AWS adds usage-based metering for SageMaker bidirectional streaming

AWS announced usage-based metering for bidirectional streaming inference on Amazon SageMaker AI, so model providers can report actual consumption during a live session to AWS Marketplace for the software charge. The earlier metering method relied on a unit count sent in an HTTP response header, which does not fit sessions whose total is unknown until they end. Providers opt in through a response header, and SageMaker AI opens a companion WebSocket where the model server sends JSON metering records as usage accrues. Endpoint infrastructure charges still apply separately. Deepgram is the launch partner, using the capability to bill speech-to-text and text-to-speech on audio processed. It is available to model providers listing on AWS Marketplace and to the customers who use those listings.