Air-gapped AI: run large language models with no network at all
RAPTOR puts the models, the interface and the data on one sealed unit. It behaves the same on a desk, in a vehicle, and in a facility that has never been connected to anything.
- Air-gapped AI
- Air-gapped AI is artificial intelligence that runs on hardware with no connection to the internet or to any untrusted network. The model weights are stored locally, inference happens locally, and prompts, documents, imagery and results never cross the boundary. Nothing is sent to a vendor for processing, and nothing has to be fetched to keep the system working.
Why teams air-gap their AI
Most AI products assume a link to someone else’s data centre. That assumption fails the moment connectivity is denied, degraded, forbidden by policy, or simply absent — which describes a forward operating base, a disaster zone, a ship at sea, a classified facility and a factory floor equally well.
An air gap is the strongest available control for that problem, because it is physical. There is no route for data to leave and no route for an attacker to arrive over the wire. The hard part was never the concept; it was building a genuinely useful AI stack that still works once the network is removed.
Classified, regulated and export-controlled material often cannot be processed by a third-party service at any price. An air gap turns that from a contractual argument into a physical fact.
Expeditionary, maritime, subterranean and disaster-response work routinely happens with no usable bandwidth. An offline platform is the only one whose performance does not degrade with the link.
Analysis that feeds a decision in seconds cannot wait on a satellite round trip. Inference next to the sensor is faster and far more predictable than inference a continent away.
No outbound calls means no telemetry, no crash reports carrying context, no training opt-out to police, and no change of terms that quietly affects your data.
What has to be true for AI to work offline
Each one is a place where an otherwise capable platform quietly stops being air-gapped.
- 01
Every model is already on the device
Local large language models, vision-language models, speech transcription, document OCR and image generation are installed as templates on the unit. There is no first-run download and no account to create.
- 02
The interface is offline too
The RAPTOR App Store — fonts, icons, scripts, documentation — is served from the node itself. A management console that needs a CDN is not air-gapped; it is merely unplugged and broken.
- 03
Nothing phones home
No analytics, no licence check, no update ping. The behaviour you observe in acceptance testing is the behaviour you get in the field, because there is no second code path that depends on connectivity.
- 04
Untrusted media is scanned before ingest
An air gap protects the network, not the file somebody carried across it. A pre-ingest scanning pipeline sits in front of the analysis stack, so media from outside the boundary is checked before it reaches a model.
- 05
Updates move deliberately
New models and templates arrive as verified, versioned bundles, applied on your schedule through your own transfer process — never as an automatic push you learn about afterwards.
- 06
Capacity is measured, not estimated
Throughput, latency and concurrent-user figures for each capability are benchmarked on the same hardware you deploy, so planning starts from real numbers instead of a vendor’s best case.
Cloud AI service versus air-gapped RAPTOR
| Cloud AI service | RAPTOR | |
|---|---|---|
| Connectivity | Required for every request | Never required |
| Where data is processed | Vendor infrastructure | The unit in front of you |
| Data leaving your control | Yes, by design | No route out |
| Behaviour when the link drops | Unavailable | Unchanged |
| Cost model | Per token, indefinitely | One-off hardware purchase |
| Performance predictability | Shared and variable | Dedicated hardware, benchmarked |
| Who decides when the model changes | The vendor | You |
Air-gapped AI: common questions
Keep reading
The other half of the question: who owns the weights, the hardware and the decisions, and why jurisdiction is not the same as hosting region.
How a capability gets from the catalogue to a running endpoint on the node, in one click and with no connectivity.
Compute, power, endurance and environmental ratings for the unit that runs it all, plus the labelled schematic.
Ready to deploy?
Talk to the team about configurations, deployment timelines and getting a unit in front of your operators.