What “air-gapped AI” actually means for a firm

“Air-gapped” is used loosely in AI marketing. Precisely, an air-gapped system has no network path at all to any external network — not firewalled, not VPN-restricted, physically or logically severed. Most deployments described as air-gapped are actually something weaker, and the distinction matters for what you can promise a client.

A spectrum, not a binary

• True air gap. The inference host has no route to the internet. Model weights and updates arrive by controlled physical media. Rare, operationally heavy, used in classified or critical-infrastructure settings.
• Private / on-prem deployment. The model runs on infrastructure the customer controls, typically on a private network with tightly restricted egress. Client data does not go to a third-party model API, but the host may still reach the internet for OS updates, telemetry or licensing.
• Hosted API with a DPA. Data goes to a third-party model provider, governed by a data-processing agreement and retention commitments. This is not sovereign and not air-gapped, however strong the contract is.

Questions that separate them

Can the inference host make outbound connections? Who operates it? Where do model weights come from and how are they updated? Is any prompt content sent anywhere for logging, moderation or analytics? The honest answer to these places a system on the spectrum.

How FinRace AI describes itself

FinRace AI runs inference on self-hosted models — client data is not routed to a third-party LLM API. Whether a given deployment is fully air-gapped depends on how it is provisioned; the security page states what the current deployment enforces rather than relying on the word “air-gapped” as a shorthand.