Key facts
| Definition | No external network dependency; models and runtime operate fully inside your perimeter |
| Deployment models | On-prem and air-gapped tiers alongside cloud and VPC options |
| Models | Open-weight models suitable for offline operation (Nemotron, Llama, Qwen, Mistral and more) |
| API contract | OpenAI-compatible API in all tiers; application code does not change |
| Updates | Delivered through your controlled process; plan cadence and verification |
| Identity and audit | SSO/SCIM and SIEM export operate inside your environment |
| Compliance posture | SOC 2 Type II and ISO 27001 readiness in progress (not yet certified) |
| Status | Live platform; enterprise deployment terms per engagement |
TL;DR
- An air gap removes network dependency; it does not remove governance or operational work.
- Use a VPC or on-prem with controlled egress unless policy truly forbids any external connectivity.
- Air-gapped models must be open-weight, versioned and vetted before they cross the boundary.
- Plan update, support and observability processes that work without the internet.
- Budget for hardware, spares and skills — the gap changes operations, not just architecture.
How it works, step by step
- Confirm the requirement: does policy forbid all external connectivity, or only data egress?
- If only egress is restricted, evaluate a VPC or on-prem deployment with controlled updates first.
- Select open-weight models that meet quality targets and licence requirements for offline use.
- Design the transfer process for weights, runtime updates and logs, including verification steps.
- Size hardware for your concurrency and context requirements, with spares.
- Stand up offline evaluation, monitoring and incident processes.
- Document the architecture for assessors, including boundary controls and change history.
Try it yourself
Open the private LLM cost estimator →
What air-gapped actually requires
True air gap means no network path in or out. That has consequences: model weights arrive on controlled media, updates follow a documented intake process with checksums and scanning, and telemetry cannot be streamed to a vendor. Licensing must work offline, and support is delivered through your own staff or on-site visits. Plugsky's on-prem and air-gapped tiers run open-weight models behind your firewall with the same OpenAI-compatible API used in the cloud, so applications built against the API remain portable.
When a VPC or on-prem is enough
Many organisations that ask for air-gapped AI actually need data residency plus isolation. A private endpoint in your VPC keeps traffic inside your network boundary while allowing managed updates over a controlled channel. On-prem goes further, keeping the data and models in your facility with egress you define. Only when policy forbids any external connectivity — classified environments, some defence and critical-infrastructure cases — does the full air gap become necessary.
The operational cost
- Update lag — model and security updates arrive on your cadence, not the vendor's.
- Observability — monitoring and log aggregation must run locally.
- Skills — GPU operations, model serving and capacity planning become your responsibility.
- Hardware — capacity planning needs headroom and spares because scaling is physical.
- Evaluation — quality checks run offline with your own tooling.
Common pitfalls
The most common mistake is over-scoping: buying an air gap for a workload that needed residency and access control. The second is under-planning the intake process, so updates stall indefinitely and the platform ages. The third is treating the air gap as a security guarantee on its own — insider risk, media handling and physical controls still apply. Decide with your security team, and document why the chosen tier is proportionate.
Honest comparison
| Capability | Plugsky | Hyperscaler AI platform | Building in-house |
|---|---|---|---|
| Air-gapped operation | On-prem and air-gapped tiers with open-weight models | Very limited offline support | Native if you build it |
| VPC isolation | Private endpoint available | Private networking available | You own the design |
| API portability | Same OpenAI-compatible API across tiers | Vendor-specific APIs | You define the interface |
| Updates | Controlled intake on your cadence | Cloud-managed updates | You manage everything |
| Support model | Defined per engagement, including restricted environments | Generally online-first | Your own team |
| Certifications | SOC 2 / ISO 27001 readiness in progress | Completed audits in many regions | Your own programme |
Frequently asked questions
Is air-gapped AI the same as on-prem?
No. On-prem means the systems run in your facility, but updates and support may use controlled external connectivity. Air-gapped means no network path in or out at all.
Which models work air-gapped?
Open-weight models that can be distributed and run offline, such as Nemotron, Llama, Qwen and Mistral families. Confirm licence terms for your commercial use.
How do updates work without the internet?
Through a documented intake process: media transfer, checksum verification, malware scanning and staged rollout. Plan the cadence so models and security patches do not drift.
Can we start on cloud and move air-gapped later?
Yes. The API contract is the same across tiers, so an application written against the OpenAI-compatible endpoint can move to on-prem or air-gapped deployment without a rewrite.
What does air-gapped AI cost?
Hardware, spares, power, licensing and operations staff dominate. Use a cost estimator to compare tiers, and remember that managed cloud tiers avoid most of these costs.
Do we lose auditability in an air gap?
No — audit logs, SIEM export and identity controls operate inside your environment. What changes is that telemetry cannot leave, so your local tooling must be sufficient.
When should we not choose an air gap?
When your requirement is residency, isolation or access control rather than absolute connectivity denial. A VPC or on-prem tier is usually proportionate and far easier to operate.