Key facts
| Availability | Enterprise and sovereign deployment option |
| Network | No inbound or outbound internet required |
| API surface | OpenAI-compatible gateway runs inside the isolated network |
| Hardware | Customer-supplied servers and GPUs |
| Models | Open-weight models served locally; hosted frontier models unavailable |
| Updates | Controlled offline transfer process |
| Observability | Local logs and metrics; no external telemetry |
| Product status | Live via enterprise agreement |
TL;DR
- Air-gapped is a real deployment option, not a marketing label.
- No internet path in or out — updates arrive through controlled transfer.
- The API stays OpenAI-compatible, so application code still works.
- You provide GPUs and operations; only open-weight models are available.
- Plan capacity, model updates and monitoring before you cut over.
How it works, step by step
- Confirm the isolation requirement: fully air-gapped or restricted-egress is often enough.
- Size GPU capacity for your peak concurrency using your target open-weight models.
- Define the media-transfer process for updates, weights and patched images.
- Deploy the OpenAI-compatible gateway and model server inside the isolated network.
- Validate the API with your existing clients, tests and evals offline.
- Document operations: monitoring, incident response, capacity headroom and refresh cadence.
Try it yourself
Open the sovereign AI readiness score →
What air-gapped means in practice
An air-gapped deployment removes the network path entirely: inference, the API gateway, logging and storage all run inside an isolated environment. Nothing calls home, which means no hosted model routing, no external telemetry and no live licence checks. Everything the platform needs — model weights, container images, patches — arrives through a controlled transfer process your security team approves. The trade-off is operational: you own capacity planning, upgrades and monitoring, and you carry the full cost of the GPUs. Capacity headroom matters more here than in cloud, because there is no burst capacity to rent when demand spikes.
What works and what changes
The interface stays familiar:
- Unchanged: OpenAI-compatible chat completions, streaming, JSON mode, function calling and embeddings, called from your existing SDKs.
- Changed: only locally served open-weight models are available; hosted frontier models are out of scope because they require an external path.
- Changed: updates are periodic and deliberate rather than continuous.
- Changed: you run observability locally, with no provider-side dashboard.
For many regulated teams that trade-off is the point: they accept a narrower catalogue and slower refresh in exchange for physical control.
What we do and what we do not do
What we do: support air-gapped deployments under an enterprise agreement, ship the OpenAI-compatible gateway for isolated networks, and define a controlled update process with you. What we do not do: provide GPUs or data-centre capacity, operate the environment remotely, or include hosted frontier models — those need connectivity by definition. If your requirement is strict isolation, plan the transfer workflow and support model up front; the docs and your account team cover the deployment specifics.
Honest comparison
| Capability | Plugsky air-gapped | Plugsky shared cloud | Fully hosted provider |
|---|---|---|---|
| Internet dependency | None | Required | Required |
| Model catalogue | Open-weight models you host | 30+ models | Provider catalogue |
| Frontier hosted models | Not available | Available | Available |
| Update cadence | Controlled transfers | Continuous | Continuous |
| Operations ownership | You, with vendor support | Plugsky | Provider |
| Best fit | Sovereign, defence, critical infrastructure | Most teams | Non-regulated teams |
Frequently asked questions
Does air-gapped mean no internet at all?
Yes. In a fully air-gapped deployment there is no inbound or outbound network path; updates and model weights arrive through a controlled transfer process.
Which models can run air-gapped?
Locally served open-weight models. Hosted frontier models are unavailable because they require an external connection.
Is the API still OpenAI-compatible?
Yes. The gateway runs inside the isolated network and keeps the same API surface, so existing SDK code works unchanged.
Who provides the hardware?
You do. Air-gapped deployments run on your servers and GPUs, with Plugsky providing the software stack and support under an agreement.
How do updates work without a network?
Through a documented offline transfer process: reviewed media carrying container images, patches and model weights, applied on your schedule.
How do I monitor it?
With local logs and metrics. There is no external telemetry, so build dashboards inside the isolated network or on a management plane you control.
Is there a middle option?
Yes. Restricted-egress deployments keep most of the isolation while allowing controlled outbound access for updates and model routing, which suits teams that do not need physical isolation.