Amazon Bedrock now offers the open‑weight Gemma 4 model family in the AWS European Sovereign Cloud (ESC), and the service runs on the next‑generation inference engine reachable via the bedrock-mantle endpoint. This change lets teams build generative‑AI workloads that stay inside the EU, meet strict digital‑sovereignty requirements, and reuse existing OpenAI‑compatible client code.
New open‑weight model availability in the European Sovereign Cloud
Gemma 4, released under the Apache 2.0 license, is the first open‑weight model family generally available on Bedrock within the ESC. The model is served by the Bedrock next‑generation inference engine, which exposes OpenAI‑compatible Responses and Chat Completions APIs at the bedrock-mantle endpoint. Existing OpenAI SDKs can be pointed at Bedrock by swapping the base URL and API key, eliminating the need for code rewrites. The Responses API also provides built‑in stateful conversation handling, storing context scoped to a Bedrock project – a logical container that isolates access, cost tracking, and usage metrics.
Data residency and security model
All inference traffic for Gemma 4 in the ESC is confined to the eusc-de-east-1 region in Brandenburg, Germany. The ESC operates exclusively with EU‑resident staff, and no data leaves the region unless a customer explicitly chooses otherwise. Bedrock enforces a zero‑operator‑access model: service operators cannot view model inputs or outputs during inference. By default, the service does not retain prompts or responses, applying a zero‑retention policy unless limited retention is required for abuse‑detection purposes (as documented by Bedrock). All data in transit is encrypted, and responses are not shared with the model provider unless the customer opts in.
Operational considerations
The inference engine follows a least‑privilege access pattern. Any operator interaction with systems that handle customer data is logged, monitored for anomalies, and audited. Bedrock projects serve as the primary boundary for access control, enabling teams to assign permissions per workload and to track usage for cost and compliance reporting. Because the engine is distributed and built for high performance, practitioners should treat the bedrock-mantle endpoint as a regional service endpoint and design their networking and IAM policies accordingly.
Related CloudNinjas coverage: AWS.
What This Means For Practitioners
- When deploying generative‑AI services in Europe, select the
eusc-de-east-1region and thebedrock-mantleendpoint to guarantee data residency. - Leverage existing OpenAI client libraries by updating the endpoint URL and API key, reducing integration effort.
- Use Bedrock projects to isolate workloads, enforce least‑privilege access, and obtain granular usage metrics for cost and compliance tracking.
- Validate that any required data retention for abuse detection aligns with organizational policies, since Bedrock may retain limited data for that purpose.
- Monitor operator logs and audit trails provided by Bedrock to ensure the zero‑operator‑access claim remains enforced in practice.
