Amazon SageMaker AI adds instance preference lists to simplify GPU provisioning for training and processing jobs

# What changed Amazon SageMaker AI now supports instance preference lists for Training and Processing Jobs. Instead of binding a job to a single instance type and waiting if that type lacks capacity, you provide an ordered list of up to five acceptable instance types. SageMaker AI evaluates the list in priority order, selects the first type with available capacity, and starts the job automatically.

# The problem this addresses During peak demand, preferred GPU instances can be unavailable. Teams tied to a single instance type often either wait or build brittle systems that poll capacity, cancel stalled requests, and resubmit with alternatives. Those workarounds add operational overhead, don't integrate cleanly with reserved capacity options, and slow down experimentation and time-critical pipelines.

# How the feature works

• Submit your job with an ordered preference list (maximum five instance types).
• SageMaker AI validates the job configuration and checks supported instance types and resource limits.
• The scheduler performs a single in-memory sweep across the ordered list and identifies the first type with available capacity.
• If a type is available, the job provisions immediately and begins execution.
• If none of the listed types have capacity at evaluation time, the job enters an event-driven queue and retries automatically. The retry window is bounded by MaxPendingTimeInSeconds, which you control.

# Concrete benefits

• Faster job starts: one API call, not multiple resubmits.
• Lower operational overhead: eliminates custom polling and resubmission scripts.
• Better use of reserved capacity: preference lists can include reservation-backed types first, then on-demand fallbacks.
• More predictable pipelines: time-sensitive jobs such as nightly retraining or scheduled processing are less likely to stall due to a single pinned instance type.

# Typical scenarios

• A team training a large model that's compatible with several GPU families submits a single job with a preference order. SageMaker AI finds an available type and starts the run without manual intervention.

• An organization with a Flexible Training Plan lists reserved types first so reservations are used when available, and falls back to on-demand types if the reservation pool is consumed.

# Practical configuration notes

• You can list up to five instance types, ordered by priority.
• SageMaker AI validates instance support and resource limits during submission.
• If no listed type is available immediately, the job goes to an event-driven queue and will retry automatically up to the limit set by MaxPendingTimeInSeconds.

# Bottom line Instance preference lists reduce the friction of securing GPU capacity for training and processing. By letting SageMaker AI evaluate multiple instance options in a single job submission, teams avoid brittle retry logic and get jobs started sooner, while still using reservations and on-demand capacity according to a defined order.