Cloud & Architecture · Paid
Amazon EC2 Capacity Blocks for ML
Amazon EC2 Capacity Blocks for ML is an AWS purchasing option that lets teams reserve GPU and Trainium instances for a fixed window on a future date, rather than hoping on-demand capacity is free when they need it. You search the available offerings, pick a start date up to eight weeks ahead and a duration, and pay the full price up front. When the block starts, the reserved instances are yours until it ends, and AWS places them close together for low-latency, high-bandwidth networking.
Who it is for: machine learning teams that need a known number of accelerators for days or weeks, such as a fine-tuning run, a large evaluation sweep, a product launch that will spike inference traffic, or a prototype that needs top-end GPUs briefly. It suits teams that cannot justify a one- or three-year commitment but also cannot risk spot interruptions or failed on-demand launches.
Key features: - Reserve capacity with a start time up to eight weeks in the future, and see offerings that can start in as little as 30 minutes. - Up to 64 instances per Capacity Block and up to 256 instances across Capacity Blocks (lower block sizes apply to some instance types and Regions). - Supported instance types include P6-B300, P6-B200, P5, P5e, P5en, P4d and P4de GPU instances and Trainium (Trn1, Trn2) instances in selected Regions, plus UltraServer blocks for some configurations. - Works with Amazon EKS managed and self-managed node groups, AWS ParallelCluster and AWS Parallel Computing Service. Instances must target the reservation ID. - Instances in a Capacity Block do not count against On-Demand Instance limits.
Constraints to plan around: the price is charged up front and cancellations are not allowed. Blocks cannot be moved or split, they end at 11:30 AM UTC, and AWS begins terminating instances at 11:00 AM UTC on the final day, so long-running inference servers need a drain plan before then.
Pricing type: paid, with a per-block price shown at purchase time; operating system charges are billed separately while instances run. See the official Capacity Blocks pricing page for current rates.
Alternatives: EC2 On-Demand Capacity Reservations (any duration, billed at the On-Demand rate whether used or not), Compute Savings Plans for steady long-term usage, EC2 Spot Instances for interruptible work, and Google Cloud future reservations in calendar mode for short GPU windows.
Official URL: https://aws.amazon.com/ec2/capacityblocks/
