on-premise

In depth

Three reasons

Three reasons genuinely justify this choice. Data that cannot leave a perimeter, for regulatory or contractual reasons, which is the most frequent case in health, in defence and on certain industrial secrets. A need for absolute stability, when a system that has been evaluated and certified must not change behaviour because a supplier updated its model. And a volume high enough and regular enough for the amortisation of hardware to beat pay-per-use billing.

The real cost

The real cost comes in four items, only one of which is visible. The accelerators, expensive and quickly depreciated. The infrastructure around them: electricity, cooling, redundancy, network. Operations, meaning people able to serve a model, to watch over it and to update it, who are rare and expensive. And the opportunity cost, often the heaviest: time spent running a platform is not spent on the product. Comparisons that keep only the first item systematically reach the wrong conclusion.

Rarely binary

The practical conclusion is that this is rarely a binary choice. Common architectures are mixed: a self-hosted model for processing that bears on sensitive data, an online service for the rest, with a layer of abstraction that routes according to the nature of the data. This approach requires classifying your data before choosing your hosting, which is in any case a precondition of any serious governance, and it avoids imposing the most constraining regime on everyone.