Air-Gapped AI Deployment

Air-gapped AI deployment means one thing: the system has no external network interface. No outbound calls during inference. No telemetry. No license validation against a vendor server. No model update pull from an external registry. The compute runs inside your environment with no dependency on anything outside it.

This is the baseline requirement for classified government networks. It's the only architecture that satisfies certain defense, healthcare, and financial environments where any external data path is categorically prohibited.

Most vendors who claim air-gapped capability do not deliver it. "Air-gapped" in their documentation means "private deployment" — which means their infrastructure, or a managed cloud in your region, or an on-prem option that still phones home for license validation or model telemetry. These are on-premises configurations, not air-gapped ones. The difference matters.

4MINDS runs fully air-gapped after initial deployment. Inference runs on your compute with no external calls required. There is no management plane in our infrastructure that touches your deployment. No sub-processors receive your data during inference. The model weights are transferred to your infrastructure at setup and remain there — fully under your control, versioned by your team.

Who requires air-gapped AI

Defense and government. Classified networks and systems handling Controlled Unclassified Information, SCI, or SAP data cannot have external network paths, full stop. "Sovereign cloud" and "European-hosted" are not air-gapped — they're cloud deployments with a different jurisdiction flag. Air-gapped means the compute is physically inside your controlled environment with no external network interface.

Healthcare — high-security environments. Clinical AI systems handling PHI in environments with strict network segmentation requirements — operating room systems, imaging networks, research environments with IRB-mandated data isolation — require inference that stays inside the perimeter by architecture, not by vendor promise.

Financial services — trading and core banking. High-frequency trading infrastructure and core banking systems run on networks with no external latency and strict controls on data egress. AI inference on those networks needs to run natively, not via API.

Critical infrastructure — energy, utilities, manufacturing. OT/IT convergence projects deploying AI on plant-floor networks or operational technology environments where external connectivity represents an attack surface.

What "air-gapped" actually requires from your vendor

An air-gapped deployment passes a simple test: disconnect the server from all external networks after initial setup. Does everything still work? For 4MINDS: yes.

The checklist for evaluating any air-gapped AI claim:

• Does inference require any outbound network call? (License validation, model telemetry, routing through a proxy?)
• Does the model update process require external connectivity? (Pulling updates from a vendor registry?)
• Does the orchestration layer require external connectivity? (Agent framework calling a cloud API for coordination?)
• Does the monitoring and logging system require external connectivity? (Telemetry shipped to a vendor dashboard?)

Each of these is a hidden external dependency. Any one of them disqualifies the deployment as genuinely air-gapped.

The 4MINDS air-gapped architecture

4MINDS deploys on your Kubernetes cluster — on-prem or in a private environment with no external network access required. The inference stack, the Ghost Weights training pipeline, the Graph RAG retrieval layer, and the eval gates all run inside your perimeter.

Ghost Weights updates the model inside your infrastructure. The shadow training pipeline runs on your data, on your compute, and produces new model versions that are evaluated and promoted entirely within your environment. The model gets smarter without any external call.

No 4MINDS infrastructure sits between your query and your model output. No sub-processors. No cloud provider in the critical path. The deployment is genuinely yours.

Zero-trust posture and audit trail

Air-gapped deployment is the perimeter control. Inside the perimeter, 4MINDS provides the governance layer:

• Every inference is logged with model version, input context, and output
• Every model update is version-controlled and retainable — rollback to any prior version on demand
• Eval gates validate model behavior against defined criteria before any version reaches production
• All agent actions are logged with a complete audit trail inside your infrastructure

The audit trail stays inside your environment. It is never shipped to a vendor dashboard or external logging service.