Amazon SageMaker Adds Instance Preference Lists to Auto-Route Training Jobs to Available GPUs
Summary
- • SageMaker now accepts a ranked list of up to 5 GPU instance types per training job
- • System auto-launches on first available type, cutting peak-demand wait times
- • Feature covers both Training Jobs and Processing Jobs; integrates with reserved capacity plans
- • Eliminates brittle manual retry scripts for teams running multi-GPU workloads
Details
Instance preference lists now live for SageMaker Training and Processing Jobs
Amazon SageMaker AI now supports specifying up to 5 ranked instance types per Training Job or Processing Job, auto-launching on the first type with available capacity.
Addresses GPU scarcity bottleneck teams face during peak demand periods
Teams previously built custom retry scripts or waited for preferred GPU types; the new feature removes this manual overhead and the operational fragility of single-instance-type pinning.
Priority-ordered evaluation scans up to 5 instance types automatically
SageMaker evaluates the preference list in defined priority order and launches on the first instance type that has available on-demand capacity, with no manual intervention required.
Integrates with Flexible Training Plans for reserved capacity with fallback
Jobs can prioritize reserved Flexible Training Plan capacity first, then automatically fall back to alternative on-demand instance types from the preference list if reserved capacity is unavailable.
Directly reduces delays in nightly retraining and production fine-tuning pipelines
Nightly retraining pipelines, production fine-tuning runs, and scheduled processing jobs benefit from lower wait times, higher capacity utilization, and more engineering time on model development.
Amazon SageMaker AI instance preference lists feature details from the AWS ML Blog announcement, published September 2026.
What This Means
AWS's new instance preference lists solve a real ML operations pain point: training jobs stalling because a specific GPU type is unavailable at peak demand. By letting teams rank up to five acceptable instance types and having SageMaker auto-route to the first available, the feature cuts idle wait times and eliminates the brittle manual retry scripts many teams rely on. For large-scale AI training shops running on AWS, the practical impact on pipeline reliability and compute utilization could be significant, particularly for teams with time-sensitive production workloads.