Amazon Bedrock explained: One AWS account for many AI models, with the controls attached

Amazon Bedrock is a managed service from Amazon Web Services (AWS) that lets a company use AI models from several labs through one account and one programming interface. AWS runs the servers, the billing and the access controls, while the model itself comes from Amazon, Anthropic, OpenAI or others. Think of it as a model food court inside a cloud provider.

Renting a model the way a company rents a database

Imagine a bank that wants a chatbot to summarize customer emails. It could train its own model, which costs a fortune. It could sign a contract with an AI lab and send data to that lab’s servers, which means a new vendor review. Or it could ask its cloud provider for a model, the way it already asks for a database.

Bedrock is that third option. The bank’s developers send text to an endpoint and get text back. They never see a graphics processing unit (GPU) or install a model. AWS describes Bedrock in its documentation as a fully managed service that gives access to foundation models from leading AI companies. A foundation model is a large, general-purpose model that can be adapted to many tasks.

One interface, many models, no servers to run

In practice, a developer picks a model by its identifier and sends a request. AWS documents several ways to do this, including a Converse operation that uses the same message format for different models, and compatibility endpoints for the OpenAI Responses and Chat Completions styles. Switching from one model to another can mean changing one line.

That matters because the best model changes every few months. A company that wired its software to one lab’s interface has to rewrite code to move. With a shared interface, it can test a cheaper or newer model against the same prompts and compare.

Bedrock is also serverless, meaning the customer does not provision machines. AWS runs the model software on its own hardware and charges for use, usually by the amount of text processed. For steady heavy traffic, AWS also sells reserved capacity under the name Provisioned Throughput. Other pieces sit on top: Guardrails for filtering harmful content, Knowledge Bases for connecting company documents, and AgentCore for building AI agents. AWS describes these as parts of the platform.

From a limited preview in 2023 to a catalog of 100+ models

AWS announced Bedrock on April 13, 2023, as a limited preview with models from AI21 Labs, Anthropic, Stability AI and Amazon’s own Titan family. Swami Sivasubramanian, then AWS vice president for data and machine learning, introduced it as part of a set of generative AI tools. The service became generally available on September 28, 2023, according to the AWS News Blog, initially in three regions.

The catalog has grown since. AWS documentation now lists more than 100 foundation models from providers including Amazon, Anthropic, DeepSeek, Moonshot AI, MiniMax, OpenAI and xAI. The documentation’s what’s-new list shows OpenAI’s GPT-5.6 Sol, Terra and Luna reaching general availability on Bedrock in July 2026, and GPT-6 Astra following in September.

Bedrock is not alone. Microsoft sells a similar service through Azure, and Google Cloud has its own platform. Each cloud provider wants the models customers choose to run next to the data customers already store there.

The governance pitch: Your data stays in your cloud account

For large companies, the main selling point is control rather than model quality. Several claims come from AWS itself and should be read as vendor statements.

• No training on your data. AWS says Bedrock does not store or use customer data to train models, and its launch post said prompts and completions are not shared with model providers.
• Model providers kept out. AWS documentation describes a Model Deployment Account in each region, owned and operated by the Bedrock team. Providers have no access to it, so they cannot see prompts or logs.
• Existing identity tools. Access is managed with AWS Identity and Access Management (IAM), the same permission system used for the rest of an AWS account.
• Private networking: AWS PrivateLink lets traffic reach Bedrock without crossing the public internet, according to AWS.
• Audit trails: AWS CloudTrail logs API activity, which helps compliance teams.

Region choice is another control. Bedrock offers cross-Region inference, which routes a request to another AWS region for more capacity. AWS documents a geographic option that keeps processing inside a boundary such as the US or EU, and a global option it says costs roughly 10% less but can run anywhere supported. Companies with data residency rules must choose carefully.

Where the story connects: From Bedrock to ChatGPT Enterprise on AWS bills

Bedrock is for developers building applications and agents. Our report on ChatGPT Enterprise becoming available through AWS Marketplace covers a different product. OpenAI’s announcement of November 9 says Marketplace buyers get the finished workplace product for staff, while developers are pointed to Bedrock.

Both routes share a billing idea. Amazon’s page says Bedrock pricing for the GPT-5.6 family matches OpenAI’s first-party rates and counts toward existing AWS spending commitments. The Marketplace listing extends that to a ready-made employee tool. The Bedrock controls listed above belong to Bedrock, though, and the Marketplace announcement does not say they apply to ChatGPT Enterprise.

Limits to weigh before committing

Bedrock does not remove every problem. A few trade-offs stand out:

• New models can arrive later. A lab’s own service may get a model or feature first. Bedrock availability varies by model and region, and some models are in limited availability.
• Lock-in moves, it does not vanish. A shared interface eases switching between models, but guardrails, knowledge bases and agent setups are AWS-specific.
• Shared responsibility. AWS states that customers remain responsible for their own configuration. A loose IAM policy is still the customer’s mistake.
• Routing trade-offs. Cheaper global routing and strict data residency pull in opposite directions.
• Vendor figures. Numbers such as how much harmful content Guardrails can block are AWS claims, not independent tests.

Two things are worth watching. One is whether AWS and OpenAI extend the same governance controls to the finished ChatGPT Enterprise product, which the announcement leaves open. The other is how fast new frontier models reach Bedrock after the labs’ own launches. A short gap makes the single-account approach easy to justify, and a long one pushes teams back to direct contracts.