What's New

Amazon SageMaker Studio IDEs — specifically JupyterLab and Code Editor — now support GPU capacity reservations through SageMaker Flexible Training Plans (FTP), enabling data scientists and ML engineers to secure predictable access to high-demand GPU instances directly from their interactive development environments. This integration extends FTP, previously available for training jobs and HyperPod clusters, into the Studio IDE workflow, offering up to 65% cost savings compared to On-Demand pricing. The experience is fully self-serve, requiring no infrastructure management from the user.

How It Works

• Purchase a Training Plan: Navigate to the SageMaker FTP console, select your desired instance type (e.g., GPU-accelerated or AWS Trainium), specify the quantity, reservation duration, Availability Zone, and start date, then complete the prepaid purchase.
• Plan Activation: After purchase, the training plan transitions to an "Active" status before it can be used; a Studio app can only attach to a plan that is in Active state.
• App Creation: When launching a JupyterLab or Code Editor space in SageMaker Studio, select the purchased training plan from the Instance dropdown — SageMaker automatically provisions the reserved instance with no manual infrastructure setup.
• Billing Model: Training plans are prepaid; you are not charged separately for compute time consumed within the Studio app while running on reserved capacity. Standard charges for storage (EBS volumes) and other resources still apply.
• Automatic Shutdown: Apps running on training plan capacity are automatically shut down 30 minutes before the capacity block expires, and the IDE proactively notifies you so you can save your work in time. You are not charged for this 30-minute wind-down period.
• IAM Permissions: The Studio execution role must include sagemaker:ListTrainingPlans and sagemaker:DescribeTrainingPlan permissions for UI access, or just sagemaker:DescribeTrainingPlan for API/CLI usage.
• VPC Requirement: The Studio domain's VPC configuration must include a subnet in the same Availability Zone specified in the training plan, with at least one free IP address available.
• Plan Immutability: The target resource type of a training plan cannot be changed after purchase, and plans purchased for Studio apps cannot be cancelled.

Why It's Important

• Eliminates GPU Availability Uncertainty: High-performance GPU instances (e.g., latest NVIDIA GPUs, AWS Trainium) are frequently oversubscribed; FTP gives teams guaranteed access during critical development or experimentation windows, removing the risk of capacity failures at launch time.
• Significant Cost Reduction for Interactive Workloads: Up to 65% savings versus On-Demand pricing makes sustained GPU-backed interactive development economically viable for organizations running long-running notebooks or iterative ML experiments.
• Bridges the Gap Between Experimentation and Training: Previously, FTP was limited to formal training jobs and HyperPod clusters; extending it to Studio IDEs means the same cost and capacity benefits now apply to the exploratory, pre-training phase of the ML lifecycle.
• Predictable Budgeting: Prepaid, fixed-duration reservations allow finance and ML platform teams to plan and allocate GPU compute costs in advance, rather than dealing with variable On-Demand billing.
• Reduced Operational Overhead: SageMaker handles provisioning, infrastructure management, and fault recovery automatically, letting data scientists focus on model development rather than infrastructure operations.

How It's Different

• Extends FTP Beyond Training Jobs: Previously, SageMaker Flexible Training Plans only supported SageMaker Training Jobs and HyperPod clusters; this is the first integration with interactive Studio IDE environments (JupyterLab, Code Editor).
• Interactive vs. Batch Workloads: Unlike training job reservations that are tied to automated batch runs, Studio IDE reservations support interactive, human-in-the-loop workflows where a developer is actively working in a notebook or editor.
• Prepaid vs. On-Demand: Standard Studio GPU instances are billed On-Demand per second; FTP capacity is prepaid upfront, trading flexibility for substantial cost savings (up to 65%) and guaranteed availability.
• Proactive Expiry Notifications: The IDE actively notifies users as the reservation nears expiration — a user-experience consideration not relevant in automated training job contexts — and provides a 30-minute grace period before automatic shutdown.
• Self-Serve Procurement: Unlike enterprise capacity agreements or private pricing negotiations, FTP is fully self-serve through the SageMaker console, lowering the barrier to securing reserved GPU capacity.
• Plan Specificity: Training plans are bound to a specific target resource type at purchase time and cannot be reassigned, meaning a plan purchased for Studio apps is dedicated to that use case and cannot be repurposed for training jobs.

When to Prefer It

• Extended GPU-Intensive Notebook Sessions: When your team runs long-duration, GPU-heavy experiments in JupyterLab (e.g., fine-tuning large models interactively, running iterative hyperparameter searches manually), reserving capacity ensures uninterrupted access and reduces cost versus On-Demand.
• Scheduled Development Sprints: When a team has a defined project window (e.g., a two-week model development sprint) requiring consistent GPU access, FTP lets you lock in capacity and cost for that entire period.
• High-Demand Instance Types: When targeting oversubscribed instance types (e.g., p4d, p5, trn1) that are frequently unavailable On-Demand, FTP guarantees the capacity will be available when your plan activates.
• Budget-Constrained ML Projects: When operating under a fixed compute budget, the prepaid model and up to 65% discount make GPU-backed Studio development financially feasible compared to On-Demand rates.
• Regulated or Predictable Workload Schedules: When your organization requires predictable, auditable compute spend (e.g., for compliance or financial reporting), prepaid training plans provide clear cost attribution.
• Avoid Using When: You need ad-hoc, unpredictable GPU access with no defined schedule — On-Demand instances remain more appropriate for sporadic, short-duration, or unpredictable workloads where the overhead of planning a reservation outweighs the savings.

Availability

• General Availability: This feature is generally available as of May 18, 2026.
• Supported IDE Types: Only JupyterLab and Code Editor app types within SageMaker Studio are supported; SageMaker Studio Classic is not supported.
• Supported Instance Types: Includes high-performance GPU instances (latest NVIDIA GPU technologies) and AWS Trainium chips; specific supported instance types are detailed in the SageMaker training plans documentation.
• Regional Availability: Supported regions are aligned with existing SageMaker Flexible Training Plans regional availability; consult the SageMaker pricing and regional services pages for the current list of supported AWS Regions.
• Pricing Model: Training plans are prepaid at reservation time; no separate compute charge is incurred during Studio app runtime on reserved capacity. Standard storage (EBS) and other ancillary resource charges still apply. Up to 65% savings versus On-Demand instance pricing.
• Plan Cancellation: Plans purchased for Studio apps cannot be cancelled after purchase, making it important to confirm instance type, duration, and AZ before completing the transaction.
• Quota Dependency: App creation will fail if the requested instance type exceeds your Studio instance type quotas; ensure quotas are sufficient before purchasing a plan.
• VPC Constraint: The Studio domain VPC must have a subnet in the same Availability Zone as the training plan's reserved capacity block, with at least one available IP address.

Related Resources

• https://docs.aws.amazon.com/sagemaker/latest/dg/reserve-capacity-with-training-plans.html
• https://docs.aws.amazon.com/sagemaker/latest/dg/training-plan-utilization-for-studio-apps.html
• https://docs.aws.amazon.com/sagemaker/latest/dg/studio-updated-spaces.html