The address your application actually calls, and everything behind it that turns a request into a response. Whether it is a hosted provider or something you run yourself, this is the boundary where availability, latency, cost and security all become concrete.
Technical view
Concerns include authentication and authorisation, request validation, timeout and cancellation semantics, streaming support, concurrency limits, and versioned routes so clients are not broken by changes. Self-hosted serving adds model loading, batching strategy and memory management. The endpoint is also the natural place for cross-cutting controls: rate limiting, guardrails, logging and cost attribution per caller.
Related
Part of AI Terminology Index, section: Running It in Production.