For most enterprise AI deployments, the question of where the model runs is a footnote. The application calls a cloud API, the API returns a response, the application moves on. The customer’s data may be encrypted in transit. The provider may sign a BAA or DPA. The arrangement is usually adequate for the use case.

For a specific category of enterprise workloads, this arrangement is not adequate at any price. The data cannot leave the network. Not “should not.” Cannot. Defense contractors, classified research environments, certain financial trading operations, foreign intelligence customers, and the highest-sensitivity healthcare and government workloads operate under data-handling constraints where any external API call is non-starter regardless of the encryption story or the contractual language around it.

For those workloads, the deployment architecture is air-gapped AI. This article covers what that means, when it is required, the operational trade-offs, and the architectural decisions that make air-gapped AI viable in 2026 in ways it was not five years ago.

What Air-Gapped Actually Means

“Air-gapped” is more than “on-premise.” On-premise means the workload runs in your data center. Air-gapped means the environment running the workload has no network connectivity to external systems. Updates arrive via physical media or controlled gateways. Outbound traffic is blocked at the infrastructure level, not just the firewall level.

In practice, air-gapped environments fall on a spectrum:

• Fully air-gapped. No network connectivity in or out. Updates via physical media. Used for classified workloads, certain defense applications, the highest-sensitivity research.
• Air-gapped with controlled gateway. Limited connectivity through a tightly controlled diode or one-way gateway, typically for receiving updates or sending alerts but not for arbitrary data exchange. Used for many regulated commercial workloads.
• Network-segmented. Sometimes loosely called “air-gapped” but technically a strict network isolation rather than true air-gap. The environment runs on a separate VLAN, has no direct internet routing, and communicates only with specifically allowlisted systems. Used for many enterprise sensitive workloads.

The architectural and operational considerations differ across these. The discussion below applies primarily to the fully air-gapped and air-gapped-with-controlled-gateway cases. Network-segmented deployments have more flexibility.

When Air-Gapped Is Required

Five categories of enterprise workload where air-gapped AI is either required or strongly preferred:

1. Classified workloads. Defense Department, intelligence community, and classified research customers operate under controls that explicitly require air-gap for systems handling classified data. AI processing of classified data is no exception.
2. FedRAMP High and IL5/IL6 workloads. Federal workloads at the higher impact levels often require dedicated infrastructure with strict isolation. AI workloads inherit these requirements.
3. Healthcare research with sensitive populations. Research involving genetic data, behavioral health, substance abuse treatment records, or HIV/AIDS status may be subject to stricter handling than baseline HIPAA. Some institutions interpret this as requiring air-gapped processing.
4. Financial trading operations. Algorithmic trading systems, particularly proprietary strategy development, often operate in network-isolated environments to prevent both data exfiltration and trading-signal leakage.
5. Sovereign-data workloads. Enterprises operating in countries with data residency laws (or under government contracts in those countries) may require that AI processing remain entirely within national borders, sometimes requiring air-gap from broader corporate networks.

If your workload is in one of these categories, you already know it. The conversation is not whether to air-gap. It is how to deliver AI capability within air-gap constraints.

What Air-Gapped AI Actually Looks Like

An air-gapped AI deployment has the same conceptual components as a cloud AI deployment, with significant operational differences.

Models

The AI models run on customer-owned hardware inside the air-gapped environment. Model files are deployed via physical media (typically encrypted drives) and loaded onto the inference infrastructure. New model versions follow the same deployment path on scheduled update cycles.

Model selection is constrained to what can be deployed locally. Cutting-edge proprietary models (the latest from Anthropic, OpenAI, or Google in their hosted-only configurations) are not available. Open-weights models from organizations that publish their weights, plus open-source models, plus models specifically licensed for on-premise deployment, are the available set.

Inference infrastructure

GPUs or specialized AI accelerators inside the air-gapped environment. Capacity planning is critical because there is no elastic scale. The peak load the environment can handle is the capacity provisioned. Workload prioritization within fixed capacity becomes an operational discipline.

Model updates

The cadence of model improvements in the public ecosystem is rapid. The cadence in an air-gapped environment is whatever the controlled-update schedule allows. Quarterly or semi-annual model updates are typical. The implication: the air-gapped environment is consistently running models that are not the latest available externally.

Orchestration and tooling

Everything that wraps the model (prompt management, retrieval-augmented generation, agent frameworks, audit logging) must also run in the air-gapped environment. Cloud-hosted orchestration platforms are not available. The deployment stack is end-to-end on-premise.

The Model-Agnostic Argument for Air-Gapped Deployments

Air-gapped deployments specifically benefit from model-agnostic platform architecture. Three reasons.

First, available models change. An open-weights model that was best-in-class in 2025 may be surpassed by a different open-weights model in 2026. Air-gapped customers cannot use the new model if their platform is locked to a specific backend. Model-agnostic architecture preserves the ability to swap backends as the air-gapped-deployable ecosystem evolves.

Second, licensing terms change. Open-weights models that are freely usable today may shift to restrictive commercial licensing for enterprise use. The platform that allows alternate backends preserves continuity.

Third, different workloads need different models. Within a single air-gapped environment, code generation workloads may want one model, document summarization may want another, structured extraction may want a third. Multi-backend support inside one platform is operationally cleaner than separate stacks.

What Most Vendors Get Wrong

The common failure mode in air-gapped AI vendor evaluations: the vendor’s pitch covers their cloud product extensively, and air-gap is treated as a footnote or “yes, that’s available.” When the customer engages, the air-gapped version is meaningfully different, less capable, and requires substantial professional services to deploy.

Questions worth asking any vendor whose product you intend to deploy air-gapped:

• Is air-gapped deployment a primary product configuration or a special engagement?
• What is the actual feature parity between the cloud and air-gapped versions of the product?
• What model backends are supported for air-gapped deployment?
• What is the model update process?
• What is the operational support model for air-gapped customers? (Hint: it cannot be “remote diagnostic access” because that defeats the air-gap.)
• Can the vendor demonstrate an existing air-gapped customer deployment?
• What is the deployment time from contract to production?
• What hardware requirements does the air-gapped version have?

A vendor whose answers to these questions are vague or require multiple follow-ups is a vendor whose product was not built for air-gapped deployment.

The Hybrid Reality

Most enterprises running air-gapped AI do not run all AI in the air-gapped environment. The hybrid pattern:

• Air-gapped AI for the workloads that require it (sensitive data, regulated processing)
• Cloud AI with BAA or DPA for workloads where the constraints are looser
• Internal policy and governance that determines which workloads route to which environment

The architectural challenge is making the developer experience consistent across both. Developers should not need to learn two different platforms depending on which workload they are building. A unified platform that supports multiple deployment targets behind the same interface is what makes hybrid practical.

Common Mistakes

1. Treating air-gap as a deployment toggle on a cloud product. Most cloud-first AI products do not deploy cleanly air-gapped. The architecture is fundamentally different.
2. Underprovisioning the air-gapped infrastructure. Without elastic scale, capacity provisioning has to anticipate peak load. Underprovisioning produces queueing under load.
3. Skipping the orchestration layer. Models alone are not a deployment. Prompt management, RAG, audit logging, and other tooling have to work air-gapped too.
4. Ignoring the model update cadence. Air-gapped customers will run older models than their cloud peers. Plan for this. Build in evaluation cycles that test new available models before deploying.
5. Forgetting about logging and audit. Air-gapped environments still have audit requirements. Logging needs to be locally stored, retained per policy, and accessible for review without breaking the air-gap.

Golden Path Digital and Air-Gapped Deployment

Golden Path Digital builds AI tooling with air-gapped deployment as a first-class configuration. AS/Forward, the IBM i modernization platform, is designed to run inside customer infrastructure with no external dependencies during operation. Model backends are configurable: hosted Claude, OpenAI, or Gemini for customers who can use them; locally-deployable open-weights models for customers who cannot. QuantaPath AI follows the same pattern for HIPAA and broader regulated workloads. 1 US patent pending on the core ingestion architecture.

The deployment model recognizes that for the customers who need air-gap, the choice is not between air-gapped and cloud. The choice is between a vendor whose product was built for air-gap from the start and one where air-gap is a retrofit.

What to Do Next

If your organization needs to deploy AI inside an environment where the data cannot leave the network, the next step is an architecture conversation. Visit goldenpathdigital.com for product details, or reach out via the contact form to schedule a discovery call.

Golden Path Digital is an enterprise software and AI modernization company headquartered in Hot Springs Village, Arkansas. AS/Forward modernizes IBM i and RPG codebases. Laravel Ascend automates Laravel application upgrades. QuantaPath AI delivers HIPAA-compliant CRM and workflow automation. Serving enterprises nationwide with US-based delivery and air-gapped deployment options. 1 US patent pending.