This article explains how to use Amazon SageMaker AI training plans to reserve GPU capacity for inference endpoints, ensuring predictable availability for time-bound workloads.

• Training plans now support inference endpoints, not just training jobs
• Search for available p-family GPU capacity using search-training-plan-offerings API
• Create a training plan reservation and receive an ARN for capacity reference
• Deploy inference endpoints using the reserved capacity ARN in endpoint configuration
• Set CapacityReservationPreference to "capacity-reservations-only" to restrict to reserved capacity
• Endpoint stops serving traffic when reservation expires if using capacity-reservations-only
• Update endpoints to new model versions while maintaining reserved capacity
• Migrate from reserved to on-demand capacity if needed beyond reservation window
• Scale endpoints within reservation limits; scaling beyond fails with validation error
• Full reservation cost charged upfront regardless of actual usage duration
• Deleting endpoint doesn't cancel or refund the training plan reservation

SageMaker AI training plans provide transparent, upfront pricing for predictable GPU availability during model evaluation, testing, and burst workloads.