Choosing 'Southeast Asia' doesn't mean your AI data is processed in Singapore

Create a Microsoft Foundry resource in Southeast Asia (Singapore) and it is natural to assume your data stays in Singapore. For stored data, that is broadly true. For the prompts and responses your model processes, it depends on a second setting that is easy to miss. Choose the widest option and a single request can be processed in any of 21 countries.

It is tempting to treat the region dropdown as the data residency decision. In Microsoft Foundry it is only half of it. This article explains the other half, using a resource in Southeast Asia (Singapore) as the example. Model availability changes often, so every figure here is as at October 2026.

Two settings, two different jobs

When you deploy a model in Foundry, two choices decide where your data goes. They are easy to confuse because both look like location settings.

The region of the Foundry resource decides where stored data sits. According to Microsoft's data, privacy and security page, data at rest stays in the Azure geography of the resource, whatever deployment type you choose. For a resource in Southeast Asia, that is the Asia Pacific geography.

The deployment type decides where prompts and responses are processed. It is easy to skip past. Microsoft's own docs call Global Standard a common starting point, and it is the widest of the three.

Three deployment types compared

Foundry offers three deployment types for a pay-per-use model, and each draws a different line around processing.

• Global: a request can be processed in any Azure region where the model is deployed. For GPT-5.4 mini today, that is 21 countries across Asia, the Americas, Europe, the Middle East and Africa.
• Data Zone (Asia Pacific): processing stays inside the Asia Pacific data zone. Today that means Singapore, Japan, South Korea, Australia and India.
• Standard (regional): processing stays in the region of the deployment.

Two details catch people out. First, Microsoft can add regions to a data zone without prior notice, so "Asia Pacific" is a boundary, not a fixed list. Second, the stored data rule does not change between types. Only the processing location moves.

Singapore-only has a price

The obvious answer is to pick Standard and keep everything in Singapore. As at October 2026, that option barely exists for language models. From a Southeast Asia resource, Microsoft lists 33 models as Global, 8 as Data Zone, and 1 as Standard. That one is an embedding model, which turns text into vectors for search. It cannot read a document or answer a question.

So the trade-off is direct: the tighter you draw the processing boundary, the fewer models you can use. Two popular models show how far apart the options can be. GPT-5.4 mini is offered from a Southeast Asia resource as Data Zone or Global. DeepSeek-V4-Flash is not offered from a Southeast Asia resource at all. Its Data Zone option exists only in the US and the EU, and its Global option needs a resource in another region, such as Japan East or Australia East.

That has a knock-on effect. Stored data follows the resource, so a DeepSeek resource in Japan East keeps its stored data in the Japan geography, not Asia Pacific. The model you pick can move your stored data, not just your processing.

If you need a chat model processed strictly in Singapore, the route today is reserved capacity (Regional Provisioned) for GPT-5.4 mini, which is billed by the hour with a minimum size per model. That is a production budget decision, not a pilot setting.

The model lists move every month. Check the region availability page for your model before you promise anything to a security team.

What stateless does and doesn't cover

Microsoft describes the models as stateless: nothing is kept inside the model, and prompts and outputs are not used to train it. Your prompts are also not shared with OpenAI, DeepSeek or any other model provider. That settles the training question.

It does not mean nothing is stored anywhere. Three things are worth knowing:

1. Abuse monitoring. If automated checks flag a request as possibly harmful, a sample can be stored for review by authorised Microsoft staff, in the geography of your resource. Customers managed by a Microsoft account team can apply to switch this storage off.
2. Responses API history. Responses are stored for 30 days by default. Set store to false if you do not need conversation history.
3. Optional features. Stored completions, batch files, fine-tuning data and assistants all keep data in the resource. You can delete it at any time, but nothing removes it for you.

One more piece of fine print. Microsoft's privacy page allows even Standard deployments to be processed between regions within the same geography for operational reasons. Treat "regional" as a strong boundary, not a promise of a single region.

Check what you actually deployed

The deployment type is visible in the SKU name of each deployment: GlobalStandard, DataZoneStandard or Standard. One command shows all of them for a resource, which is quicker than clicking through the portal and easy to paste into a change ticket.

az cognitiveservices account deployment list \
  --name "{foundry-resource}" \
  --resource-group "{resource-group}" \
  --query "[].{deployment:name, model:properties.model.name, type:sku.name}" \
  --output table   # the "type" column is where processing can happen

Four checks before go-live

Run these before your next AI pilot reaches a security review.

1. Which region is the Foundry resource in?
2. What deployment type is each model deployment?
3. Is your chosen model offered from your region, and as which type?
4. Which features store data, where, and for how long?

None of these take long to answer. Having the answers ready before the security review saves a round of back-and-forth, and it shows the reviewers you have read the same fine print they have.