Key facts
| Definition | A model deployment in a network with no external connectivity |
| Why it matters | Satisfies the strictest isolation and sovereignty requirements |
| What is included | Weights, serving stack, API endpoint, authentication and logging |
| Update path | Signed or scanned artifacts transferred through controlled media |
| API compatibility | OpenAI-compatible endpoint, so applications keep working |
| Operational needs | Hardware sizing, model lifecycle, monitoring and incident response on site |
| Plugsky offer | Air-gapped deployment is available as an enterprise option; confirm scope and terms with the team |
| Common trap | Underestimating update and observability work when nothing can phone home |
TL;DR
- Air-gapped means no network path out — not just a private subnet.
- Weights, serving and logs all live inside the isolated network.
- Updates require a deliberate, reviewed transfer process.
- Applications keep an OpenAI-compatible endpoint, so code changes are minimal.
- Plan hardware, monitoring and support before deployment day.
How it works, step by step
- Confirm that policy, not preference, requires full disconnection.
- Size hardware for peak concurrency plus headroom for model updates.
- Deploy the serving stack and OpenAI-compatible endpoint inside the isolated network.
- Provision credentials, RBAC and local audit logging for all clients.
- Define the update process: artifact review, transfer, validation and rollback.
- Rehearse failover and incident response without vendor remote access.
Try it yourself
Open the private LLM cost estimator →
What air-gapped actually means
An air gap is a physical or logical absence of any network path to the outside world. A private VLAN with a NAT gateway is not air-gapped. In a true deployment, model weights arrive through controlled media, the serving stack never resolves external hostnames, telemetry is stored locally, and support is delivered through documented, reviewed procedures rather than remote sessions.
Architecture and operations
- Serving: inference runs on hardware inside the network, sized for peak load.
- API: an internal OpenAI-compatible endpoint serves applications and tools.
- Storage: models, prompts and logs reside on internal systems under local retention rules.
- Updates: reviewed bundles move in, are validated and can be rolled back.
- Observability: metrics and audit logs collected internally, exported only by policy.
The trade-off versus a private VPC is stark: maximum isolation for maximum operational effort.
Common mistakes
- Assuming network segregation alone satisfies the requirement without reviewing the full data path.
- Forgetting model lifecycle — base models and serving stacks still need updates.
- Underprovisioning hardware, then discovering the workload cannot scale when it matters.
- No local observability, so failures surface only from user reports.
- Ignoring the human process: who approves an update bundle, and how rollback is executed.
Air-gapped deployments with Plugsky
Plugsky offers air-gapped deployment as an enterprise option alongside VPC and on-prem. Because the API remains OpenAI-compatible, existing applications and tooling reconnect to the internal endpoint with a base URL change, and platform concepts such as scoped keys, RBAC and audit logging still apply. Capacity is dedicated, so sizing, model selection and update cadence are agreed up front. Engage the team early so architecture, legal terms and support procedures are settled before the isolated environment is built.
Honest comparison
| Level | Connectivity | Update path | Operational burden |
|---|---|---|---|
| Public API | Internet | Continuous | Lowest |
| Dedicated VPC | Private network path | Automated | Moderate |
| On-prem | Corporate network | Controlled | High |
| Air-gapped | None | Physical media | Highest |
Frequently asked questions
What is an air-gapped LLM?
A language model deployed in a network with no external connectivity, where weights, serving infrastructure, data and logs all stay isolated from the internet.
How is air-gapped different from on-prem?
On-prem typically retains corporate network connectivity. Air-gapped means no external network path at all, so updates and observability must work through controlled local processes.
How do models get updated in an air-gapped environment?
Through a formal process: a signed or scanned artifact bundle is reviewed, transferred through controlled media, validated in a staging area and applied with a rollback plan.
Does an air-gapped deployment change my application code?
Not materially. The endpoint stays OpenAI-compatible, so applications change the base URL and credentials rather than the integration.
What are the main costs?
Dedicated hardware sized for peak load, staff time for updates and monitoring, and slower access to new models — balanced against the strongest possible isolation.
Does Plugsky offer air-gapped deployments?
Yes, as an enterprise option. Confirm scope, hardware sizing, update process and contractual terms with the team before planning the build.