QDNAAdvisory and architecture for LLM inference and training platforms, on-premises or hybrid

Amazon Bedrock

External APIs

Amazon Bedrock

Amazon Bedrock aggregates models from several providers, including Anthropic, Meta, Mistral and Amazon, as SaaS behind a single managed API. Useful for a load spike or a model missing from the local catalogue. Data leaves your infrastructure, so this is reserved for non-sensitive workloads.

Key points: Managed · Multi-provider · Off-premises · Non-sensitive.

Models served: Claude by Anthropic, GPT-5.6 and gpt-oss by OpenAI, Llama by Meta, Mistral, Amazon Nova and Titan, DeepSeek, Qwen, GLM by Z.ai, Kimi by Moonshot AI, Gemma by Google, Grok by xAI, Cohere, AI21 and Stability AI (AWS catalogue as of 2 September 2026)

Location: United States, subject to the Cloud Act.

Where the data goes

The AWS documentation distinguishes two regimes that should not be conflated. In standard inference, data stays in the Region you call. With cross-Region inference profiles, it moves inside a geography, for example all European Regions, without leaving it; a global profile routes it to any commercial Region, with an announced saving of about 10 % (AWS documentation as of 2 September 2026). The second regime improves availability and blurs the exact location.

Retention and training

AWS places Bedrock under shared responsibility and documents a retention mode set per account or per project: "none" for zero data retention, "default" for the model's own policy, "aws_review" for human review by AWS. The model provider has no access to prompts or completions, but some models require the review mode: Claude Fable 5 and Claude Fable 5.1 impose retention of up to thirty days within the AWS boundary (AWS documentation as of 2 September 2026). The question to ask is therefore not what Bedrock does, but which model, in which Region, under whose upstream terms.

When to use it from a sovereign platform

When the company already runs on AWS and billing, authentication and private networking exist. The benefit is integration, not sovereignty: the choice of a cross-Region inference profile must be explicit, because it changes the answer to the location question.

Official sources

Talk to a specialist

A no-obligation conversation.

Book a call