Choose a deployment type
A deployment makes a model available to your project. Foundry offers several deployment types, and they differ in where your data is processed, how you pay, and how much throughput you get.
| Deployment type | Where data is processed | How you pay | Best for |
|---|---|---|---|
| Global Standard | Any Azure region | Pay per token | Most projects and experiments, with the highest default quota |
| Global Provisioned | Any Azure region | Reserved capacity (PTUs) | Steady, high-volume traffic that needs consistent latency |
| Standard | Within an Azure geography | Pay per token | Data that must stay in one geography |
| Data Zone Standard | Within a data zone, such as the US or the EU | Pay per token | Data that must stay in one zone |
| Global Batch | Any Azure region | Pay per token, with a 50% discount and up to 24-hour turnaround | Large jobs that can wait |
Start with Global Standard for experiments. Choose a provisioned type only when your traffic is steady and predictable.
Quota and token limits
Quota sets how many tokens per minute (TPM) a model can use in your subscription. Since May 2026, Foundry tracks quota for each subscription. Deployments of the same model and version share one quota pool across regions, so one team's deployment can use up quota that another team needs.
- Tokens per minute (TPM) caps how many tokens a deployment can process each minute. Going over it returns a 429 Too Many Requests error.
- A total token quota caps usage over a period, such as a day or a month. Going over it returns a 403 Forbidden error.
- An administrator can set both limits for each project in AI Gateway: open Manage, then AI Gateway, then Token management. AI Gateway has to be set up first.
Use these limits to stop one project from taking more than its share of a shared budget.
API keys or Microsoft Entra ID
Foundry accepts two ways to prove who is calling. An API key is one secret for the whole resource. Microsoft Entra ID gives each user or app its own identity and access token.
- API keys are quick for a first prototype on your own machine. They cannot show which person made a call, and they are hard to limit.
- Microsoft Entra ID is the choice for production. You can limit access by role, and an app can use a managed identity with no secret in its code.
For keyless calls, give the user or app the Cognitive Services User role on the Foundry resource, then request a Microsoft Entra token for Foundry. Never paste a key into a shared file or a code repository. For Foundry project work, use the Foundry User role. The Azure AI Developer role is not meant for Foundry project work.
Keep costs under control
- Pay-per-token deployments bill for every token you send and receive, so test with short prompts first.
- Provisioned deployments bill for reserved capacity whether or not you use it.
- Instant access lets you try some models without creating a deployment. During preview, it works only in Foundry projects in West US 3.
Set a budget in Azure Cost Management. A budget sends alerts, but it does not stop spending. Cost data can take 8 to 24 hours to appear, so set your first alert low.
When a test ends, delete the resources you no longer need. Deleting a resource stops its usage charges, but storage, networking, reserved capacity, and support plans can keep billing. Check the cost view after you clean up.