Infrastructure

AI deployment options:
five places it can run.

This page exists so the decision can be made in one sitting. Each option is described by what it gives you and what it takes, with no recommendation attached — because the right answer genuinely varies and we build all five.

THE COMPARISON

	Public API	Your cloud	Private DC	On-premise	Air-gapped
Data leaves your perimeter	Yes	To your own tenancy	To your own rack	No	No
Frontier capability available	✓	✓	—	—	—
Choose the physical location	—	Region only	✓	✓	✓
Single tenancy	—	—	✓	✓	✓
Cost shape	Per token	Per token	Fixed	Fixed	Fixed
Capital outlay	None	None	Leased or owned	Owned	Owned
Vendor can deprecate your model	✓	✓	—	—	—
Works without an internet link	—	—	—	✓	✓
Operational burden on you	None	Low	Moderate	High	Highest
Time to first pilot	Days	Weeks	Weeks	4–10 weeks	8–16 weeks

Structural properties of each deployment model. Timelines are ranges we scope to, not measured client outcomes.

CHOOSING WITHOUT AN ARGUMENT

Work down the constraints in this order, and the answer usually falls out before you reach the bottom.

1. Is there a hard rule about where the data may go? A regulation, a customer contract, a classification. If yes, that eliminates options before anything else is considered, and no capability argument overturns it.

2. Does the system have to work without connectivity? A factory floor, a vessel, a remote site, an isolated control network. If yes, you are on-premise or air-gapped and the rest is detail.

3. Is the volume sustained and high? If yes, fixed-cost hosting will beat per-token pricing, and the crossover is worth calculating properly rather than estimating. If the volume is low or spiky, the API wins and will keep winning.

4. Does the workload need frontier capability for every request? Rarely true once tested, and if it is genuinely true then hosted is the answer. If it is true for a minority, that is a routing problem rather than an infrastructure one.

5. Is there anyone to operate it? The question that quietly decides more of these than the other four combined. If there is no team and no budget for managed operation, choose the option with the lowest operational burden that satisfies your constraints — and be honest about it now rather than in eighteen months.

THE COMMON ANSWERS

QUESTIONS

QUESTIONS — 3

If the system was built for it, yes and relatively cheaply. That means the model behind an interface, no dependency on a provider-specific feature, prompts and evaluation sets kept portable, and infrastructure defined as code. We build that way by default, because the most common regret we hear about is a system welded to a provider chosen in week one.

Most mature deployments do. Local for the bulk and sensitive work, hosted for the hard minority. That is the hybrid routing pattern and it is usually the honest answer to a question phrased as either-or.

We would expect it to split by constraint rather than by preference — regulated and high-volume work toward the controlled end, early-stage and exploratory work toward the hosted end. We would rather set out the trade-offs than tell you what is popular, because popularity is not evidence about your constraints.

Keep reading

Compare

On-Premise vs Cloud AI

Data Residency & Sovereignty

→

30-minute technical call · no deck