AWS Marketplace now supports usage-based software metering for models running through Amazon SageMaker AI bidirectional streaming. The change lets providers charge for actual consumption during a live audio or data stream instead of relying only on hourly software pricing.
Ordinary request-response inference can place a billable unit count in the response header. A bidirectional session may last seconds or many minutes, so its final usage is not known when the stream begins. SageMaker now opens a companion WebSocket for metering records alongside the connection carrying inference data. Both share a request identifier, allowing providers to associate incremental charges with the active session.
Deepgram is the launch partner, using the feature for speech-to-text and text-to-speech products. Providers can report units as they accrue, with a limit of one metadata message per second and 512 bytes per record. Usage metering applies to the Marketplace software charge only; customers still pay SageMaker endpoint infrastructure costs while an endpoint is provisioned. The change aligns billing more closely with consumption but does not eliminate idle infrastructure expense.