A client wants to self-host rather than use a commercia...

Four questions, in that order:

1. Can the data leave the environment? If not, self-hosting (or an in-perimeter managed option) stops being a preference and becomes a hard requirement.
2. What’s the actual volume? Self-hosting has to clear a real utilization bar to beat pay-per-token pricing — GPUs sitting mostly idle are a worse deal than the API they were meant to replace.
3. What capability does the task genuinely need? A narrower open-weight model is often enough for a well-scoped task, and “needs the biggest frontier model” is worth actually testing rather than assuming.
4. Who operates the deployment after handover?

That last one is the forward-deployed-engineer question, and it’s the one most people skip. A self-hosted deployment is infrastructure the client’s own team has to run, patch, and scale after I leave. It’s often the right call to ship something slightly less optimal that actually survives handover, over the more elegant option that quietly becomes unmaintained six months later.