Amazon Bedrock: Models, Governance, and Buyer Questions

Amazon Bedrock

Amazon Bedrock is a managed Amazon Web Services platform for accessing and operating foundation models from Amazon and third-party providers. It combines model APIs with AWS identity, security, governance, evaluation, knowledge, agent, and operational services.

Bedrock is a distribution layer as well as a technical platform. It can let an existing AWS customer procure and operate several model families through an approved cloud path.

Product surface

The current Amazon Bedrock overview describes more than 100 foundation models and compatibility with several API styles. The listed providers include Amazon, Anthropic, DeepSeek, Moonshot AI, MiniMax, OpenAI, and others.

The exact catalog varies by region, access mode, model provider, and date. "Available on Bedrock" should always name the model version and region relevant to the claim.

flowchart LR
    A["Application"] --> B["AWS identity and network controls"]
    B --> C["Amazon Bedrock API and operations"]
    C --> D["Selected foundation model"]
    C --> E["Guardrails, knowledge, agents, evaluation, and monitoring"]
    D --> F["Application output"]
    E --> F

Model choice and lifecycle

Bedrock separates the platform from any one model family. That can help a customer compare or route among models without leaving the AWS operating environment.

It does not make every model interchangeable. APIs, tool behavior, context, performance, price, regions, safety behavior, and license terms can differ.

AWS publishes a model lifecycle policy with Active, Legacy, public extended access, and End-of-Life states. The page warns that Bedrock lifecycle dates may differ from provider dates. Buyers need an upgrade plan rather than assuming a model endpoint is permanent.

Security and privacy

AWS describes encryption, KMS key control, identity policies, and PrivateLink connectivity in its Bedrock security material. Its current product page says customer inputs and outputs are not shared with model providers or used to train base models.

These are AWS statements about the service. A buyer should review the current service terms, documentation, architecture, region, logging, abuse detection, retention exceptions, subprocessors, and the specific model provider agreement.

Model access through a cloud does not transfer responsibility for application permissions, retrieval quality, harmful outputs, human oversight, or workflow controls.

Marketplace and ecosystem

The Amazon Bedrock Marketplace extends discovery and deployment to additional specialized models. AWS says marketplace models can be invoked through Bedrock APIs and used with tools such as Agents, Knowledge Bases, and Guardrails.

This illustrates how a cloud channel can compound. The customer keeps identity, billing, networking, operational tooling, and procurement in one environment while the model catalog changes.

The same concentration can create switching burden. Applications may depend on AWS-specific identity, agents, data services, monitoring, and contracts even when a model itself is portable.

Anthropic relationship

Anthropic's April 2026 announcement says more than 100,000 customers run Claude on Bedrock and describes expanded capacity using AWS Trainium. AWS maintains current Claude model cards for the versions offered through the service.

The relationship does not make Claude exclusive to AWS. Anthropic reports distribution through Google Cloud and Microsoft Azure as well.

Buyer questions

A buyer should identify the required model and region, data path, identity boundary, networking, supported API, logging, retention, guardrails, evaluation method, quotas, latency, pricing, lifecycle, support, portability, and exit plan.

The business decision is not only which model performs best. It is whether the full Bedrock system meets the task, risk, cost, and operating requirements.

This profile reflects current AWS and Anthropic sources reviewed on July 27, 2026. It does not independently validate vendor performance or customer counts. AI assistance was used for research organization, drafting, and validation. Publication remains unauthorized.