Amazon (NASDAQ: AMZN) has expanded Amazon Bedrock support for OpenAI GPT-5.6 Terra and Luna in AWS GovCloud (US), a change that gives regulated organizations a more direct route to deploy the models while retaining cloud-native operational controls. The update pairs model access with cross-region inference inside GovCloud, targeting a capacity and resiliency problem that can complicate production AI work in sovereign environments.
AWS said the models now run on the Bedrock Runtime endpoint through the Responses, Converse and Chat Completions APIs. That matters because teams can use OpenAI-compatible interfaces or AWS’s Converse interface while keeping model calls inside the same Bedrock service layer they use for other supported models. For a public-sector program or contractor subject to U.S. data-residency requirements, the choice of API is not merely a developer preference; it affects how an application is integrated, monitored and governed.
The new cross-region inference option automatically routes requests across both AWS GovCloud (US) Regions. AWS said the service is designed to provide higher throughput and additional redundancy without requiring customers to manage capacity across the two regions themselves. The routing remains within GovCloud, which AWS describes as preserving U.S. data sovereignty while broadening access to Bedrock capacity during demand spikes.
That operational model is important for AI applications that cannot simply fail over to a general commercial-cloud region when demand rises. A document-processing workflow, service-desk assistant or mission-support application may face sharp bursts of traffic yet still need to follow a prescribed regional boundary. Cross-region inference does not remove the need for an organization to establish its own authorization, data-handling and model-use controls, but it can reduce the capacity-planning work required to keep a governed application responsive.
AWS also said the native API support connects these model calls to existing account-level controls in Bedrock. Customers can record invocations in Amazon S3 or Amazon CloudWatch Logs, view invocation counts, token counts, latency, throttles and errors in Amazon CloudWatch, and allocate spend through AWS Cost Explorer and the Cost and Usage Report. Those details turn model access into an operating service that finance, security and platform teams can observe alongside other cloud consumption. They can also give security and FinOps teams a shared record for reviewing unusual model use or rapidly rising token demand.
The update identifies GPT-5.6 Terra and Luna as the supported OpenAI models. AWS directs customers to Bedrock model cards for the model-specific information, including supported regions, endpoints, APIs, features, inference profiles and pricing. Enterprises should still validate those details against their selected deployment pattern before moving a workload into production, particularly where throughput, logging destinations or procurement rules are constrained.
The strategic shift is less about adding another model selector than about making a frontier-model integration fit established sovereign-cloud operations. Earlier enterprise AI deployments often required teams to trade capacity flexibility against locality and familiar observability. By tying OpenAI model access to GovCloud routing, logs, metrics and cost records, AWS is positioning the model layer to behave more like a governed cloud service than a separate, opaque endpoint.