# Azure ML Online Endpoint

> Azure ML online endpoints are the HTTPS front door for real-time inference. The endpoint itself is free, while the deployments behind it bill per instance-hour. ZopNight discovers each endpoint with its deployment list and traffic split, so inference spend rolls up under the serving topology it belongs to.

Source: https://zop.dev/integrations/azure/ml-online-endpoint
Updated: 2026-08-19

---

Online endpoints serve real-time model inference over HTTPS, fronting one or more deployments on provisioned compute. The endpoint routes traffic; its deployments hold the billable instances.

## Router on top, meters underneath

An online endpoint is the stable HTTPS address and auth boundary for real-time scoring; it carries no instance meter of its own. The money sits one level down, in the deployments it routes to, each a provisioned set of CPU or GPU instances billing continuously. An endpoint with a 90/10 traffic split across two deployments is therefore two separate bills behind one URL, and the 10-percent deployment often costs as much as the 90.

## Serving topology as ZopNight records it

Discovered via the AML enricher with deployment and traffic-split context, letting inference spend roll up under the serving topology. That rollup is what makes an inference bill explainable: instead of a flat list of VM charges, each charge hangs off the endpoint (and so the model, the team, the purpose) it serves.

## Why there is no stop verb here

ZopNight treats endpoints as discovery and topology only, because billing sits with the deployments behind them. Deleting or pausing an endpoint would break the serving contract without touching the instance meters. Cost action on real-time inference means resizing or removing deployments, and the endpoint page is where those candidates become visible.

## Endpoint-level signs of waste

Watch for endpoints whose traffic split still routes a share to a superseded deployment, the classic leftover of a blue-green rollout that was never finished; endpoints kept alive for demos with fully provisioned deployments behind them; and duplicate endpoints serving the same model for teams that stopped talking to each other.

## Locating endpoints in ML studio

Azure ML studio → Endpoints → Real-time endpoints shows each endpoint with its deployments and traffic percentages. Read the split first; it is the map of where the instance-hours go.