How does government run on-prem AI with Plugsky?

Key facts

Government fit	Accredited applications call one API inside the agency perimeter
Deployment	Same OpenAI-compatible API in your VPC, on-prem or air-gapped
Data path	Prompts, embeddings and logs never leave the agency network
Models	30+ models; air-gapped bundles confirmed under the licensing agreement
Live endpoints	Chat, streaming, function calling, JSON mode and embeddings
Identity	Agency-held scoped keys, SSO with SCIM and RBAC for operators
Audit	Authentication, key lifecycle and admin events exportable to SIEM
Pricing	Flat monthly self-serve plans; enterprise deployment scoped on the pricing page

TL;DR

• Match deployment to classification: on-prem for controlled data, air-gapped for disconnected sites.
• Keep keys in agency custody and rotate them under documented dual control.
• Reuse the OpenAI-compatible API so accredited applications stay unchanged.
• Give assessors one evidence path: auth, key and admin events in the agency SIEM.
• Agree model bundle and update procedures before the accreditation review, not after.

How it works, step by step

1. Record the classification and handling rules for each candidate workload; these decide whether VPC, on-prem or air-gapped deployment is permissible.
2. Design the network zone and egress rules so only accredited subnets can reach the model endpoint, and only approved destinations are reachable from it.
3. Put key custody in the agency's secret management with documented rotation, dual control and an auditable issue and revoke process.
4. Integrate SSO with SCIM for operator access and RBAC for policy changes, and keep human and service identities separate.
5. Export authentication, key lifecycle, admin and inference events to the agency SIEM, and align prompt retention with records policy.
6. Run a threat model and a limited pilot on unclassified data with a named owner and defined exit criteria.
7. Prepare the accreditation pack: topology, data flows, controls, evidence sources, update and incident procedures.

Try it yourself

Open the sovereign AI readiness score →

When on-prem or air-gapped is the only option

Shared cloud is acceptable for many government workloads, but classification, sovereignty or network isolation rules the door for others. On-prem deployment keeps inference and logs inside the agency network while applications continue to call one OpenAI-compatible endpoint. Air-gapped deployment goes further: the platform runs in a disconnected enclave, with model bundles supplied under the licensing agreement and updates carried in through the agency's own process.

Because the API surface is identical across variants, an application accredited against a pilot deployment does not need rework when it moves to a more isolated environment — the topology changes, the integration does not.

What assessors will ask

Accreditation reviews focus on control ownership and evidence. Be ready with a concrete story for each area.

• Key custody: agency-held secrets, documented rotation, dual control, prompt revocation on role change.
• Identity: SSO with SCIM for staff, scoped keys for services, RBAC for administrative actions.
• Data flow: which subnets reach the endpoint, where logs and embeddings live, what retention applies.
• Evidence: authentication, key and admin events in the SIEM, joined with model metadata.
• Change: how platform and model bundle updates are approved, applied and rolled back.

Operations, cost and honest trade-offs

Self-hosting is not free: the agency owns capacity, patching, monitoring and incident response, and the procurement case should include those costs alongside licensing. Compare against VPC deployment and the flat monthly plans on the live pricing page before assuming on-prem is cheaper — it is chosen for control, not economy.

Scope statements should be exact: Plugsky provides the model API, deployment variants, scoped authentication and audit events. It does not provide identity proofing, classification authority, records policy or accreditation. Endpoint coverage is the same everywhere: chat, streaming, JSON mode, function calling and embeddings are live, while audio, images, files and batch are labelled coming soon.

Honest comparison

Concern	Plugsky on-prem or air-gapped	VPC deployment	Shared public cloud
Data location	Agency network or enclave	Agency cloud tenancy	Vendor infrastructure
Classification fit	Controlled through classified, subject to assessment	Controlled and lower	Unclassified or low sensitivity
Key custody	Agency secret management	Agency cloud secrets	Vendor-managed keys
Update path	Agency-controlled import and rollout	Provider-managed	Vendor-managed
Effort	Highest: capacity and accreditation	Moderate	Lowest

Frequently asked questions

Can air-gapped deployments receive updates?

Yes, through the agreed release process under the licensing agreement: bundles and platform releases are imported and applied under agency change control rather than pulled from the internet.

Does the API change between cloud and air-gapped?

No. The same OpenAI-compatible API is exposed, so applications and evaluations do not change when the deployment topology changes.

Who holds the API keys?

The agency does, in its own secret management, with documented rotation, dual control and revocation procedures that assessors can inspect.

What is in the accreditation evidence pack?

Topology and data flows, identity and key controls, SIEM evidence sources, retention settings, and change and incident procedures.

Which capabilities are available on-prem?

Chat completions with streaming, JSON mode and function calling, plus embeddings, are live. Audio, images, files, batch, moderation, fine-tuning and assistants are labelled coming soon.

How do we avoid over-buying capacity?

Size on peak concurrency and context length with a headroom margin, and keep a pilot scope small until utilisation data exists.

Is there a lower-effort intermediate step?

Yes. VPC deployment keeps workload data inside an agency tenancy with less operational burden than full on-prem; use it where policy allows.