Key facts
| Definition | A model API accessible only from your private network or environment |
| Deployment models | VPC, private link, on-prem and air-gapped |
| Why it matters | Keeps prompts, outputs and logs inside a compliance boundary |
| Access control | Network-level restriction plus scoped API keys, RBAC and audit logs |
| API compatibility | OpenAI-compatible, so applications change only the base URL |
| Trade-offs | Higher operational cost and slower model updates than a shared endpoint |
| Verification | Network diagrams, endpoint tests and egress monitoring prove isolation |
| Status | Private deployment options are available; confirm current regions and terms with sales or docs |
TL;DR
- Private means network reachability, not just a dedicated URL.
- VPC and private link cover most regulated workloads.
- On-prem and air-gapped serve the strictest isolation requirements.
- The API stays OpenAI-compatible, so code changes are minimal.
- Verification is part of the deliverable: prove traffic cannot leave.
How it works, step by step
- Classify the data the workload processes and the isolation level it requires.
- Choose the deployment model: private link, dedicated VPC, on-prem or air-gapped.
- Provision the endpoint inside your network and restrict ingress to known clients.
- Scope API keys and roles to the smallest set of models and actions needed.
- Test that the endpoint is unreachable from the public internet and log egress attempts.
- Review patching, model updates and support access against your policy.
Try it yourself
Open the private LLM cost estimator →
What makes an endpoint private
Privacy is a network property. A dedicated URL on the public internet is not private — it is just uniquely addressed. A private endpoint is reachable only through your VPC, a private link, a VPN or a node with no external route at all. Combined with scoped credentials and audit logging, that means prompts, completions and telemetry stay inside your boundary, and the provider's shared infrastructure never sees the traffic.
Deployment models compared
- Private link: the provider's service exposed into your cloud network without public egress.
- Dedicated VPC: model serving runs in your own cloud account and region.
- On-prem: serving runs in your data centre on your hardware.
- Air-gapped: no internet at all; updates arrive through controlled media.
Each step increases isolation and operational burden. Match the level to the data classification, not to ambition.
Common mistakes
- Calling a dedicated public URL 'private' because the hostname is unique.
- Forgetting support access: remote provider engineers can defeat in-network deployment if not contracted otherwise.
- Allowing broad egress from the workload subnet, which reopens a data path.
- Using a single all-powerful key for every service in the VPC.
- Assuming private means unlimited scale — capacity is provisioned, not elastic.
Private endpoints on Plugsky
Plugsky supports VPC, on-prem and air-gapped deployments, keeping model serving and data in your environment while the API remains OpenAI-compatible. Applications change only the base URL, and platform controls such as scoped keys, RBAC and audit logs still apply. Because capacity is dedicated rather than shared, plan sizing and model updates with the team. For regulated buyers, pair the deployment with contractual terms and evidence such as network diagrams and egress monitoring to satisfy procurement review.
Honest comparison
| Model | Reachability | Ops burden | Typical fit |
|---|---|---|---|
| Public API with region selection | Internet | Lowest | Commercial apps |
| Private link or VPC | Your cloud network only | Moderate | Regulated enterprises |
| On-prem | Your data centre | High | Strict data controls |
| Air-gapped | No external network | Highest | Government, critical infrastructure |
Frequently asked questions
What is a private AI endpoint?
A model API reachable only from your own network — through a VPC, private link, VPN or fully disconnected environment — so requests and responses never traverse the public internet.
Is a dedicated API URL the same as a private endpoint?
No. A unique public URL is still public. Privacy requires network-level restriction such as private link, VPC peering or an air-gapped deployment.
Does a private endpoint change my application code?
Usually not beyond configuration. Plugsky's private deployments remain OpenAI-compatible, so you change the base URL and keep the SDK.
What are the trade-offs?
Higher operational cost, provisioned rather than elastic capacity, and slower updates. In exchange, data never leaves your controlled network.
How do I verify isolation?
Test that the endpoint is unreachable from the internet, monitor egress from the workload subnet, and document network diagrams plus access controls for audit.
Can I use Plugsky models on a private endpoint?
Yes. VPC, on-prem and air-gapped deployment options are available; confirm current regional availability and contractual terms with the team.