Teams with data residency obligations often pick Amazon Bedrock because AWS has regions in their country. Then a developer copies a model ID from a sample, the ID starts with global., and prompts from a Sydney workload get processed in Ohio or Frankfurt. The call still goes to the Sydney endpoint and CloudTrail still logs it in Sydney, so nobody notices.
Where a Bedrock request is processed depends on the model ID you pass, not the endpoint you call. See also Do you pick the best Large Language Model for your project?
| Model ID | Where the request can run | Meets a residency requirement? |
|---|---|---|
No prefix, such as anthropic.claude-... | The region you call (in-region) | ✅ Yes |
Geographic prefix: au., eu., us., jp. | The regions listed in that geographic profile | ✅ Yes, if every listed region is acceptable |
global. | Any commercial region that the global profile supports | ❌ No |
Recent Claude models have no in-region ID in Australia, so the choice there is between au. and global.. AWS prices global profiles about 10% lower, so expect someone to suggest switching.
Northwind's claims app calls Bedrock in Sydney with modelId: "global.anthropic.claude-sonnet-5-5".
❌ Figure: Bad example - Bad example - The call is made in Sydney, but AWS can process it in any commercial region
Northwind's claims app calls Bedrock in Sydney with modelId: "au.anthropic.claude-sonnet-5-5", and an SCP denies global profiles across the organization.
✅ Figure: Good example - Good example - Requests run in Sydney or Melbourne, and a global model ID fails instead of going offshore
Code review will miss some model IDs, so put the rule in a service control policy.
Geographic profiles need every destination region allowed. An au. profile called from Sydney routes between Sydney (ap-southeast-2) and Melbourne (ap-southeast-4). If your region SCP allows only Sydney, the request is denied. Allow both regions, or add an exception on the bedrock:InferenceProfileArn condition key for the profiles you approve.
Global profiles are evaluated with the region set to unspecified. A region SCP that denies everything outside your allow list already blocks them. That protection disappears the day someone adds unspecified to make a global model work, so add the explicit deny that AWS documents:
{"Sid": "DenyBedrockGlobalInference","Effect": "Deny","Action": "bedrock:*","Resource": "*","Condition": {"StringEquals": {"aws:RequestedRegion": "unspecified"},"ArnLike": {"bedrock:InferenceProfileArn": "arn:aws:bedrock:*:*:inference-profile/global.*"}}}
Do not rely on the prefix alone, because the destination list can differ by source region. The au. profile for Claude Sonnet 4.5 stays in Sydney and Melbourne when called from Sydney, but it can also run in New Zealand when called from there.
Print the destinations for each profile you plan to use. The region in each model ARN is a place your request can run:
aws bedrock get-inference-profile \--region ap-southeast-2 \--inference-profile-identifier au.anthropic.claude-sonnet-5-5 \--query "models[].modelArn"
Run it from every source region you call Bedrock from, and keep the output as evidence. To see where requests actually ran, query CloudTrail: additionalEventData.inferenceRegion records the processing region.
New models often launch with us. and global. profiles only, and get an au. profile months later, if at all. Check the Regional availability table on the model card before you promise a model to a client, and see Do you plan for AI model quotas and region availability?
If the model is not available in your geography, use one that is, or get it from a provider that hosts it in your country. Offshore processing needs the data owner's sign-off, recorded in your risk register.
Prompts can still end up elsewhere:
bedrock-mantle endpoint: some models are served from this second endpoint, with its own IAM actions and VPC endpoint, and no invocation logs. Test your SCPs and IAM policies on both endpointsbedrock:InvokeModel and bedrock:InvokeModelWithResponseStream only on approved model and profile ARNs. See Do you give least-privilege access to your AI resources?