Deployment Shapes — Hosted, VPC and On-Prem

The first question an enterprise customer's security team asks is not "how does your product work." It is "where does it run." The answer you give decides three quarters of the engineering effort, the whole shape of the support contract, and how long it takes to ship a bug fix to a user who found one.

There are three answers that occur in the wild. An FDE should know all of them and know how to pick between them, because the person who gets to pick in the room is usually the one who has already read this lesson.

The three shapes

Each shape exists because the one above it could not meet a specific requirement. The right instinct is to default to the leftmost shape the customer's actual constraints allow — and the word "actual" is doing real work in that sentence.

The question to ask before promising anything

A deal is closing, an AE says the customer wants VPC. Before quoting a timeline, the question an FDE asks is this:

Whose security review said so, and would a hosted deployment also pass it?

The three answers you will hear:

• "Their CISO's team required it after reviewing the questionnaire." This is the real version. Honour it.
• "Nobody asked for it, but the AE thought it would help the deal land." This is where you intervene. VPC is slower to ship to, more expensive to operate, and buys the customer nothing they did not already have — and once you have promised it, you do not get to undo it without a renegotiation.
• "The customer asked because our competitor offers it." This one deserves a conversation with the customer about what they are actually trying to buy. The answer is sometimes genuine (regulatory, usually) and sometimes a procurement reflex that disappears when someone names the cost.

An FDE who asks this question early is usually the most expensive person in the room, and that is the right instinct to carry into a kickoff.

What each one costs you, honestly

Three dimensions: the money, the ship speed, and the incident response.

Hosted is the cheapest to run, because one environment serves everybody and every improvement reaches every user the moment you deploy it. It is the fastest to ship to, because "deploy to production" is a command you run yourself. It is the easiest to support in an incident, because your on-call engineer can see the state of the system in real time and act on it.

VPC is the most interesting one, because it is the one teams most often underprice. The compute is the customer's — not a saving you keep, because the support contract still has to cover you running it. The ship cycle becomes theirs — your Friday release is their next change-window Tuesday, and your hotfix is their urgent-change process. The incident response becomes a joint exercise, because your engineer cannot log into their environment. The honest version is: VPC is a 2–4× multiplier on your cost to serve a customer, and the cost is almost entirely in people, not in machines.

On-prem is a step beyond VPC. Updates become a release train. The product versions you support in the field drift apart. Support happens over screen shares and shipped bundles. Observability depends on telemetry the customer chooses to export. It is a real business, but it is not the business a frontier lab's FDE is usually optimising for, and it is almost always a mistake to take it on unless the customer's environment is genuinely air-gapped — defence, certain regulated government, specific pharmaceuticals and banks. In anything else, VPC is closer to the security posture for a fraction of the operational cost.

The hybrid shape nobody puts on the slide

The deployment matrix almost always has a fourth entry in practice: hosted application, customer-hosted model. The code runs in your SaaS, the model endpoint lives in the customer's cloud or an endpoint they trust. This covers a common objection — "the model vendor is in a jurisdiction we cannot use" — without the full VPC commitment, because the application code is still yours to ship.

The variants that work in the field:

• Hosted app, model in a region the customer specified.
• Hosted app, model on a provider endpoint the customer owns.
• Hosted app, model on an inference server the customer hosts behind their own gateway.

Not every customer will accept these, but a surprising number of enterprise security teams prefer them to VPC once they understand the trade — because the thing they actually cared about was the model vendor, not the application.

What to take away

There are three real deployment shapes — hosted, VPC and on-prem — and one useful hybrid. Default to hosted because it is the cheapest to run, the fastest to ship to, and the easiest to support in an incident. Move to VPC when a customer's security review actually requires it, not when an AE promised it, and know that it typically multiplies your cost to serve by two to four times, almost all of it in people. Only reach for on-prem when the customer's environment is genuinely air-gapped, because VPC gives nearly the same security posture for a fraction of the operational cost. And remember that the shape of the model call is a separate question from the shape of the application — the hybrid of hosted app with a customer-pinned model endpoint quietly solves many of the objections that look like VPC-or-nothing.

Next: once a system is deployed, how do you see what it is doing, and what does an incident response actually look like when the thing on fire is the agent itself.