Notes · AI models

Azure OpenAI: where your data is really processed

“We use GPT on Azure, in the Swiss region.” It sounds reassuring. Whether it is true does not depend on the region alone. What decides it is a setting many people overlook: the deployment type.

Region and deployment are two different things

In Azure you first create a resource in a region, for example Switzerland North. Inside it you create a deployment for each model, and that is where you choose the type. Microsoft distinguishes:

• Global (Global Standard, Global Provisioned, Global Batch): prompts and responses may be processed in any Azure region where the model runs.
• Data Zone (for example Data Zone Standard): processing stays within a zone defined by Microsoft, namely the US, the EU or Asia Pacific. The EU zone follows Microsoft's EU Data Boundary. Besides the EU member states, it includes the EFTA states, so Switzerland too. Microsoft can add regions to a zone without prior notice.
• Standard and Regional Provisioned: processing stays within the Azure geography you chose. Within that geography Microsoft may move between regions for operational reasons.

Stored data, such as uploaded files, stays in your resource's geography for every type. So the type decides where the computing happens, not where data is kept. Both matter to you.

What this means for a Swiss company

Global means a request from Fribourg may be processed in a data centre in America or Asia. Microsoft recommends Global Standard as the starting point for most workloads, because new models appear there first and the price is lowest. For customer data it is rarely the right choice.

Data Zone EU means processing stays in the EU and EFTA area. For a lot of personal data that is enough, because the Federal Council recognises the data protection of the EU member states as adequate.

Standard in Switzerland North means processing stays in Switzerland. But new models arrive in a fixed order, global first, then data zone, regional last and with no guaranteed date. According to Microsoft's availability table on 7 November 2026, Switzerland North offered regional deployments of gpt-4.1, gpt-4o and gpt-5.1, among others. The newer models from gpt-5.4 onwards were only available there as Data Zone or Global. So if you want the newest model, you often have to go to the EU area. The table changes all the time. I check it again before every decision.

One point that often worries people: according to Microsoft, prompts and responses are not used to train the base models. For abuse monitoring, Microsoft may store flagged requests, and it does so in your resource's geography.

How to check it yourself

1. Look at the resource's region. The Azure portal shows it in the resource overview.
2. Look at the type of each deployment. The Foundry portal shows it on the deployment. In templates and scripts it is called the SKU, such as GlobalStandard, DataZoneStandard or Standard. If you see Global anywhere, processing may leave Europe. The same goes for DeveloperTier.
3. Lock it for the future. Azure Policy can stop anyone from creating a global deployment later. Microsoft documents a rule on the deployment type for this.
4. Check the table before every model change. Is the new model available in your region with your type? If not, decide deliberately, rather than someone switching the type “just for now”.

What I do

• We decide together which data needs which type. General text can use a different type than personal data.
• I set up Azure OpenAI in your own Azure subscription. Region and type are written in a template your IT can read and verify.
• An Azure Policy blocks global types where they are not allowed.
• Before a model change I check availability and tell you what changes for the data location.
• I write the entry for your record of processing activities with you.