Air-gapped & on-prem AI

What "Air-Gapped AI" Actually Means (and When You Really Need It)

"Air-gapped" has become an AI marketing word. Most of the time it's used loosely, and the loose version is exactly what a security team can't accept. Here's what it really means.

If your organization handles classified, criminal-justice, export-controlled, or protected health data, someone has probably told you their AI is "on-prem" or "air-gapped." Before you believe it, it's worth being precise about what those words mean, because the difference between "runs on your servers" and "cannot reach the internet at all" is the difference between compliant and not.

A spectrum of isolation, not a single label

Deployment isolation isn't binary. It's a spectrum, and vendors often use the strongest-sounding word for a weaker reality:

  • Cloud (rented). The model runs on the vendor's infrastructure; your data travels to it. Most convenient, least isolated. Fine for public, low-sensitivity work.
  • Private cloud / VPC. The model runs in a cloud tenancy you control. Better, but it's still the internet, still a third party's hardware, still a network path off your premises.
  • On-premises. The model runs on servers you own, inside your building. Your data doesn't leave your infrastructure during use, but the machine may still have a network connection.
  • Air-gapped. The strongest form: the system runs on hardware with no network path in or out: no internet, sometimes no connection to any other network at all. Data physically cannot leave, because there's nowhere for it to go.

When a requirement says "air-gapped," it usually means the last one, literally. "On-prem but internet-connected" does not satisfy it.

Why owning the model is what makes real isolation possible

Here's the key point most people miss: you can only air-gap something you fully possess. You cannot air-gap a rented cloud model. Its whole design is to be reached over a network. Real isolation requires that the model's weights live on your hardware and that running them needs nothing external.

That's exactly what a private, owned, fine-tuned model gives you. The weights are static files. Once they're on the machine, generating answers requires no outside call: no license server, no telemetry, no "phone home." A model that can run with the network cable unplugged is one you can genuinely air-gap. A model that can't isn't a candidate, no matter what the brochure says.

The one honest catch: pre-staging

Air-gapped doesn't mean nothing ever crosses the boundary. It means every crossing is deliberate, controlled, and auditable. There's exactly one thing you have to bring in before you go dark:

  • The model and its software, once, up front. You download the base model and the necessary software on a connected machine, verify it, and carry it into the secure environment by whatever approved means your policy allows. After that, training and running the model need no network at all.

That's it. The client's data never crosses the boundary. The training runs inside. The finished model stays inside. The only ingress is a one-time, controlled transfer of the base ingredients, the kind of thing your security process is already built to handle.

Who actually needs this

Not everyone. Air-gapping adds operational overhead, and for public or low-sensitivity work it's overkill. But for a specific and important set of organizations, it's the requirement, not the upsell:

  • Defense and intelligence: classified networks, ITAR/EAR-controlled data.
  • Law enforcement and justice: CJIS-governed criminal-justice information.
  • Healthcare: PHI where policy forbids any external processing.
  • Critical infrastructure and finance: systems that must keep operating, and keep data in, regardless of the outside world.

For these buyers, "we'll host it securely in our cloud" isn't a smaller version of what they need. It's a different thing entirely.

The bottom line

"Air-gapped" is a precise technical claim, not a vibe: no network path in or out, data physically unable to leave. You can only make that claim honestly about a model you own and can run with nothing external, which is exactly why rented cloud AI can't offer it. If true isolation is your requirement, the question to ask any vendor is blunt: can this run with the network cable unplugged, and can you prove it? If the answer is anything but a clear yes, it isn't air-gapped.

Need AI that runs with the cable unplugged?

We build private models our clients own and run entirely on their own infrastructure, on-prem or fully air-gapped, with data never leaving their control. Start free with a readiness scorecard, or book a short call.