SageMaker AI
Amazon SageMaker AI is a managed machine learning service covering data preparation, model building, training, tuning, deployment and monitoring.
A note on the name
Section titled “A note on the name”On 3 December 2024 AWS renamed the model-building service from Amazon SageMaker to Amazon SageMaker AI, and reused “Amazon SageMaker” for a larger platform that unifies data, analytics and AI. That platform includes:
- Amazon SageMaker AI — build, train and deploy models (what this section documents)
- Amazon SageMaker Lakehouse — unified access across S3 data lakes, Redshift and other sources
- Amazon SageMaker Unified Studio — the single development environment for analytics and AI, generally available since March 2025
- Amazon SageMaker Catalog — data and AI governance, built on Amazon DataZone
- SQL analytics, data processing and Amazon Bedrock alongside them
The API namespaces, CLI commands, CloudFormation resource types (AWS::SageMaker::*), managed policy names and service endpoints were all left unchanged for backward compatibility. Only the product name moved.
Where older material says “SageMaker Studio”, the current environment is SageMaker Unified Studio.
Core surfaces
Section titled “Core surfaces”Training
Section titled “Training”- Managed training jobs on provisioned compute that is released when the job finishes
- Built-in algorithms for common tasks, plus bring-your-own algorithms and containers
- Distributed training across multiple instances
- Automatic model tuning (hyperparameter optimisation) jobs
Deployment
Section titled “Deployment”Two shapes, with two more options between them.
Real-time inference — synchronous predictions behind an endpoint.
- Create a model
- Create an endpoint configuration
- Create the endpoint
- Call
InvokeEndpoint()
Batch transform — asynchronous predictions over a whole dataset at once, with no endpoint to keep running.
Serverless inference — endpoints with no infrastructure to configure, which scale to zero. Suited to workloads with idle periods that can tolerate a cold start.
Asynchronous inference — queues requests and processes them in the background. Suited to large payloads, long processing times and near-real-time rather than interactive latency.
Other features
Section titled “Other features”- Inference pipelines — chaining preprocessing, model and postprocessing containers behind one endpoint
- Model Registry — versioning and approval workflow for models
- SageMaker Pipelines — orchestration of the training and deployment workflow
- SageMaker Neo — compiling a model for a specific target architecture. See SageMaker Neo.
- JumpStart — deploying pre-trained and foundation models from a catalogue
Notebook Instances — a standalone managed Jupyter instance — remain available but are the legacy surface; new work starts in SageMaker Unified Studio or JupyterLab within a SageMaker AI domain.
Deployment type summary
Section titled “Deployment type summary”| Batch (offline) | Real-time (online) | |
|---|---|---|
| Usage | Asynchronous or batch | Synchronous, real time |
| When | Generate predictions for an entire dataset at once | Respond to a request as it arrives |
| Method | SageMaker AI batch transform | SageMaker AI hosting services (endpoint) |
| Input format | Varies by algorithm | Varies by algorithm |
| Output format | Varies by algorithm | JSON string |
| Cost shape | Pay for the duration of the job | Pay for the endpoint while it exists |
Integration with other AWS services
Section titled “Integration with other AWS services”- S3 — training data and model artefacts
- ECR — custom container images for training and inference
- IAM — execution roles and permissions
- CloudWatch — metrics and logs from training jobs and endpoints
- Step Functions and SageMaker Pipelines — workflow orchestration
- EventBridge — reacting to job and endpoint state changes
Data labelling
Section titled “Data labelling”Amazon SageMaker Ground Truth built labelled training datasets using a private workforce, a vendor workforce or Amazon Mechanical Turk, with automated labelling to reduce the human effort.
Ground Truth is no longer open to new customers; AWS moved it to maintenance on 30 July 2026 and does not plan to add features. Existing customers can continue to use it. SageMaker Ground Truth Plus, the fully managed variant, reached end of support on 30 June 2026. Amazon Mechanical Turk itself closes on 30 September 2026, so from that date its workforce option disappears from Ground Truth and Augmented AI; only private and vendor workforces remain.
Several other SageMaker AI features moved to maintenance on the same date and are closed to new customers: A2I, Clarify, Debugger, Model Monitor, Role Manager, Geospatial, Mechanical Turk and Studio Lab.
For new labelling work, plan on a third-party labelling platform or an internal tool writing manifests into S3 in the format the training job expects.
Cost notes
Section titled “Cost notes”- Real-time endpoints bill continuously while they exist, whether or not they are being called. Delete endpoints created for experiments.
- Use managed Spot training for interruptible training jobs.
- Batch transform and serverless inference avoid the standing cost of an always-on endpoint.
- Autoscaling on endpoints matches capacity to traffic rather than to peak.