Feature × Audience

How can MSPs deliver air-gapped AI with Plugsky?

MSPs deliver air-gapped Plugsky as a managed sovereign service: a standard reference install per client estate, open-weight models sized to their GPUs, and a repeatable update bundle pipeline you operate on their behalf. One OpenAI-compatible integration serves every client, while per-client workspaces, keys and audit streams preserve isolation without maintaining separate codebases.

Key facts

Delivery modelRepeatable reference install per client estate
API continuityOne OpenAI-compatible integration across all clients and tiers
TenancyWorkspace and scoped keys per client, with per-client audit exports
DeploymentOn-prem and air-gapped tiers alongside cloud and VPC options
UpdatesStandard bundle pipeline with checksums, staging and rollback
Key managementCustomer-managed keys in the client's KMS or HSM
AuditAccess, key and inference events exported to client SIEM
StatusLive platform; enterprise deployment terms per engagement

TL;DR

  • Industrialise one reference install instead of bespoke builds per client.
  • Operate a standard update bundle pipeline with staging and rollback.
  • Keep per-client workspaces, keys and audit streams for isolation.
  • Support remote estates with bundle-based diagnostics, not outbound access.
  • Charge for the managed service, not the inference.

How it works, step by step

  1. Define a reference architecture: model set, GPU sizing bands, gateway, key integration and log path that cover most client estates.
  2. Qualify clients against it — data sensitivity, network classification, hardware — and identify the exceptions.
  3. Build a provisioning kit: install runbook, model inventory, evaluation suite and client audit export.
  4. Operate a shared update pipeline producing verified bundles for every client, with staging validation and rollback.
  5. Keep a workspace and scoped key set per client, even in single-tenant estates, so reporting and offboarding stay clean.
  6. Package service tiers: deployment, quarterly model refresh, evaluation reporting and incident support.
  7. Rehearse recovery and failover with each client before go-live, and record the evidence in your service file.
1Define a referencearchitecture: modelset, GPU sizing2Qualify clientsagainst it — datasensitivity,3Build aprovisioning kit:install runbook,4Operate a sharedupdate pipelineproducing verified5Keep a workspaceand scoped key setper client, even in6Package servicetiers: deployment,quarterly model

Try it yourself

Open the self-hosting break-even calculator →

Productising air-gapped delivery

Air-gapped projects usually fail on bespoke engineering, not on models. MSPs win by standardising: one reference architecture with defined model sets and GPU sizing bands, one provisioning kit, one update process. Clients then receive a known quantity with documented behaviour rather than a research project.

Plugsky helps because the API contract is stable across tiers. Your integration, evaluation harness and runbooks are written once and reused in cloud, VPC, on-prem and disconnected estates, so each new client is a deployment exercise rather than a development one.

Update bundles as a managed service

The recurring value in an air-gapped estate is the refresh cycle: new models, runtime fixes and security patches delivered safely. Build a pipeline that produces verified bundles with checksums and release notes, validates them on your own staging estate or the client's mirror, and installs with a rollback path.

  • Schedule: quarterly model refreshes and emergency security releases as separate tracks.
  • Validate: run the client evaluation suite against the candidate bundle before promotion.
  • Document: every installation leaves an updated manifest and change record.

Support and isolation without egress

Remote support in a disconnected estate is a process, not a connection. Ask clients to export diagnostics, metrics and audit summaries; analyse offline; and return an installable fix. Keep per-client workspaces and keys so access and reporting stay separated even when you operate multiple estates.

On the commercial side, air-gapped work justifies recurring fees because it is genuinely continuous: capacity reviews, refresh cycles, evaluation reporting, incident response and audit support. Price those as service lines and be explicit about what a client's own team owns, from hardware to key custody.

Honest comparison

MSP concernPlugsky air-gappedRe-selling cloud AIBespoke builds
RepeatabilityOne reference install and kitSingle hosted tenancyEach client differs
Update deliveryStandard verified bundlesProvider-managedAd hoc per client
IsolationWorkspace and keys per clientShared platformVaries
Support modelExport-based remote diagnosticsOnline accessWhatever you improvise
Margin shapeRecurring service feesThin resale marginProject-based
API portabilityOpenAI-compatible offlineVendor APIYou define it

Frequently asked questions

Do we need a separate codebase per client?

No. Keep one integration against the OpenAI-compatible endpoint and vary workspaces, keys, models and hardware per client. That is what makes the service profitable.

How do we update remote estates?

With a standard bundle pipeline: checksums, release notes, staging validation against the client evaluation suite, then installation with a tested rollback path.

How do we support a network with no egress?

Through exported diagnostics and metrics, offline analysis and installable fixes. Agree the process in the contract; expect it to be slower than online support.

Who owns the keys?

The client does. Customer-managed keys live in their KMS or HSM, and key administration should be separated from platform administration.

How do we price managed air-gapped AI?

As recurring service fees covering deployment, refresh cycles, evaluation reporting, incident support and audit assistance. Avoid pricing on inference volume, which is not the value.

Can we use the same team for cloud and disconnected clients?

Yes, if the integration, evaluation suite and runbooks are shared. The interface is identical; only the deployment environment differs.

How do we win the first air-gapped client?

Start with a deployment on their staging network using open-weight models and a bounded workload, and prove the update and recovery processes. That evidence wins the larger rollout.