Deepgram Boosts Amazon SageMaker AI Support with AWS IAM Temporary Delegation

Deepgram, a developer of AI speech models (Nova, Flux, Aura-2), announced an integration with AWS's IAM Temporary Delegation mechanism to enhance support for deployments on Amazon SageMaker AI. Previously, troubleshooting issues required Deepgram engineers to coordinate screen-sharing sessions with clients, which took days. The new procedure works as follows: a Deepgram engineer sends a /delegate_access command in the support ticket system, the client specifies the Amazon Resource Name (ARN) of their SageMaker AI endpoint, after which a delegation request appears in their IAM console. The client can see the exact permissions (read-only access to one endpoint and one CloudWatch log group for 12 hours) and can approve or deny the request. Upon approval, AWS issues temporary STS tokens to Deepgram, which automatically expire after the set duration (up to 12 hours). All actions using delegated credentials are logged in the client's AWS CloudTrail. The integration is based on a pre-registered permission template, DeepgramSageMakerReadOnlyTroubleshooting, and uses Amazon Simple Notification Service (SNS) for token exchange. Prerequisites include an active Deepgram subscription, a deployed model on SageMaker AI, IAM permissions to approve delegation requests, and enabled CloudTrail. Deepgram also offers a 14-day trial of its models, though the client pays for the AWS infrastructure costs (GPU instances, logs, networking). As a result, the time for initial incident analysis has been reduced from days to minutes, and security has been improved by eliminating long-lived cross-account roles.