Deepgram has introduced enhanced metrics for Amazon SageMaker AI deployments, offering detailed insights into billing, feature usage, and engine behavior. These metrics are published directly into the user’s CloudWatch account, providing visibility into how Deepgram’s speech-to-text and text-to-speech models are being used without requiring additional agents or permissions. This advancement supports data residency and compliance efforts while maintaining managed control over deployment and monitoring.
The new metrics include usage and billing data that align with AWS Marketplace metered billing, allowing users to reconcile their AWS bills against actual traffic down to the model and transport. These metrics are available through the Deepgram container, which writes CloudWatch Embedded Metric Format (EMF) records to container stdout. SageMaker AI forwards these records to the endpoint’s CloudWatch log group, where they are automatically extracted into metrics. This method works under AWS Marketplace network isolation, as metrics travel over the existing SageMaker-to-CloudWatch logging path rather than any outbound network connection.
The source explains that Deepgram Enhanced Metrics provide visibility into what users are being billed for and how their traffic uses Deepgram’s features. These metrics are published under the Deepgram/SageMakerInference namespace, with dimensions covering category, model, and transport. They do not include endpoint names or instance IDs, so users cannot filter the stream to a single endpoint or instance. For per-endpoint, per-instance, or per-GPU breakdowns, users can use Prometheus and OpenTelemetry metrics.
Source: awsml
