AWS EC2 User Guide - Capacity Blocks for ML: 8-week booking horizon, 64/256 instance caps

PRIMARY AWS documentation, fetched 2026-10-04. A rare case of a hyperscaler publishing a HARD capacity lead time rather than leaving it to 'contact sales': 'You can reserve a Capacity Block with a reservation start time UP TO EIGHT WEEKS IN THE FUTURE. Each Capacity Block can have up to 64 instances, and you can have up to 256 instances across Capacity Blocks.' Why that number matters and what it does NOT mean: 8 weeks is the MAXIMUM booking horizon, i.e. a FLOOR on how far ahead scarce GPU capacity must be committed, not a typical wait. It is the contractual expression of scarcity - AWS will not take a GPU reservation further out than 8 weeks because it does not want to warrant capacity that far forward. Two further constraints from the same page: Capacity Blocks are 'available for select instance types in some AWS Regions' (so it is not universal even within one account), and AWS recommends them 'when you need GPUs to support your ML workloads for days or weeks at a time and don't want to pay for a reservation while your GPU instances aren't in use' - i.e. the product is deliberately positioned as the bridge between spot/on-demand and a 1-3 year reservation. Cost basis matters here: the repo's aws-ec2-* records price Capacity Blocks alongside on-demand per GPU-hour (p5.48xlarge Capacity Block $41.528 vs $55.04 on-demand, about 25% off), so the block is the discount tier that requires committing 8 weeks ahead. Also note the page's own framing that instances inside a block are co-located in an EC2 UltraCluster for low-latency petabit-scale non-blocking networking, which is the infrastructure reason a block is worth its premium for multi-node training.