article --summary
The summary
The post argues that accelerator costs are high while most organizations lack visibility into who uses their GPUs, whether devices are healthy, and where money is lost. Shared clusters and failures often leave teams guessing, and restarting whole instances wastes healthy hardware.
Amazon CloudWatch accelerated compute insights, launched at KubeCon North America 2026, brings accelerator telemetry, Kubernetes context, and pricing data into one operational view covering NVIDIA GPUs, Trainium, and Inferentia on EKS. It attributes usage by pod, namespace, and team, handles MIG and other partitioning schemes, and flags waste in right-sizing, training efficiency, and underused inference endpoints.
The tool can also route health events to the AWS DevOps Agent, which proposes evidence-backed fixes for approval. The article then walks through installing the CloudWatch Observability add-on on EKS to start collecting data.
Editorial note: This is a summary of an article first published by The New Stack. The full piece is on their site.