AWS launches EC2 Capacity Blocks for ML and SageMaker training plans to reserve GPU capacity for short-term workloads

Amazon EC2 Capacity Blocks for ML lets customers reserve GPU capacity for a specific time window, with durations from 1–14 days (in 1-day increments) or 15–182 days (in 7-day increments), at 40-50% lower hourly rates compared to on-demand pricing. For example in US East (N. Virginia), p5.48xlarge costs $34.608/hour with Capacity Blocks versus $55.04/hour on-demand.

SageMaker training plans provide access to reserved GPU capacity within the Amazon SageMaker AI managed environment for training jobs, HyperPod clusters, and inference workloads, priced 70-75% below on-demand rates, with upfront payment required.

Capacity Blocks support selected instance families such as P5, Trn1 and Trn2, but do not cover every GPU instance type and cannot be used with Amazon SageMaker-managed instance types such as ml.p4dn or ml.p5, while SageMaker training plans cover a range of accelerated computing options including the latest NVIDIA GPUs and AWS Trainium accelerators (except G-type instances except G6).