Infrastructure · We run it ourselves
Amazon's cloud. We use one narrow part of it — Bedrock, which serves several vendors' models behind one API without the prompt leaving the account.
Naming AWS on a capability page is nearly meaningless, because AWS is two hundred services and nobody uses more than a handful. So this entry names the one we actually run and states plainly what we do not, which is most of it.
01
The question a model integration really raises is where the prompt goes. A client's pricing logic, contract text or customer conversation sent to a public endpoint is a disclosure decision, and it is usually made by a developer choosing a library rather than by anyone who would recognise it as one. Bedrock's answer is that inference runs inside the account and the region chosen, and that prompts are not used to train the underlying models.
02
It decouples the choice of model from the shape of the integration. Several vendors sit behind one API, so replacing a model is a configuration change rather than a rewrite — which matters because the best model for a task now is reliably not the best model for it in six months, and a system that cannot switch will be carrying last year's answer quality until somebody funds a migration.
03
Region selection is a compliance argument before it is a latency one. For a client whose data cannot leave the Gulf, or cannot leave the EU, the deployable set of options is small and it is decided at the infrastructure layer rather than in the application.
04
We go direct to a model vendor when the reverse is true: when the newest capability matters more than the residency guarantee, or when the volume is small enough that the operational overhead of a cloud account is not worth carrying. There is no vendor commission on either route, so the recommendation is not shaped by which one pays.
05
Cost behaves differently from a subscription. Inference is billed by token, which means a feature nobody uses costs nothing and a feature that succeeds gets expensive in proportion — a property worth designing around from the start rather than discovering in the second month's invoice.
An entry that cannot point at something you can open, run or read does not compile. That is a property of the type, not a promise in a paragraph.
Invariant
We use Bedrock and the AWS SDK. We do not run a large AWS estate, and this entry is not a claim to one: no production Kubernetes on EKS, no multi-account landing zone, no data lake. A client whose problem is a sprawling AWS bill or a migration off legacy infrastructure needs a specialist cloud practice, and saying so costs us a project we would have run badly.