Skip to content
Echnotek
Deployment Models

AI Deployment Models: Private Cloud, Customer VPC, On-Premise and Air-Gapped, Explained

"Private AI" isn't one shape of infrastructure — it's four, and they're not interchangeable. Where the model runs, who operates it, and how far a connection can reach outside it are different decisions with different tradeoffs, and picking the wrong one is a common way private AI projects stall after the architecture is already agreed.

This page explains each of the four on its own terms — what it actually is, how it's operated, and who it's genuinely the right fit for — rather than as a single row in a comparison table on a different page.

Four models. Not interchangeable. One sentence each below.
At a glance

The Four Models, Quickly

The full profile of each is below — this is the version you'd screenshot into a slide.

Deployment model quick comparison
ModelInfrastructure ownerBest fit
Private CloudEchnotek, dedicated tenancyFast production launch, named contracted processor is acceptable
Customer VPCYou, in your own cloud accountData can't leave infrastructure you already own
On-PremiseYou, on your own hardwareRegulation requires physical control, not just a cloud account
Air-GappedYou, fully offlineConnectivity itself — not just data exposure — is the risk
The four models

Every Model, in Full

Click through each one — what it is, how it's operated, and what it actually trades off against the others.

A single-tenant deployment that Echnotek hosts and operates in a region you specify — no shared inference, no multi-tenant model instance, but the infrastructure itself is ours, run under contract. This is the fastest of the four to reach production, because there's no cloud account setup or hardware procurement on your side: the architecture, sizing and operations are handled end to end. The tradeoff is that you're accepting a named, contracted processor rather than holding the infrastructure yourself — which is a real constraint for organisations whose policy requires the latter, and a non-issue for most others.

FASTEST TO PRODUCTION, A NAMED PROCESSOR.

Every layer — the model, the retrieval index, the agent runtime — deploys directly into your own AWS, Azure or GCP account: your VPC, your IAM policies, your bill, your logs. Echnotek builds and operates the system there, but you hold the infrastructure, the keys and the network boundary throughout. This is the most common choice for enterprises that already run cloud infrastructure and want the system to live inside the account structure their own platform team already manages, without taking on the operational burden of running GPUs themselves.

YOUR ACCOUNT, OUR OPERATIONS.

The full stack runs on hardware inside your own data centre — GPUs sized and provisioned by Echnotek, deployed on-site, with your infrastructure team trained to operate it independently afterward. This is the model regulated and defence-adjacent buyers reach for when a policy or regulator specifically requires physical control of the hardware, not just administrative control of a cloud account — a distinction that matters more than it sounds like it should to some compliance frameworks. It's also the slowest of the four to stand up, since it depends on hardware procurement and installation rather than provisioning into an existing account.

PHYSICAL CONTROL, LONGER LEAD TIME.

The same architecture as the other three, with no route to the public internet at all — model updates, monitoring exports and any other data movement happen through controlled, manual transfer rather than a live connection. This is the model for classified, defence-adjacent or maximum-restriction environments, where the risk being managed is connectivity itself, not just what data might be exposed over it. It's the most operationally demanding of the four: routine tasks like updating a model version become a scheduled, manual process rather than a deployment pipeline.

MAXIMUM ISOLATION, MANUAL UPDATES.
Choosing between them

Which One, in One Sentence Each

Most organisations know within a sentence which of these four they actually need — the rest is confirming it.

01
Speed matters most, a contracted processor is acceptable

Choose Private Cloud — it's live fastest, with the infrastructure operated for you.

02
You already run cloud infrastructure, and data can't leave it

Choose Customer VPC — the system deploys into the account you already control.

03
A regulator or policy requires physical hardware control

Choose On-Premise — the only option where the hardware itself, not just the account, is yours.

04
Connectivity itself is the risk being managed

Choose Air-Gapped — the only option with no route to the public internet at all.

FAQ

Deployment Models FAQ

Yes — this is a common path. Launching in a private cloud to validate the system quickly, then migrating to on-premise once hardware is procured and a longer-term policy requirement is confirmed, reuses almost all of the architecture and build work.

Who owns the infrastructure. Private cloud runs on infrastructure Echnotek owns and operates under contract; customer VPC runs inside a cloud account you already own, with Echnotek operating the system there. Both are single-tenant and equally private — the difference is whose account the bill and the keys sit in.

Eventually, if you want to — Echnotek sizes, procures and deploys the hardware and trains your infrastructure team to operate it independently. Some clients keep us operating it under contract instead; both are available on the same on-premise model.

For most, yes — customer VPC or private cloud already satisfies the data-control requirement without the operational overhead of manual updates. Air-gapped earns its cost specifically when connectivity itself, not just data exposure, is the thing a regulator or policy is managing.

Private cloud is fastest — often weeks. Customer VPC adds time for account and IAM setup on your side. On-premise depends on hardware procurement and installation, which is typically the longest lead time of the four. Air-gapped adds process overhead for the manual transfer procedures, on top of whichever physical deployment it's built on.

Where this fits

See These Four Applied

This page explains the four models on their own terms — here's where they show up applied to a real decision.

Start here

Not sure which of the four fits your constraints?

Bring your infrastructure and security leads to a discovery call. We'll walk through what you already run, what's actually required, and which of these four models fits — before anything is built.

Start with a conversation

Let’s talk now