Air-gapped AI: when a regulated buyer actually needs it

Air-gapped AI is artificial intelligence that runs on hardware with no connection to the internet or any outside network, so no prompt, document, or model weight can ever leave the building. A regulated buyer genuinely needs it when the data is classified, legally confined to a jurisdiction, or would cause serious harm if it escaped the perimeter, and when a contract or regulator demands provable isolation rather than a supplier's promise. For most organisations that is the strict end of the spectrum, not the default posture. The honest test is whether you could accept any third party ever seeing the input, and whether you must be able to prove that they cannot.

• Air-gapped means owned, offline, and provable: the AI, the data, and the audit trail never leave your controlled environment.
• It is the strict end of a spectrum that runs from public cloud AI, through private cloud, to on-premises, to fully air-gapped.
• Most regulated work is well served by on-premises or private deployment; true air-gapping is for the highest-sensitivity cases.
• The real test is legal and contractual: can any outside party ever see the input, and must you prove that they cannot?
• Mickai runs a capable model entirely on hardware you own, so sovereignty becomes a deployment choice rather than a leap of faith.

What does air-gapped actually mean?

An air gap is a physical and logical separation between a system and any untrusted network, most importantly the public internet. In an air-gapped deployment the model weights, the inference engine, and every prompt and response sit inside a boundary that data cannot cross. There is no telemetry, no remote logging, and no silent update reaching back to a vendor. Isolation becomes a property you can inspect, not a policy you are asked to trust.

When does a regulated buyer actually need it?

You need air-gapping when the cost of a single leak is catastrophic and when isolation must be demonstrable to an auditor, a regulator, or a counterparty. That describes classified defence material, certain patient records, live investigations, safety-critical control systems, and commercial data whose exposure would end a contract or a company. If a credible adversary or a trusted insider could exfiltrate the input, and the law or the contract says that must be impossible, the perimeter has to be closed.

The practical test is three questions. Is the data legally or contractually forbidden from leaving a defined boundary? Would exposure cause irreversible harm? And must you prove isolation rather than simply assert it? Three yes answers point clearly to an air gap.

When is air-gapping overkill?

Air-gapping is overkill when the underlying requirement is control and residency rather than absolute isolation. A firm that must keep data in the United Kingdom, avoid any training on its content, and retain a full audit trail can often meet every obligation with an on-premises or private deployment that never touches a shared cloud. Closing the gap entirely adds real operational cost, because patching, model updates, and threat intelligence all have to be carried in by hand.

The mistake is treating the air gap as a marketing badge. The right question is not whether isolation sounds reassuring, but whether a regulator or contract genuinely requires it. Buying more isolation than the risk warrants slows delivery without reducing real exposure.

Does the EU AI Act or UK regulation force air-gapping?

No regulation names an air gap as such. The EU AI Act sets obligations by risk category, and its high-risk obligations are deferred to 2 December 2027, so 2026 is a preparation window rather than a live high-risk deadline. UK data protection law and sector rules focus on lawful basis, data residency, and demonstrable control. Air-gapping is one strong way to satisfy those duties for the most sensitive workloads, not a rule in its own right.

How does Mickai deliver air-gapped AI without losing capability?

Mickai is a Sovereign Intelligence Operating System, a SIOS that runs a capable model directly on hardware you own, with its assistant, studios, and audit subsystems all inside your perimeter. Because the whole environment is designed to run offline, an air-gapped installation keeps the same interface and the same studios as a connected one, with no feature stripped out to make isolation possible. Micky Irons, founder of Mickai, built the platform so that sovereignty is the starting point rather than an afterthought.

That posture is backed by an intellectual property estate of 104 filed UK patent applications, approximately 2,340 claims, owned by Mickai LTD, covering the isolation, entitlement, and audit techniques that make provable offline operation practical. Updates arrive as signed packages a buyer chooses to admit, so the perimeter stays closed and the audit trail stays intact.

Frequently asked questions

Is air-gapped AI the same as on-premises AI?

No. On-premises means the system runs on your own hardware, but it may still reach the internet for updates, licensing, or telemetry. Air-gapped goes further, because there is no network path to the outside world at all. Every air-gapped deployment is on-premises, but not every on-premises deployment is air-gapped.

Can an air-gapped model still be updated?

Yes, but updates are carried across the boundary deliberately rather than downloaded automatically. New model versions and security patches arrive as signed packages that an administrator reviews and admits. This keeps the buyer in control of exactly what enters the environment and preserves the audit trail.

Do I lose accuracy by running AI offline?

Not meaningfully. A capable model that runs locally can match the quality most regulated tasks require, from drafting to analysis to document review. The trade is operational rather than intellectual, because you take on patching and updates in exchange for isolation, not weaker answers.

How do I prove to an auditor that the system is really isolated?

Provable isolation comes from architecture and evidence, not assurances. That means a documented network boundary, a tamper-evident audit trail of every action, and signed records showing that no data left the perimeter. Mickai is built to produce this evidence by default, so isolation can be demonstrated rather than merely promised.

Is air-gapping only relevant to defence and government?

No. Defence and national security are the clearest cases, but the same test applies to healthcare records, financial and legal privilege, critical infrastructure, and any commercial data whose leak would be irreversible. The question is the sensitivity of the data and the strength of the obligation, not the sector on the letterhead.