Start here. This is the direct spoken answer to practice first.
Overview
A gateway can simplify shared model access, but it also becomes another critical dependency.
I introduce an AI gateway when several applications need consistent authentication, quotas, routing, telemetry, provider policy, or protocol adaptation. A single small service can usually keep those concerns in its own model adapter. The gateway must solve a real cross-cutting problem because every request now depends on its latency, availability, and configuration.