Running AI workloads on Azure

Azure does not sell you a model. It sells you a resource, and then a deployment inside that resource, and then a quota allocation attached to the deployment — three objects that are created separately, billed separately and fail separately. Most of the confusion people have with Azure OpenAI comes from treating those three as one thing, which is what the portal’s single create-and-deploy wizard quietly encourages. The first page in this cluster pulls them apart, and nearly every other page here depends on that separation being clear.

The rest works outward from the request. What the rate limiter counts and when it counts it, what a provisioned throughput unit actually buys per minute, how a content filter interposes itself between the model and your stream, what changes when the endpoint is only reachable inside a virtual network, and what a durable orchestrator gives you that a retry loop does not. Where a figure is Microsoft’s, it is dated and linked; where a figure moves, the page says which quota line or API to read for the current one rather than guessing on your behalf.