Build1 publisher3 min readPublished
AWS puts Gemma 4 on Bedrock inside its EU-only Brandenburg region
Amazon says the Apache 2.0 model family is generally available on Bedrock's next-generation inference engine in eusc-de-east-1, reached through an OpenAI-compatible endpoint and run by staff who reside in the EU.
The Engineer · Build desk
What happened
- AWS has made Gemma 4 generally available on the Amazon Bedrock next-generation inference engine in the AWS European Sovereign Cloud, the first open weight model family offered there.
- The sovereign cloud became generally available in January 2026 with its first Region in Brandenburg, Germany, and all inference for these models stays inside eusc-de-east-1.
- Day-to-day operations, including data centre access, technical support and customer service, are controlled only by AWS employees who reside in the EU.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision EU teams that bought GPUs because managed inference sat outside their residency boundary now have a managed option to price against. The Apache 2.0 license means the same weights can run on both sides of the comparison.
- constraint The zero-retention default covers stateless calls only; a team that turns on stateful conversation management has to document the Bedrock project as the place its conversation data sits.
- exposure Procurement language that requires operators to hold EU citizenship is not met by today's controls, because AWS describes current staffing as a blended team during a transition.
- capability Because the endpoint speaks the OpenAI APIs, a sovereignty decision made this quarter can be unwound with the same edit that made it.
Two defaults in the announcement need reading together. Amazon says Bedrock uses a zero data retention model, so by default it does not store inputs or outputs [7]. It also says the Responses API supports stateful conversation management, which rebuilds context without you passing conversation history with each request [4]. Rebuilding context requires having kept it. Turn on stateful conversations and an auditor will ask about the project: stored responses are scoped by Amazon Bedrock project, described as a logical boundary that represents a workload for access control, cost tracking and usage monitoring [5].
That server-side state is also the part of a migration that does not travel. The compatibility claim is about code. The bedrock-mantle endpoint supports the OpenAI Responses and Chat Completions APIs, and AWS says an existing OpenAI SDK codebase moves by changing only the base URL and the API key [3]. Conversations the engine rebuilt for you live in a Bedrock project [5].
The security posture is stated in absolutes with one carve-out. AWS says no service operators can access model input or output during inference, and that prompts and responses are encrypted in transit and by default are not shared with the model provider [6][16]. Limited retention might apply for certain models for abuse detection [8]. The post does not say whether Gemma 4 is one of them.
On residency the wording is specific. All inference stays within eusc-de-east-1 [9], and customer-created metadata including roles, permissions, resource labels and configurations stays in the EU [11]. AWS describes the cloud as having no critical dependencies on non-EU personnel or infrastructure [12]. It says only AWS employees who reside in the EU control day-to-day operations, including access to data centres, technical support and customer service [13].
Residence and citizenship are different tests. AWS says it is gradually transitioning the sovereign cloud to be operated exclusively by EU citizens located in the EU [14]. Until that is done, it says, it will continue to work with a blended team of EU residents and EU citizens located in the EU [14].
Whether this displaces a self-managed fleet depends on conditions you can check. Gemma 4 has to be the model you want, because it is the only open weight family the post names as generally available [1][17]. Your capacity and failover plan has to survive one region, since Brandenburg is the first region and the only one named [10]. Your compliance argument has to accept controls the operator asserts and audits itself. Access follows a least privilege model, and access to systems that store or process customer data or metadata is logged, monitored for anomalies and audited [15]. Gemma 4 is released under Apache 2.0 [2], so the same weights run on hardware you own.
What to watch
- Whether a second sovereign Region appears. A second Region would allow failover inside the same jurisdictional boundary.
- Whether the Bedrock abuse detection documentation lists Gemma 4 among the models with limited retention.
- When the transition to operation exclusively by EU citizens located in the EU is declared complete.