AWS showed how Amazon Bedrock analyzes video in three modes and calculates cost

AWS showed how to build scalable video analysis on Amazon Bedrock without a separate computer vision team. The company described three architectural approaches — frame-based, shot-based, and through multimodal embeddings — and immediately linked them to accuracy, latency, and cost.

Why Video is Still Difficult

Video has long become a standard format for surveillance cameras, media production, social networks, and corporate communications, but extracting useful signals from it remains challenging. Manual review doesn't scale well, and classical rule-based systems only see pre-defined patterns. Even when data is already collected, quickly understanding what is happening in a long video is still difficult. At large volumes this quickly becomes an expensive and slow operation.

AWS is betting on multimodal foundation models in Amazon Bedrock. Such models process visual and text data together: they can describe scenes in natural language, answer questions about video content, and notice subtle events that are difficult to formalize with ordinary rules. The point of this approach is that video analytics can now be assembled like a constructor from ready-made services, rather than as a separate research project with a large ML team.

Three Analysis Modes

The first option is frame-based workflow. The system takes frames at fixed intervals, removes similar and duplicate images, and then sends the remainder to the model for image understanding, while audio is transcribed separately through Amazon Transcribe. To filter out unnecessary frames, AWS provides two modes: Nova Multimodal Embeddings with 256-dimensional vectors and semantic similarity, or OpenCV ORB without additional calls to Bedrock. The first understands the meaning of a scene more accurately, the second is faster and cheaper. This mode is suitable for cameras, process control, and compliance verification.

The second option cuts video not into individual frames, but into short clips or segments equal in length. This is shot-based workflow: it preserves temporal context within a fragment and is better suited for media content, library cataloging, and highlight search. Segments can be built along natural scene boundaries using PySceneDetect or simply divide the video into equal intervals, for example 10 seconds. The first method is better for films, presentations, and vlogs, the second is for surveillance, sports, and live broadcasts.

• Frame-based mode — for precise monitoring and searching for specific events in time.
• Shot-based mode — for scenes, chapters, and long videos where context within a fragment matters.
• Embedding mode — for semantic search by queries like text or reference image.

The third option AWS calls multimodal embedding. In it, video is transformed into representations suitable for search: you can find fragments by text query, by similar image, and even do cross-modal search between different data types. In this architecture, Amazon Nova Multimodal Embedding and TwelveLabs Marengo are supported, and a unified interface allows changing the model for the task without complete pipeline reassembly. This is especially useful for archives with thousands of hours of content.

Infrastructure and Price

The entire system is built on AWS serverless services. Step Functions orchestrates frame-based and shot-based scenarios, Lambda performs processing, S3 stores raw results and artifacts, DynamoDB stores structured metadata for queries by video, timecode, and analysis type. For integration, a programmatic API is provided, and for the interface — a React application through CloudFront with authentication through Amazon Cognito.

Services for Nova, TwelveLabs, and recommendations through Bedrock Agents are separated. The practical focus of the article is not only on analysis quality, but also on cost control. AWS has built in token usage tracking and cost estimation for each processed video, including breakdown by Bedrock models and transcription through Transcribe.

This is important because different scenarios have radically different tradeoffs: in some places maximum accuracy is needed, in others minimum latency, and in others price on large volumes is more important. As a starting point, AWS also released the solution as an open source CDK package and included examples for cameras, chapter analysis, and moderation of user-generated content.

What It Means

AWS is essentially offering not one "magical" model for video, but a set of clear templates for different tasks. For business this is a good signal: video understanding is gradually transforming from expensive custom development into engineering assembly, where you can pre-select the right balance between quality, response speed, and budget.