Azure OpenAI Deployment
Does ZopNight manage Azure OpenAI Deployment?
An Azure OpenAI deployment bills one of two ways: Standard pays per token consumed, while a provisioned-throughput deployment bills every hour for its full PTU reservation whether or not a single token flows. ZopNight discovers each deployment with model, version, and PTU/TPM capacity and flags underutilized provisioned throughput.
Rules that fire on Azure OpenAI Deployment
No active rule family targets Azure OpenAI Deployment today. Rules that used to are retired, and retired rules publish no pages and fire no findings. Scheduling and permissions coverage are unaffected.
At a glance
| Field | Value |
|---|---|
| Scheduling notes | discovery and cost visibility only. |
An Azure OpenAI deployment is a named model instance (e.g. GPT-4o) with either pay-per-token Standard billing or provisioned throughput units billed hourly. PTU deployments bill their full reservation whether or not tokens flow.
Standard tokens versus provisioned throughput
The billing split matters more here than on almost any other Azure type. A Standard deployment costs exactly what it serves: tokens in, tokens out, metered per model at published rates. A provisioned-throughput deployment inverts that: you reserve PTU capacity for predictable latency and throughput, and the reservation bills by the hour at its full size regardless of load. Quiet nights, idle weekends, and abandoned experiments all bill identically to peak traffic. The trade is deliberate (guaranteed capacity in exchange for a standing cost), but it only pays off when utilization stays high enough to beat the per-token alternative.
Capacity detail ZopNight records per deployment
Each deployment is discovered through the account’s data plane by the Azure OpenAI enricher, carrying its model name, model version, and capacity figures (PTU for provisioned deployments, TPM rate limits for Standard ones). Cost Management billing attributes actual spend, and recommendation rules single out underutilized provisioned-throughput capacity. Deployments are not schedulable: there is no stop verb on a deployment, and scaling PTUs up or down is a manual capacity decision, so ZopNight surfaces the underutilization as a recommendation rather than acting on it automatically.
How PTU reservations leak money
Three patterns dominate. A deployment sized for a demo or load test that nobody deleted afterwards, billing its reservation for months. Provisioned capacity bought for business-hours traffic, where roughly two-thirds of every day contributes nothing but still bills. And a model upgrade that left the old deployment running in parallel “just in case”. Two reservations now serve one workload.
Managing deployments in AI Foundry
Azure AI Foundry portal → Deployments lists every deployment under the selected resource with its model, version, and provisioned capacity; the same list is reachable from the Azure portal via the account’s overview.