Amazon SageMaker Costs: Notebooks, Endpoints and HyperPod
SageMaker spend spreads across notebooks, training, endpoints and HyperPod clusters. Each has a different off switch, or none.
Instance time, everywhere
SageMaker notebook instances and real-time inference endpoints are both charged for the instance type you choose, based on the duration of use. A notebook left running over a weekend and an endpoint serving no requests both bill the whole time.
What ZopNight can switch off
- Notebook instances are on ZopNight’s Start/Stop supported list, so they can follow a schedule, group or override like any VM.
- HyperPod clusters support start and stop as a scale to and from zero.
ZopNight’s SageMaker coverage goes beyond basic discovery: dedicated providers read endpoints, notebook instances, HyperPod clusters, domains and Studio apps, cost calculators cover training-job instance-hours, feature-group online storage and serverless inference, and endpoint Invocations and ModelLatency metrics feed the rules.
ZopNight’s SageMaker rules
Scheduling and idle:
- SageMaker Notebook Missing Auto-Stop prices a recurring stop across the notebook’s measured idle window.
- SageMaker Endpoint Non-Production Scheduling Opportunity and SageMaker HyperPod Cluster Non-Production Scheduling Opportunity. The HyperPod rule prices the measured off-hours idle fraction and reads an environment tag before it falls back to the cluster’s name.
- Idle SageMaker Endpoint, No Invocations and Idle SageMaker HyperPod Cluster, Near-Zero Utilization.
Sizing and pricing:
- SageMaker Endpoint Low Invocations, wired to a resize remediation, and SageMaker Endpoint Instance Count Over-Provisioned.
- Over-Provisioned SageMaker HyperPod Cluster and SageMaker Transform Job Over-Provisioned.
- SageMaker Training Job, Enable Managed Spot Training.
Posture:
- Notebook root access, direct internet access, no VPC, no lifecycle configuration, and volumes not encrypted with a customer-managed key; endpoint data capture disabled; HyperPod clusters outside a VPC.
Related
Fact sheets: notebook instances, endpoints and HyperPod clusters. Scheduling notes: notebooks and HyperPod. Broader AI spend is on AI cloud management.