Feature × Audience

Do startups need air-gapped AI, and how does Plugsky fit?

Most startups do not need air-gapped AI. Building for disconnected deployment early slows you down and burns capital on hardware you cannot utilise. The practical strategy is hybrid: develop against Plugsky's OpenAI-compatible cloud endpoints, keep the integration portable, and add an on-prem or air-gapped tier when a signed enterprise or public sector deal actually requires it.

Key facts

When neededRarely early; driven by customer contract or regulation, not by preference
Default pathBuild on the cloud endpoints with plugsky-micro and plugsky-lite free, no card
Trial14-day full-access trial for heavier prototyping
PortabilityOpenAI-compatible API means the same code runs in every tier
Hardware realityAir-gapped means GPUs you own, sized for peak with no bursting
Hybrid optionRoute sensitive workloads to a private or on-prem tier, keep the rest hosted
Update burdenYou own model and runtime refreshes in a disconnected estate
StatusOn-prem and air-gapped tiers are live; enterprise terms per engagement

TL;DR

  • Do not build for an air gap before a customer requires one.
  • Develop on the free cloud tier and keep the API integration portable.
  • A private endpoint in your VPC or the customer's solves most residency asks.
  • Only go disconnected when the contract pays for the hardware and support burden.
  • Because the interface is OpenAI-compatible, adding the tier later is a deployment, not a rewrite.

How it works, step by step

  1. Ask what the customer actually requires: region-locked processing, VPC isolation or genuinely no egress.
  2. Build your product against the hosted OpenAI-compatible endpoints using the free models for day-to-day development.
  3. Keep base URL, model names and feature flags in configuration so deployment target is not hard-coded.
  4. For most enterprise objections, offer a VPC or on-prem tier with controlled egress before discussing a full air gap.
  5. If a deal requires no egress, price the hardware, deployment and update burden into the contract.
  6. Deploy open-weight models sized to the customer's hardware, and gate features the local models cannot support.
  7. Keep one evaluation suite running against every tier so quality expectations stay honest.
1Ask what thecustomer actuallyrequires:2Build your productagainst the hostedOpenAI-compatible3Keep base URL,model names andfeature flags in4For most enterpriseobjections, offer aVPC or on-prem tier5If a deal requiresno egress, pricethe hardware,6Deploy open-weightmodels sized to thecustomer's

Try it yourself

Open the private LLM cost estimator →

Why air gaps rarely fit startups

An air-gapped estate is hardware you buy, size for peak, patch yourself and cannot burst. For a startup, that capital and operational load competes directly with product development, and the market that requires it is small relative to the effort. Most early AI startups should let the customer's requirement pull them into disconnected deployment, not push it speculatively.

The classic mistake is building the hard version first: a bespoke on-prem release with a divergent codebase that slows every subsequent feature for the majority of customers who never needed it.

The hybrid path that actually sells

Start hosted. Use plugsky-micro and plugsky-lite on the free plan to build and demo, and move to a paid plan as usage grows. When a customer raises residency, offer the middle options first: region-locked planes, a private endpoint in your VPC or theirs, or an on-prem deployment with controlled egress. These satisfy most procurement and security reviews at a fraction of the air-gap cost.

  • Keep it configurational: base URL, model alias and deployment flags, never hard-coded.
  • Gate features: smaller open-weight models will not match frontier quality on every task.
  • Price honestly: deployment, refresh cycles and support are recurring costs, not one-off favours.

If a deal truly requires an air gap

Then treat it as a deployment project with a business case. Size GPUs for the customer's concurrency, plan the update bundle process, decide who owns keys and recovery, and confirm who pays for refresh cycles after go-live. Plugsky runs the same OpenAI-compatible endpoints offline with open-weight models, so your application layer does not need a second implementation.

Negotiate scope tightly: one or two workloads, one network segment, a named evaluation set and a support agreement. A successful bounded deployment is a reference; an unbounded one is a sink for engineering time you should be spending on the core product.

Honest comparison

Startup decisionStart hostedPrivate or on-prem tierFull air gap
Upfront costNoneModerateHigh — hardware and setup
Time to first customerDaysWeeksMonths
Satisfies residencyRegion-locked planesOften yesBeyond most requirements
Update burdenProvider-managedSharedEntirely yours
Feature qualityFull model rangeFull model rangeLimited to local models
When to chooseDefault for startupsEnterprise security reviewRegulated contract only

Frequently asked questions

Do we need GPUs to start?

No. The hosted API handles inference, and the free plan with plugsky-micro and plugsky-lite needs no card. Buy hardware only when a customer contract requires disconnected operation.

How do we answer a residency question in a sales call?

Offer region-locked planes first, then a private endpoint or on-prem deployment with controlled egress. Most buyers accept these; a full air gap is a last resort with real cost implications.

What breaks if we build air-gap first?

You slow every feature for every customer, carry hardware cost early and create a divergent codebase. The requirement usually belongs to one deal, not the product.

Can we support an air-gapped customer later without a rewrite?

Yes. The API is OpenAI-compatible across tiers, so adding a disconnected deployment is configuration, packaging and hardware — not a second implementation.

Who pays for model refreshes in a disconnected estate?

The customer, contractually. Refresh cycles, evaluation and support are recurring work; price them into the agreement rather than absorbing them.

What about a hybrid model?

Common and sensible: keep most workloads hosted and route sensitive ones to a private or on-prem tier through the same API. You get residency where it matters without running everything yourself.

How do we evaluate local model quality?

Keep one evaluation suite per feature and run it against every tier. Gate features behind capability checks when the local open-weight model cannot meet the bar.