NIST glossary, citing CNSSI 4009-2022, defines an air gap as an interface where systems are not physically connected and any logical connection is not automated. That is a stricter idea than offline AI, which only needs to work without a connection.
How does an air-gapped LLM work in practice?
At runtime, an air-gapped LLM behaves like any local AI setup: weights, inference server and data sit on your own machines. What changes is the supply chain. Everything the system needs, from model files to patches, has to arrive through a controlled, manual transfer.
Think of it as three supply lines:
- Models. Download on a connected machine, carry across on approved media, drop into the runtime's model directory. Ollama reads models from the folder set in
OLLAMA_MODELS(Ollama FAQ). NVIDIA documents a fixed routine for NIM: fetch a profile withdownload-to-cacheon a connected host, copy the cache, then start the container on the isolated host without anNGC_API_KEY(NVIDIA NIM air gap deployment). - Software.
pip downloadsaves packages to a directory that you can later pass topip install --find-links(pip documentation). Container images and OS packages need an internal mirror of their own. - Licences. Anything that phones home to validate a licence or needs an online account will stop working. Ask vendors about offline activation before you buy, not after.
Tools also need to be told not to reach out. With HF_HUB_OFFLINE=1, the Hugging Face library makes no HTTP calls and serves only cached files (Hugging Face docs). Ollama has a local-only mode via OLLAMA_NO_CLOUD=1 (Ollama FAQ).
Air-gapped vs on-premise vs offline
These terms describe increasing levels of isolation. An on-premise LLM runs in your own data centre but usually sits on the corporate network, often with internet access. Offline AI does not need a connection while running. Air-gapped rules out a connection physically and governs every transfer with a procedure.
So "air-gapped" describes how you operate a system, not a product feature. The same installation can be connected during testing and isolated in production.
What to watch out for
The transfer path is the weak spot. A USB drive that carries a model can carry malware just as easily. Typical safeguards are approved media, checksums for every file, a scanning station before import, and a log of who moved what and when.
There are trade-offs in daily use too. Agents cannot call web search, external APIs or hosted models, and security fixes arrive only as fast as your transfer process runs. Air-gapping is worth it when data must provably never leave the network, as in defence, critical infrastructure or regulated research. For many mid-sized companies, a local deployment behind a strict firewall, as outlined under sovereign AI, is enough. For running models without internet in general, see AI without internet.
Frequently asked questions
Is a system behind a strict firewall air-gapped?
No. A firewall filters a connection that exists. An air gap means there is no physical link at all, and every transfer is manual.
Can I use hosted models such as ChatGPT in an air-gapped setup?
No. Only models whose weights you can download and run entirely on your own hardware will work.
How often should you update an air-gapped AI system?
On a fixed schedule that the transfer process can handle, plus ad hoc for critical security fixes. The cadence is an operational decision, not a technical limit.
Sources
- NIST CSRC Glossary: air gap (from CNSSI 4009-2022) csrc.nist.gov
- Hugging Face Hub: Environment variables (HF_HUB_OFFLINE) huggingface.co
- Ollama FAQ docs.ollama.com
- NVIDIA NIM for VLMs: Air Gap Deployment docs.nvidia.com
- pip documentation: pip download pip.pypa.io