Amazon SageMaker AI
Platform to build and deploy ML
A complete ML platform (formerly Amazon SageMaker): data preparation, notebooks/Studio, training on managed instances, tuning, deploying inference endpoints, and MLOps (Pipelines, Model Registry).
- Cost$$$ High
- Difficulty●●●●●
- Costs money when idle?Yes
- Managed by AWS
- In examsCLF · SAA
Think of it asLike an all-in-one manufacturing plant: raw materials, machinery, and delivery lines are all there, you design the product yourself.
When to use it
✓ Use it when
You need to train, tune, and deploy your own ML models.
✗ Avoid it when
You only need to call existing models or generative AI → Bedrock / specialized AI services (Rekognition, Comprehend…).
Common uses
- Train churn prediction models
- Deploy real-time inference endpoints
- MLOps pipelines for periodic retraining
What does it cost?
- Pricing model: Charged per instance hour for notebooks, training, and endpoints, plus storage; serverless inference and Savings Plans are available as options.
Forgetting to shut down endpoints or notebooks/Studio apps is the most common cost leak; use Spot for training to reduce costs.
When the question says build, train and deploy your own ML model, MLOps, choose SageMaker AI; for foundation models through an API, Bedrock; for image or text recognition without ML expertise, Rekognition or Comprehend; for inference on a whole dataset with no endpoint, batch transform.
Frequently asked questions
How is SageMaker AI different from Bedrock and prebuilt AI services (Rekognition, Comprehend)?
Rekognition, Comprehend, Transcribe and similar are prebuilt AI APIs for one specific problem, needing no ML knowledge. Bedrock gives access to foundation models through an API. SageMaker AI is for when you build, train, fine-tune and deploy your own models on your own data, which needs more ML knowledge but is the most flexible.
What inference options does SageMaker AI offer?
Real-time endpoints for low latency and always-on serving. Serverless inference for spiky traffic, scaling to zero when idle. Asynchronous inference for large or long-running requests, queued. Batch transform to run inference on a whole dataset at once without a permanent endpoint.
How do I reduce cost with SageMaker AI?
Real-time endpoints bill even when no requests arrive, so delete endpoints you do not use or switch to serverless or batch. Training can use managed Spot for lower cost, tolerating interruptions with checkpoints. Remember to stop notebooks or Studio apps you are not using.
