Home › AWS › AWS data services › SageMaker AI
AWS data services · ML and AI · about 10 min read
Amazon SageMaker AI (Studio, Feature Store, endpoints) and Amazon SageMaker Unified Studio
AWS's machine learning service for training, tracking and serving models, now called SageMaker AI, inside a wider SageMaker with Unified Studio.
In plain words
What it is
Machine learning, or ML, means training a model on past data so that it can predict something about new data. A plant wants to predict which curing presses will fail next week. A retailer wants to forecast how many tyres of each size each store will sell. Doing this well takes several steps: preparing features (the input columns the model learns from), training, tracking experiments, registering the chosen model, serving it so apps can ask for predictions, and watching it once it is live.
Amazon SageMaker AI is AWS's managed service for those steps. It gives you notebooks and an IDE in SageMaker Studio, training jobs that start servers, run your code and shut down, a Feature Store with an offline store in S3 and an online store for fast reads, a model registry, managed MLflow for experiment tracking (now called MLflow Apps, serverless), and endpoints that serve predictions in real time, on demand, asynchronously or in batches.
A concrete example: a tyre maker trains a model to predict press failures from vibration and temperature readings. In SageMaker AI the data scientist reads features from S3, runs a training job on a GPU instance, logs the run in MLflow, registers version 3 of the model, and deploys it to a real-time endpoint. The maintenance app calls the endpoint every hour for each press. The endpoint runs on instances that bill every hour whether or not anyone calls it.
The names changed. In December 2024 AWS launched a next-generation Amazon SageMaker that is a broad data and AI platform, with SageMaker Unified Studio as one workspace over EMR, Glue, Athena, Redshift, Bedrock and SageMaker AI, and a lakehouse layer built on Iceberg. The original ML service was renamed SageMaker AI and also works on its own. In mid-2026 AWS closed several SageMaker AI features to new customers, including Model Monitor, Clarify, Ground Truth and Debugger.
A fully equipped rented workshop for building and running models, next door to the warehouse where your materials are kept.
Big picture
Where it fits
The lit lane is where SageMaker AI works. Boxes on the left feed it; boxes on the right use what it produces.
SageMaker AI sits in the Use lane, where models are trained and called. On the reference platform that lane is mostly Databricks. Features are gold tables (prod_gold.features), built and tested like any other gold table. Models are tracked with MLflow and registered in Unity Catalog with the same owners and grants as the data. Forecasts run as batch scoring jobs, and Model Serving answers apps that need a result in seconds. The design rule is that nothing is copied out to a separate ML system.
SageMaker AI is therefore a 'keep if you have it' service here. If a team already runs SageMaker endpoints that work, there is no reason to break them. Point their training at gold tables through Delta Sharing or the Iceberg view of the tables, and move them when they need rework anyway. The site's AWS versus Databricks table says the same: use Databricks for ML on lakehouse data, and call Bedrock models from Databricks where the company already uses Bedrock.
SageMaker Unified Studio and the SageMaker lakehouse are a different question. They compete with Unity Catalog and the Databricks workspace as the place where data people work. Running both as primary platforms means two catalogs, two permission models and two places to look for lineage. The platform picks Unity Catalog, so Unified Studio is not part of the design.
How it works
The ideas you need
| Idea | What it means |
|---|---|
| SageMaker AI | The renamed original ML service: training, tuning, registry, MLflow, Feature Store and endpoints. |
| SageMaker Unified Studio | One web workspace in the next-generation SageMaker for data, analytics and AI work across several AWS services. |
| Training job | A managed run that starts instances, executes your training code on data from S3, saves the model, and shuts the instances down. |
| Feature Store | A store of model inputs with an offline part in S3 for training and an online part for low-latency reads at prediction time. |
| Model registry | A catalogue of model versions with approval status, used to decide which version goes to production. |
| MLflow Apps | Serverless managed MLflow in SageMaker AI for tracking experiments and models, replacing MLflow tracking servers. |
| Real-time endpoint | Instances that keep a model loaded and answer requests in milliseconds; they bill every hour they run. |
| Serverless and asynchronous inference | Serverless scales to zero between calls with cold starts; asynchronous queues large requests and writes results to S3. |
| Batch transform | A job that scores a whole dataset in S3 and writes predictions back, with no endpoint left running. |
Choosing
When to use it, and when not
Use it when
- Your team already runs SageMaker training pipelines and endpoints in production and they meet their targets.
- You need special hardware or large distributed training, such as many GPU nodes for a foundation-model fine-tune, through SageMaker HyperPod or training jobs.
- You want to deploy an open-weight model from SageMaker JumpStart on your own instances inside your VPC.
- The workload is AWS-only, with no Databricks, and data already lives in S3 and the Glue catalog.
- An app needs a model behind an endpoint with deep AWS networking controls such as VPC-only access and per-endpoint IAM.
Avoid it when
- The training data is in Unity Catalog gold tables and the team uses Databricks every day; copying data out adds a second copy, a second set of grants and broken lineage.
- You need model monitoring for a new account; SageMaker Model Monitor and Clarify are closed to new customers since mid-2026.
- You want a second primary workspace and catalog (SageMaker Unified Studio and its lakehouse) next to Databricks and Unity Catalog.
- Predictions are needed once a day for a forecast; a batch scoring job in Databricks is simpler than an endpoint.
Production grade
What goes wrong in production
| Problem | What breaks | What to do |
|---|---|---|
| Idle endpoints | Real-time endpoints for old experiments keep running on GPU instances and bill around the clock. | Tag every endpoint with an owner and expiry, review weekly, and use serverless, asynchronous or batch inference where traffic is light. |
| Data copied out of governance | Features are exported from gold to an S3 bucket for training, outside Unity Catalog grants, masks and lineage. | Read gold through Delta Sharing or the Iceberg view with read-only rights, and never make an untracked copy; better, train in Databricks. |
| Feature skew | The online feature store computes a feature differently from the offline training data, so live predictions are wrong. | Compute features once in gold, publish the same table to the online store, and test that training and serving values match. |
| Features closed to new customers | A new account cannot turn on Model Monitor, Clarify or Ground Truth, which older designs relied on. | Use AWS's suggested open-source monitoring with MLflow and Evidently, or Databricks data quality monitoring and inference tables. |
| Two model registries | Some models are registered in SageMaker and others in Unity Catalog, so nobody knows which version is live. | Pick Unity Catalog as the registry of record; if SageMaker serves a model, record its endpoint and version against the Unity Catalog entry. |
| Renamed products confuse teams | People mix up SageMaker AI, the next-generation SageMaker, SageMaker Unified Studio and SageMaker Studio. | Write down which parts you use: here, only SageMaker AI endpoints you already have. |
Alternatives
What to compare it with
| Alternative | Pick the alternative when | Pick SageMaker AI when |
|---|---|---|
| Databricks Mosaic AI (MLflow, feature tables, Model Serving) | Your data is in Unity Catalog and you want one set of grants, lineage and release path for data and models. | You need hardware or training scale Databricks does not offer in your region. |
| Amazon Bedrock | You want to call or fine-tune foundation models through an API without managing instances. | You train your own classical model, such as gradient boosting on sensor features. |
| Azure Machine Learning | You are on Azure. | You are on AWS. |
| Google Vertex AI | You are on Google Cloud. | You are on AWS. |
| Open source: MLflow with KServe or BentoML on EKS | You want full control and portability and can run Kubernetes. | You do not want to run and secure the serving platform yourself. |
Cost
How you pay
SageMaker AI charges per second for the instances you use: notebook and Studio spaces, training jobs, processing jobs and endpoints, with GPU instances costing far more than CPU ones. Serverless inference is charged by compute time and data processed, asynchronous and batch by the instances they start. Feature Store charges for reads, writes and storage. Serverless MLflow Apps come at no extra charge, but the storage they use is billed. Savings Plans for SageMaker can cut steady usage. Check the SageMaker AI pricing page for your region.
What drives the bill
- Real-time endpoint instance hours, especially GPU endpoints left running.
- Training instance hours and the number of tuning trials.
- Studio and notebook spaces left running.
- Feature Store online reads and writes at high volume.
Cost traps
- An endpoint per experiment that nobody deletes.
- Studio spaces with GPU instances left open over a weekend.
- Data copied to S3 for training, then kept forever.
- Paying for both Databricks Model Serving and SageMaker endpoints for the same model during a long migration.
Prices change; check the AWS pricing page for your region before deciding.
Security and governance
Keeping it safe
- Run Studio domains, training jobs and endpoints in private subnets in VPC-only mode, with VPC endpoints for S3, SageMaker API and runtime.
- Give each execution role read access only to the training data prefix it needs, and use Delta Sharing or read-only Iceberg access for gold.
- Encrypt volumes, model artifacts and Feature Store data with customer managed KMS keys.
- Use IAM conditions to stop endpoints being created without a VPC config or without tags.
- Never train on personal data that has not been tokenised; the platform swaps names and contacts for tokens before bronze.
- CloudTrail records SageMaker API calls; include endpoint creation and model registration in the access review.
Operating it
What to watch, and the limits
| Watch | Why |
|---|---|
| Invocations and InvocationsPerInstance | Traffic on each endpoint; near zero means it is a candidate for deletion or serverless. |
| ModelLatency and OverheadLatency | Time spent in the model and in SageMaker, in microseconds. |
| Invocation4XXErrors and Invocation5XXErrors | Bad requests and model or container failures. |
| CPUUtilization and GPUUtilization | How busy endpoint instances are, for right-sizing. |
| Drift in prediction inputs | Feature distributions compared with training data, with MLflow and Evidently or Databricks monitoring, since Model Monitor is closed to new customers. |
| Endpoint count and age by tag | Endpoints without an owner or past their expiry. |
| Limit | Value and notes |
|---|---|
| Serverless inference payload | 4 MB per request and response |
| Serverless inference timeout | 1 minute for the container to respond |
| Asynchronous inference | Large inputs passed through S3; invocation timeout up to 3,600 seconds |
| Features closed to new customers | Model Monitor, Clarify, Ground Truth, Debugger, Augmented AI and others, announced 30 June 2026 |
Limits and quotas checked 2026-11-11; many can be raised through Service Quotas.
Decisions
Situations and verdicts
| Situation | Choice | Verdict |
|---|---|---|
| Press-failure model for six plants, data in gold | Feature table in gold, MLflow in Databricks, registered in Unity Catalog, daily batch scoring | Pursue: one place for data, model and grants. |
| Existing SageMaker endpoint for demand forecasting that works | Keep it, read features from gold through Delta Sharing, move when it needs rework | Keep if you have it: no value in breaking a working model. |
| New model monitoring setup in a new AWS account | Databricks data quality monitoring and inference tables, or open-source Evidently with MLflow | Drop SageMaker Model Monitor: closed to new customers. |
| Fine-tuning a large open-weight model on many GPUs | SageMaker training jobs or HyperPod, results registered in Unity Catalog | Pursue with care: right for scale, but keep the registry of record in one place. |
FAQ
Questions people ask
What is the difference between SageMaker and SageMaker AI?
SageMaker AI is the original ML service, renamed in December 2024. Amazon SageMaker now names a broader data and AI platform that includes SageMaker Unified Studio, a lakehouse layer, and SageMaker AI as one part.
Is SageMaker Model Monitor deprecated?
It is in maintenance: closed to new customers since mid-2026, with existing customers able to continue and no new features planned. AWS suggests open-source monitoring with MLflow and Evidently, plus CloudWatch.
Should we use SageMaker or Databricks for ML?
If the data is in Unity Catalog and the team works in Databricks, use Databricks so features, models and grants live together. Keep SageMaker where it already runs well or where you need its special training hardware.
Can SageMaker read Delta tables?
Yes, indirectly. Read them as Iceberg through UniForm, through Delta Sharing, or with Spark in a processing job. Use read-only access and avoid copying the data.
Is SageMaker Unified Studio a replacement for Databricks?
It covers similar ground: notebooks, SQL, a catalog and AI tools in one workspace. Running both as primary platforms means two catalogs and permission models, so pick one; this platform picks Databricks and Unity Catalog.
Sources
Checked against
- AWS docs: What is Amazon SageMaker AI
- AWS docs: What is the next generation of Amazon SageMaker
- AWS docs: Services in maintenance (SageMaker AI features)
- AWS docs: SageMaker Model Monitor availability change
- AWS docs: Invoke a serverless endpoint
- AWS What's New: serverless MLflow in SageMaker AI
Checked 2026-11-11.
Related
Where it connects
AWS and Databricks reference platform · Amazon Bedrock · Amazon DataZone and SageMaker Unified Studio catalog · Medallion architecture · Delta vs Iceberg · All AWS data services
← Comprehend · DataZone →