Amazon launches Bedrock

Amazon announced Bedrock, a managed service giving AWS customers API access to a menu of foundation models rather than a single model built by Amazon. At launch, Bedrock offered AI21 Labs’ Jurassic-2 language models, Anthropic’s Claude, and Stability AI’s image-generation models including Stable Diffusion, accessible through one interface with AWS handling hosting, scaling and integration into a customer’s existing cloud infrastructure. Amazon said its own Titan family — a text model and a text-embeddings model — would join the service after a limited preview.

The announcement came alongside other generative-AI tooling for AWS customers, including EC2 instances built around AWS’s own Trainium and Inferentia chips for training and running models more cheaply than on general-purpose GPUs, and CodeWhisperer, AWS’s code-completion assistant.

Bedrock’s design choice was the notable part: where Microsoft had tied its cloud strategy tightly to a single partner, OpenAI, and Google was building and pushing its own PaLM models, Amazon positioned AWS as neutral ground — a marketplace offering several labs’ models rather than betting the business on one house model or one external partner. That let a customer already running infrastructure on AWS switch between models, or use several for different tasks, without leaving the platform. It also gave AI21, Anthropic and Stability distribution through the largest cloud provider by market share, at a point when each still depended heavily on external compute and enterprise reach that a hyperscaler could supply.

Bedrock became one of the clearer test cases for whether cloud infrastructure or model ownership would capture more of the value in generative AI, a question that stayed open as the “picks and shovels” argument for infrastructure providers competed with the pricing power frontier labs continued to command through the following years.