Amazon SageMaker adds billing by use for live voice AI
Model providers can report the audio that customers use in a live session, and Deepgram is the first partner.
Amazon tells of billing by use for bidirectional streaming (live data in two directions) on Amazon SageMaker AI. Model providers on AWS Marketplace can report the use of a live session for their software charge. Deepgram, a company that makes real-time voice AI, is the first partner. The charges for the endpoint infrastructure continue to apply.
How it works
A live session can have a different length each time, and the total is not known until it ends. SageMaker opens a second WebSocket connection (a live data connection) adjacent to the data stream. The model server sends one metering record at a time on it, with at most one message each second.
The numbers
If a product has a charge for each second of audio, 12 minutes of metered audio is 720 billable seconds. A record can be 512 bytes at most. The rule of the provider tells which audio is metered.
What is not known
Amazon tells that a sent record does not prove that the use was billed. Customers must compare the units with their AWS Marketplace reports. If the connection breaks, inference continues by default. Amazon gave no prices.
Insight: Amazon keeps the meter and the bill
The provider reports use, but the charge goes through AWS Marketplace and into the AWS bill of the customer. In October, Scaled Cognition released its voice model at $9 for each million characters. This reads as a toll gate: the one who has the bill is between the model makers and the buyers. Buyers get one bill, but must check the numbers.
Sources
Posted
