Some data cannot leave a perimeter: classification, trade secret, or a simple documented refusal of legal exposure. The model then has to be installed where the data is.
The model runs in your subscription, in the region you chose. The simplest option, and sufficient in most cases.
OVHcloud, Scaleway, Infomaniak or Delos, depending on the jurisdiction and the qualifications your sector requires.
On your own hardware, with no outbound link. The most demanding: manual updates, operations on you, a far higher cost per request.
By language, context size and available hardware. A smaller, well-evaluated model often beats a larger, badly integrated one.
Where does the constraint come from: a law, a contract, an internal policy, or a habit? The answer changes the cost by a wide margin, and sometimes the constraint falls away.
Request volume, acceptable latency, hardware needed. This is the item projects most underestimate.
Model, inference service, indexing, monitoring. Described in code like the rest of the infrastructure.
Training for your team, an update procedure, and an exit plan if the requirement disappears.
No request leaves the perimeter. It is verifiable, and we verify it in front of you.
Why this degree of isolation, and what it costs. That is the document that protects you when the decision is reopened.
If the requirement changes, you know how to move back to a managed service without rebuilding everything.
Tell us where the constraint comes from. Sometimes the analysis concludes a managed service is enough, and that is good news for your budget.
See your external surface →