Cross-Region inference is Bedrock's answer to a common scaling problem: your workload bursts past what one region will serve. It works through inference profiles — AWS-managed objects that pair a foundation model with a set of AWS regions requests may be routed to. Instead of invoking a model in exactly one region, you invoke the profile and let AWS pick a destination with available capacity. For bursty enterprise workloads, this raises effective throughput without any application-side load balancing.
Two flavors: geographic and global
AWS provides two types of managed cross-region profile:
| Type | Routing scope | Notes |
|---|---|---|
| Geographic | Regions within one geography (US, EU, APAC, etc.) | Data stays within the geography — the usual choice when residency matters |
| Global | Any supported commercial AWS region | Widest capacity pool; AWS cites roughly 10% cost savings versus geographic/standard pricing |
On the legacy Bedrock surface these appear directly in the model ID as prefixes: global.anthropic.claude-opus-4-6-v1 for global routing, or us./eu./jp./apac. variants for geographic routing. The newer "Claude in Amazon Bedrock" surface expresses the same idea as Global versus Regional endpoints, with geographic profiles available for US, EU, JP, and AU — and a 10% pricing premium for single-region (regional) endpoints over global ones on Claude Sonnet 4.5 and later models. Either way the economics push the same direction: letting AWS route widely is the cheaper default; pinning a region is what costs extra.
What actually happens to your request
Routing is AWS-internal. Cross-region traffic stays on the AWS network — it never crosses the public internet — and is encrypted in transit. Requests can be served in regions you never manually enabled in your account. Pricing is based on the region from which you call the inference profile, with no additional routing charge. And every cross-region request is logged by CloudTrail in the source region, with the additionalEventData.inferenceRegion field recording where inference actually ran — useful evidence when auditors ask where data was processed.
IAM and SCP requirements
Two policy details routinely trip teams up. First, invoking through a cross-region inference profile requires allowing bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream on both resource types: the foundation-model ARN (arn:aws:bedrock:*::foundation-model/...) and the inference-profile ARN (arn:aws:bedrock:{region}:{account-id}:inference-profile/*). A policy scoped only to foundation models fails once you switch to a profile.
Second, Service Control Policies that restrict regions must accommodate routing: geographic profiles need all destination regions allowed, and global profiles require "aws:RequestedRegion": "unspecified" to be permitted. A region-pinning SCP written before your Claude rollout is a common silent cause of throttled-looking failures. See SCP controls for Bedrock.
Limits and quota interaction
Inference profiles do not support Provisioned Throughput — they are an on-demand mechanism. On the quota side, the legacy bedrock-runtime endpoint tracks a distinct "Cross-Region InvokeModel tokens per minute" quota per model, separate from single-region on-demand quotas, and the bedrock-mantle endpoint tracks its own input/output token-per-minute quotas entirely. If you're hitting limits, identify which endpoint and which quota name before filing an increase — the processes differ (Service Quotas console for bedrock-runtime; an AWS Support case for bedrock-mantle). More in requesting Bedrock quota increases.
Where to go next
If you need routing you control — specific regions, separate cost tracking — read application inference profiles. For the pricing angle, see the platform overview and Bedrock billing mechanics.