Amazon SageMaker AI

What it is

Amazon SageMaker AI is a fully managed service for data scientists and developers to build, train, and deploy ML (Machine learning)Software that learns patterns from example data to make predictions or decisions, instead of following rules written by hand. A model is first trained on data, then used to make predictions on new data, which is called inference.Glossary entry (ML) models into a production-ready hosted environment. It provides managed notebooks and IDEs, managed training on scalable infrastructure, built-in algorithms, support for your own algorithms and frameworks, and managed endpoints for predictions.

Naming: on December 3, 2024, Amazon SageMaker was renamed Amazon SageMaker AI. The name "Amazon SageMaker" now refers to a broader unified platform for data, analytics, and AI, which includes SageMaker AI. API (Application programming interface)A defined way for one program to ask another for data or actions, such as a set of HTTPS requests and responses. Every AWS service is controlled through its API, which the console, the CLI, and the SDKs call for you.Glossary entry namespaces, CLI commands, and policy names still use sagemaker. Exam questions may simply say "SageMaker"; they mean the ML service described here.

Choose SageMaker AI when a question involves building or training a custom ML model on your own data, or hosting a model for inference. When a pretrained AI service already does the job, such as detecting objects in images (Rekognition) or extracting text from documents (Textract), use that service instead: no ML expertise is needed.

Build, train, deploy

Build:
- SageMaker Studio and notebooks give you managed environments to explore data and write code.
- SageMaker Canvas lets business analysts create ML predictions with no code, such as churn or demand forecasts. Data Wrangler, for preparing data, is now part of Canvas.
- Feature Store stores and shares ML features, with an online store for real-time inference and an offline store for training.
Train:
- A training job runs on managed instances that SageMaker AI starts and stops for you, reading data from Amazon S3 and writing the model back to S3.
- Use built-in algorithms, popular frameworks, or your own container.
- SageMaker JumpStart provides pretrained, open-source models and solution templates that you can tune and deploy.
- Managed Spot Training runs training jobs on EC2 Spot capacity, cutting cost by up to 90%. SageMaker AI handles interruptions.
Deploy: choose the inference option that matches the traffic (see the next section).
Automate: SageMaker Pipelines orchestrates ML workflows (prepare, train, evaluate, register, deploy) on serverless infrastructure. Model Registry tracks model versions and approvals.

Inference options

Option	Best for	Key facts
Real-time inference	Interactive, low-LatencyThe time between sending a request and receiving the response, usually measured in milliseconds. Distance and processing add to it; caching and serving users from nearby locations reduce it.Glossary entry predictions	A persistent endpoint on instances you choose; supports autoscaling
Serverless Inference	Traffic with idle periods that can tolerate Cold startThe extra delay when AWS Lambda must create a new execution environment, load the code, and run its initialization before handling a request. Later requests reuse the environment and are faster; provisioned concurrency avoids cold starts.Glossary entry	No instances to choose; scales automatically, down to zero; pay for use. No GPU (Graphics processing unit)A processor with thousands of small cores that run many calculations in parallel. GPUs speed up machine learning, graphics, and video work; EC2 instance families such as P and G include them.Glossary entry, VPC configuration, or network isolation
Asynchronous Inference	Large payloads (up to 1 GB) or long processing (up to one hour)	Queues requests; writes results to S3 and can send success and error notifications through SNS; can scale to zero when idle
Batch transform	Predictions on a whole data set, no persistent endpoint	Reads from and writes to S3

When the model needs a GPU, must reach resources in your VPC, or must run with network isolation, Serverless Inference is ruled out: use a real-time endpoint (or asynchronous inference for large or long requests).

Training and deploying on SageMaker AI. A training job, on managed instances that can use Spot capacity, reads training data from Amazon S3 and writes model artifacts back to S3. The model is deployed to a real-time endpoint that an application calls for predictions, or used by batch transform to score a whole data set.

Feature availability changes

SageMaker Ground Truth (data labeling), SageMaker Model Monitor (data and model quality monitoring), and SageMaker Clarify (bias detection and explainability) are no longer open to new customers. Existing customers can keep using them, but no new features are planned. Older exam questions may still name them for labeling, drift detection, and bias, so recognize what they did, but do not choose them for a new design.

Security, resilience, performance, and cost

Security

IAM execution roles give notebooks, training jobs, and endpoints access to S3 and other services.
Run training and endpoints inside your VPC, and use interface VPC endpointA private connection from a VPC to an AWS service or another endpoint service without crossing the internet. Gateway endpoints, for Amazon S3 and DynamoDB, are route table targets; interface endpoints are network interfaces that use AWS PrivateLink.Glossary entry (AWS PrivateLink) to call SageMaker AI APIs privately.
To invoke an endpoint from SubnetA range of IP addresses in a VPC, inside one Availability Zone. A subnet is public when its route table sends internet traffic to an internet gateway, and private when it does not.Glossary entry without internet access, create an interface VPC endpoint for the SageMaker AI Runtime (runtime.sagemaker), which serves InvokeEndpoint. The API endpoint (api.sagemaker) covers management calls only, and an endpoint's own VPC configuration only lets the model container reach your VPC resources; it gives clients no private path to the endpoint.
Network isolation (internet-free mode) blocks containers from reaching the internet, which they can do by default.
Encryption at restEncrypting data where it is stored, such as on disks, in databases, or in backups, so that it cannot be read without the key. On AWS it usually uses keys in AWS KMS.Glossary entry with AWS KMS for storage volumes, model artifacts, and outputs in S3; TLS (Transport Layer Security)The protocol that encrypts traffic between a client and a server and lets the server prove its identity with a certificate. It replaced SSL, though the older name is still common; HTTPS is HTTP over TLS.Glossary entry in transit.
CloudTrail records API calls.

Resilience

Deploy production endpoints on two or more instances; SageMaker AI automatically tries to spread them across AZ (Availability Zone)An isolated location within an AWS Region, made up of one or more data centers with their own power, cooling, and networking. The AZs in a Region are linked by fast, low-latency connections, so spreading resources across them protects against the failure of one site.Glossary entry, so an instance or Availability Zone failure does not take the endpoint down.
Model artifacts and data live in S3, so you can redeploy at any time.

Performance

Autoscaling adjusts endpoint instance counts with the workload.
Choose GPU or accelerated instances for deep learning training and inference, and distributed training for very large models and data sets.

Cost

Pay for instance time for notebooks, training, and hosting, plus storage.
Cut costs with:
- Managed Spot Training.
- Serverless or Asynchronous Inference for intermittent traffic (asynchronous can scale to zero).
- Batch transform instead of an always-on endpoint.
- SageMaker AI Savings Plans.
Stop idle notebooks and delete unused endpoints.

Monitoring

CloudWatch metrics and logs for training jobs and endpoints (invocations, latency, errors). CloudTrail records API activity.

SageMaker AI or an AI service?

If the requirement is	Choose
Build, train, and host a custom ML model on our data	Amazon SageMaker AI
No-code predictions for business analysts	SageMaker Canvas
Pretrained open-source models to fine-tune	SageMaker JumpStart
Detect objects, faces, or unsafe content in images and video	Amazon Rekognition
Extract text, forms, and tables from documents	Amazon Textract
Find sentiment, entities, or key phrases in text	Amazon Comprehend
Generative AI with foundation models through an API	Amazon Bedrock

Reverse mindmap: when the question points here

Is it Amazon SageMaker AI?

Yes, when the question says
- Build and train a custom ML model leads to SageMaker AI
- Data scientists need managed notebooks and training leads to SageMaker AI
- Host a model behind an endpoint leads to SageMaker AI inference
Choose the inference option
- Low-latency, steady traffic leads to Real-time endpoint with autoscaling
- Intermittent traffic, cold starts are fine leads to Serverless Inference
- Needs GPUs, VPC access, or network isolation leads to Real-time endpoint, not serverless
- Payloads up to 1 GB or jobs up to an hour leads to Asynchronous Inference
- Notify when an async prediction finishes or fails leads to Async Inference SNS topics
- Score a whole data set in S3 leads to Batch transform
Save and simplify
- Cut training cost by up to 90% leads to Managed Spot Training
- No-code ML for analysts leads to SageMaker Canvas
- Start from a pretrained model leads to SageMaker JumpStart
- Automate train, evaluate, deploy leads to SageMaker Pipelines
Secure it
- Containers must not reach the internet leads to Network isolation
- Private API access from a VPC leads to Interface VPC endpoint
- Invoke an endpoint from private subnets leads to SageMaker AI Runtime interface endpoint
No, when the question says
- Pretrained vision, text, or speech task leads to The matching AI service

Key takeaways

SageMaker AI (renamed from Amazon SageMaker in December 2024) builds, trains, and deploys custom ML models.
Training jobs read data from S3 and write models to S3; Managed Spot Training saves up to 90%.
Inference options: real-time (low latency), serverless (idle periods; no GPUs, VPC configuration, or network isolation), asynchronous (large or long requests, results in S3 with SNS notifications, scales to zero), and batch transform (whole data sets).
Private clients invoke endpoints through an interface VPC endpoint for the SageMaker AI Runtime.
Canvas offers no-code ML; JumpStart offers pretrained models; Pipelines automates workflows.
Ground Truth, Model Monitor, and Clarify are no longer open to new customers.
Prefer a pretrained AI service when one already solves the problem.