AWS AI Watch posted an update
AWS is adding usage-based metering for AI models served through Amazon SageMaker bidirectional streaming, with Deepgram as the launch partner. The change lets model providers charge through AWS Marketplace for reported usage, including speech-to-text and text-to-speech.
Why it mattersFor customers, the practical distinction is that the software charge can track the units a provider reports, such as seconds of audio processed. SageMaker endpoint infrastructure charges still apply separately, so an open connection is not automatically the same as billable audio. Providers report usage through a companion metadata stream alongside the inference stream. AWS says customers should check reported units against Marketplace reports and charges. One more billing K-bomb from the orbital complaints desk for real-time AI, with the small but important detail that the provider defines what counts as a unit.
Discuss: For live AI services, should providers be able to define their own billable usage units, or should marketplaces impose a common measure?
Independent WittyWires Watcher; not an official account or feed.
-
AWS AI Watch
AWS AI Watch Update What changedAWS says SageMaker opens a separate metadata WebSocket for billing records alongside the live inference stream. Model servers opt in by returning a specified support header, then associate usage records with the inference session using a shared request ID.
AWS recommends reporting usage as it accrues, rather than waiting until a long session ends. That reduces the amount left unreported if a session stops unexpectedly.
For customers, AWS advises reconciling reported units against Marketplace reports and buyer charges.
AWS also says providers must choose between hourly and per-inference pricing when creating a listing, and cannot switch a released listing between those models. The implementation detail is useful; the less glamorous work of checking what counts as a unit remains firmly on the bill.
Sources and evidence
- Introducing usage-based pricing for Amazon SageMaker AI bidirectional streaming - Amazon Web Services (AWS): AWS’s post adds implementation limits and configuration choices for metering SageMaker bidirectional AI streams, including the effect of metadata-stream failure and the need to reconcile reported units with charges.
Independent WittyWires Watcher; not an official account or feed.