Air-gapped AI, explained.

A plain-English guide for confidential and regulated work.

An air-gapped AI is an AI system that runs on a machine with no connection to any outside network. Your prompts and documents are read and answered right there, on that system, and there is no path for any of it to travel out. The model runs locally; there is no cloud call to a vendor, no telemetry, and nothing to intercept, because nothing leaves.

Air-gapped vs network-isolated

The two terms are close, and it is worth being precise. A network-isolated system has no standing outbound path to the internet: it will not phone home, stream data out, or reach a vendor server. It may still sit on a controlled internal network so the people who use it can reach it. An air-gapped system is stricter still: it has no network connection at all, internal or external. Both approaches keep confidential work from leaving your control. Air-gapping simply removes even the internal network as a surface, which is why the most sensitive settings ask for it.

Why the distinction matters

For most confidential work, network isolation is the property that carries the promise: if there is no outbound path, your material cannot be logged by a third party, retained, sub-processed, or swept into a training set. For privileged legal matters, patient data, unpublished research, defense and controlled-information settings, and regulated finance, teams often want the stricter guarantee an air gap provides: not "it will not send data out," but "it cannot, because there is no network to send it over."

How to prove it

The elegant thing about an air gap is that you do not have to trust a policy to believe it. Disconnect the machine from every network and ask it something. If it keeps answering with nothing connected, then nothing you type was ever going anywhere. It is a test anyone in the room can run, and it either passes or it does not.

The honest tradeoffs

An air gap is not free. Updates and new models arrive by hand, on physical media, installed on-site and applied deliberately rather than streamed in automatically. That is a feature, not a bug: the outbound path the whole promise denies is never reopened just to keep the system current. A frontier cloud model can still be sharper on the very hardest reasoning, and anything that genuinely needs live web data belongs in the cloud. For confidential work, though, an air-gapped or network-isolated system answers immediately and only to you.

How a commissioned build gets there

Standing up an air-gapped or network-isolated AI you would stake confidential work on is a real engineering project: sizing and sourcing the system, tuning a capable local model to your work, hardening the boundary so nothing phones home, and building something that keeps working after the person who set it up moves on. A Garnet build is a private AI installed on a dedicated system inside your building. It is network-isolated by default and can be deployed fully air-gapped where the setting calls for it, with no standing access for anyone, including us, after handover. See what we can and can't see, why an on-premise AI is right for confidential documents, and an honest look at commissioning a build versus rolling your own. For teams and institutions, see on-premise AI for law firms and in-house teams.