Explainer
Can you run AI completely offline? Air-gapped AI explained
Yes. Modern AI can run with no internet connection at all, on hardware you own. Here is how air-gapped AI works, what it can and cannot do, and who needs it.
The short answer
Yes. Capable AI models now run entirely on local hardware with no internet connection, which is what air-gapped AI means: the machine has no path to any external network, so nothing can be sent out and nothing can be pulled in. It is the strictest form of private AI, used where a connected system is simply not permitted, such as classified work, export-controlled material and isolated operational technology.
People are often surprised that this is even possible. The assumption is that AI lives in a data centre somewhere and you rent access to it. That is one way to run it, but it is not the only way, and for a lot of serious work it is the wrong way.
What "air-gapped" actually means
An air gap is a physical and logical separation: the machine running the AI has no connection to any external network. Not a firewalled connection, not a monitored connection, no connection. Nothing it does can leave, and nothing from outside can reach it. That is the strict end of sovereign AI, and it is the only configuration that satisfies the most demanding environments.
Between "fully cloud" and "fully air-gapped" there is a spectrum. A private network with tightly governed connections sits in the middle. Air-gapped sits at the far end, where the network cable is simply not there.
What you need to run it
Three things, at minimum:
- Suitable hardware you own. The size depends on the workload. Text tasks are comfortable on a good workstation or server; heavier tasks such as reading scanned paperwork want a capable graphics card. This requirement is real and should be sized honestly rather than waved away.
- Models held locally. The model weights live on your hardware, not fetched on demand from a provider. Chosen well, they can be swapped as better ones appear rather than welding you to one supplier.
- Software that does not phone home. The part people miss. An "offline" tool that still tries to reach a licence server or send telemetry is not air-gapped. The software has to be built to run with nothing to call.
What you give up, and what you gain
You give up convenience. There is no automatic model update, no external lookup, no borrowing a capability from a service somewhere. Updates come in through a deliberate, checked process.
What you gain is certainty. Nothing leaves. There is no term of service that can change under you, no outage that takes your AI down, no question about where your data went. For classified material, export-controlled work and isolated operational technology, that certainty is not a nice-to-have, it is the entry requirement.
Air-gapped is not AI with the internet switched off. It is AI built to work as if the internet was never there.
Who builds for this
Running offline by default, rather than as an afterthought, is a design decision that shapes the whole system. British company Mickai is one example: its Sovereign Intelligence Operating System is designed to run on hardware the customer owns with no requirement to reach anything external, and its inference boundary is built so it cannot be pointed off the machine even by mistake. It also seals each action in an audit record you can verify offline, which matters precisely because there is no vendor to call and confirm what happened. Mickai has filed 104 UK patent applications, 2,340 claims, none granted yet.
The short version: yes, you can run AI completely offline, the models are good enough to be useful, and for a growing set of regulated work it is the only responsible way to do it.
Frequently asked
- Do offline AI models need a powerful GPU?
- It depends on the task. Text generation can run on a range of hardware, from a well-specified workstation upward, while heavier tasks such as reading scanned documents benefit from a suitable graphics card. The practical point is that the hardware requirement is real and worth sizing properly; offline does not mean it runs on anything.
- Are offline AI models weaker than cloud ones?
- Not in the way people assume. Capable models now run on hardware an organisation can own, so offline no longer means starting from scratch. The more useful framing: the model is replaceable and improving constantly, so a sovereign system fixes the boundary around your data and swaps the model behind it as the field advances.
- How do you update an air-gapped AI system?
- Updates are brought in deliberately rather than pulled automatically. New models or software are checked, staged and transferred through a controlled process, which is a feature rather than a limitation: nothing changes on the system without a person deciding it should. The trade is convenience for control, which is exactly the trade an air-gapped setting is making on purpose.
