Why it matters
Regional deployments are often chosen by default or carried over from older setups, even when the workload has no requirement about where prompts are processed.How nOps finds it
nOps reads the last 7 complete days of pay-as-you-go Azure OpenAI token charges per account, model and deployment type, billed on at least 4 of those days. For each token type it compares the list price per million tokens of the current deployment type with Global Standard for the same model, version, context and token type, and applies that difference to what you paid, scaled to a month. Tokens covered by a commitment (provisioned throughput) are not counted, and deployments mostly on commitment are left out. The saving must be at least $5 a month, and accounts with 90% or more of their AI spend covered by credits are left out.Before you start
Confirm the workload has no data residency or processing-location requirement that rules out Global deployments.How to implement
- Confirm with the workload owner and your compliance team that prompts may be processed in any Azure region.
- In Azure AI Foundry (or the Azure OpenAI resource), create a new deployment of the same model and version with the Global Standard deployment type.
- Point the application at the new deployment name, ideally gradually.
- When no traffic reaches the old deployment, delete it.
- Update any Bicep, ARM or Terraform that defines the deployment so it is not recreated.
Risks and considerations
- Global deployments can process prompts in any Azure region. Stored data stays in the account’s geography, but this may not meet some data residency rules.
- Global Standard has its own quota, so check the account has enough before moving traffic.
- To roll back, point the application back at the Regional or Data Zone deployment.
FAQs
Does my data leave my region?
Does my data leave my region?
Prompts and responses may be processed in another Azure region. Data stored at rest stays in the geography of the account.
Related
- Move Low-Complexity Azure OpenAI Traffic to a Smaller Model
- Move Latency-Tolerant Azure OpenAI Workloads to Batch
- Supported recommendations