Serving infrastructure turns a trained artifact into a reliable production dependency. The model has to load, accept requests, fetch or receive features, batch work efficiently, use expensive hardware well, expose metrics, scale under load, and degrade safely when something downstream fails.

None of that is the model's job, and all of it determines whether the model is usable in production.

"Deploy it behind an API" is only the start. The infrastructure question is the request path itself: where the time goes, where the bottlenecks are, what signals tell you to scale, and what the system does when a dependency times out.