Menu
Products

Azure OpenAI

5 min readintermediateUpdated 28 Sept 2026
1 · In one line

Azure OpenAI lets companies use OpenAI's models through Microsoft's Azure cloud, with Azure billing, safety filters and data controls.

1 · What it is

Azure OpenAI is a way for companies to use OpenAI’s models without leaving Microsoft’s Azure cloud. OpenAI makes the models. Microsoft runs copies of them on its own computers and sells access as part of Azure. The bill arrives on the company’s normal Azure account, and Microsoft gives support if something breaks.

Why not just call OpenAI directly? Microsoft’s terms say your prompts and answers are not shown to other customers or to OpenAI. They are not used to train models unless you agree. The models here also do not talk to OpenAI’s own services. Microsoft does still store and process data to run the service and to watch for misuse.

Using it starts with a deployment. First a team creates a resource, which is like its own account space inside Azure. Then it deploys a model into it, which means switching on a copy the team can call. The deployment type decides two things: where the prompt is processed, and how you pay. You can pay per token (a token is a small chunk of text), reserve capacity, or send big jobs in a cheaper batch. For place, “Global” may use any Azure region (a region is a group of data centres in one part of the world), while “Data Zone” keeps processing inside an area such as the EU. Microsoft suggests Global Standard as the starting point for most work.

Here is an everyday example. A school’s help desk wants a bot that answers questions about timetables. Its developer already wrote code for OpenAI. With the newer v1 API (the set of rules programs use to talk to the service), they keep the same OpenAI code library and change the web address to their own Azure resource, ending in /openai/v1. When a student types a question, a content filter checks it first. The filter uses classifier models, which sort text into categories. It looks for hate, sexual content, violence and self-harm, at four levels from safe to high. Then the model writes an answer, and the filter checks that too before the student sees it.

There are limits. Filters are automatic classifier models. They also skip embedding models and Whisper audio models. Each Azure subscription has quotas that cap how much it can use. And not every model is offered in every region.

2 · Why it exists

Companies want OpenAI's models, but on their own cloud terms.

Data worriesA company may not want its prompts shared with other customers or used to train models.
Harmful outputA model can produce hateful, sexual, violent or self-harm content unless something checks it.
Where data goesSome teams must control which part of the world processes their data.
3 · How it works

Follow one request through Azure OpenAI.

Azure OpenAI request flowAn app sends a prompt to its own Azure endpoint. A content filter checks the prompt. The highlighted model deployment, hosted by Microsoft inside Azure, writes the completion. A content filter checks the completion before the answer returns. The model does not talk to OpenAI's own services. YOUR APPINSIDE AZURERESULT PromptPrompt filterModelAnswer filterAnswerDeployment type: where it runs and how you pay to your own/openai/v1 addresshate · sexualviolence · self-harmOpenAI model runby Microsoftsame fourcategoriesback tothe appGlobal · Data Zone · per token · reserved · batch no calls to OpenAI's own services
  1. 1 · deployA team creates a resource in Azure and deploys a model, choosing a deployment type.
  2. 2 · sendThe app sends a prompt to the resource's own web address, and it can use OpenAI's own client library.
  3. 3 · filterA content filter checks the prompt for harmful content.
  4. 4 · generateThe model, hosted by Microsoft inside Azure, writes the completion.
  5. 5 · checkThe filter checks the completion too before the answer returns.

Same models, different host: Microsoft runs them inside Azure, not OpenAI.

4 · Where it's used
WhoWhat they askWhat it works with
Hospital IT team“Can we summarise notes without our text reaching OpenAI or other customers?”The data, privacy and security terms for models sold by Azure
Developer with OpenAI code“How few lines must change to point our app at Azure?”The v1 API and the OpenAI client library
European bank“Can prompts be processed only inside the EU?”A Data Zone deployment type
Support chatbot team“Why did the service block that reply?”Content filter categories and severity levels
5 · What it solves, and what it doesn't
solves
  • It runs OpenAI models inside Microsoft's Azure cloud.
  • It keeps prompts and answers away from other customers and from OpenAI.
  • It adds content filters on both the prompt and the answer.
  • It lets teams choose where prompts are processed and how they pay.
doesn't solve
  • Filters are automatic classifier models.
  • The filter does not check embedding or Whisper audio models.
  • Throughput is capped by quotas on your Azure subscription.
  • Model choice varies by region, so not every model is everywhere.
6 · Go deeper
Microsoft FoundryChatGPTLLMAmazon BedrockVertex AI

Sources used

This explainer is written in original language. The links below support its factual claims.

  1. officialAzure OpenAI in Foundry Models, Microsoft Azure · read 28 Sept 2026
  2. docsFoundry Models sold by Azure, Microsoft Learn · read 28 Sept 2026
  3. docsData, privacy, and security for Models sold by Azure in Microsoft Foundry, Microsoft Learn · read 28 Sept 2026
  4. docsContent filtering for Microsoft Foundry Models (classic), Microsoft Learn · read 28 Sept 2026
  5. docsDeployment types for Microsoft Foundry Models, Microsoft Learn · read 28 Sept 2026
  6. docsAzure OpenAI in Microsoft Foundry Models v1 API, Microsoft Learn · read 28 Sept 2026
  7. docsAzure OpenAI quotas and limits, Microsoft Learn · read 28 Sept 2026
Suggest a correction