AWS Details GPT-6 Cross-Region Inference From Five Asia-Pacific Regions
Bedrock can route GPT-6 Sol and Luna requests from five Asia-Pacific source regions, but global processing creates policy and data-location work for enterprise teams.
Edited by Tyronne Panaino
AWS documented on October 12 how Amazon Bedrock customers can invoke OpenAI GPT-6 models through global cross-Region inference from parts of Asia-Pacific. The AWS technical article says GPT-6 Sol and GPT-6 Luna can be called from AWS source regions in Singapore, Jakarta, Malaysia, Thailand and Taipei, while GPT-6 Astra is available through this path from Singapore.
The distinction matters to engineering, security and compliance teams because the source region remains the operational home of a request even when Bedrock chooses another supported commercial region for processing. AWS presents the profiles as a way to reach broader capacity when a single region cannot meet demand, but the same design means organizations must understand both where control records stay and where inference can occur.
Model access differs by source region
Sol and Luna have the broadest source-region coverage in the guidance: all five listed Asia-Pacific regions can submit requests through their global profiles. Astra has a narrower starting point and is documented only from Singapore. That makes model selection part of regional architecture rather than a simple swap between three otherwise identical endpoints.
AWS says Singapore is enabled by default. Jakarta, Malaysia, Thailand and Taipei are opt-in regions that an account must activate before use. A deployment plan therefore needs to check both the desired model and whether the source region is enabled for the account. The source also warns that profile availability and region support can change, so the current model card remains a necessary pre-deployment check.
The source region keeps the operational records
A global profile replaces a direct model identifier when a request is sent to Bedrock. AWS then selects a supported destination region for processing. Billing, quota accounting, CloudWatch metrics and CloudTrail events stay with the source region, giving operators one place to monitor calls even when inference is routed elsewhere.
The same Bedrock Runtime endpoint supports the Converse API, OpenAI Responses API and OpenAI Chat Completions API for these profiles. That common surface reduces the number of separate integrations an engineering team must operate, although the fetched evidence is a first-party configuration guide rather than an independent compatibility test.
CloudTrail provides the important link between the two sides of the route. AWS says each successful request can record the processing location in the `additionalEventData.inferenceRegion` field while the event itself remains available from the source region. Teams that need an audit trail can use that record to check where individual calls were processed instead of assuming that the source region was also the destination.
Enterprise region controls need a scoped exception
The routing model can conflict with organization-wide service control policies. AWS says global profile requests set `aws:RequestedRegion` to `unspecified`, so a policy that allows only named regions can deny a call even when the source region is approved. The article recommends scoping an exception to the Bedrock inference profile instead of broadly weakening the region allowlist.
That is a practical security boundary, not a minor deployment detail. A broad exception could affect services beyond the intended model path, while an inference-profile condition can keep the permission tied to the approved Bedrock resource. Organizations still need to test their own identity policies, service control policies and streaming permissions; the article does not establish that an example policy will fit every account structure.
Global routing does not equal local data residency
AWS states that classifier-flagged traffic retained for automated abuse detection is processed in the destination region. For a global profile, that destination can be any supported commercial AWS region where the model is deployed. The operational logs may remain at home, but the inference and associated retention path can cross the geographic boundary represented by the source region.
This is the main limitation for regulated or residency-sensitive workloads. The article advises organizations with data-location requirements to involve risk and compliance teams early. It does not independently verify availability, compare latency with single-region inference, publish outage data or prove that a particular regulatory obligation is satisfied.
The next useful checkpoints are measured availability from each source region, evidence of how routing behaves during capacity pressure, and organization-specific validation of identity, logging and residency controls. Until those checks exist, the AWS article is best read as a detailed operating map for the service rather than a performance or compliance guarantee.
Status
Learning. AWS has published a detailed first-party guide to the profiles, interfaces and policy boundaries. Internal confidence is medium because availability and behavior are described by the service provider and were not independently tested for this article.
Sources
Update note: Last reviewed 2026-10-13. We will revise this post if AWS changes the profile coverage or independent testing establishes the routing, reliability or data-location behavior.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.