Amazon SageMaker: AWS's machine-learning platform, explained.

IWhat Amazon SageMaker is

SageMaker AI is a managed service for the whole machine-learning lifecycle on AWS (build, train, tune, deploy and operate) without provisioning, patching or scaling servers yourself.

In one line: SageMaker AI takes a model from a blank notebook to a monitored production endpoint, with AWS handling the infrastructure at each step. A raw EC2 GPU instance is a bare machine you configure and look after. SageMaker gives you managed pieces (a training job, an endpoint, a pipeline) that start the right compute, run your code and shut it down.

It is broad on purpose: about a dozen capabilities under one IAM model, one bill and one console. A team can experiment, train at scale, govern models and serve them without stitching together separate tools.

It works with any framework: PyTorch, TensorFlow, JAX, Hugging Face, XGBoost, scikit-learn and custom containers. It is used for tabular ML (fraud, churn, forecasting), deep learning (vision, recommendations) and training or fine-tuning large models.

A simple mental model: Bedrock is AI as an API call; SageMaker is the full ML factory. Section V covers the difference.

IIThe key components

Each component maps to a stage of ML work. You will not use every one on every project.

• Studio: The web IDE: notebooks, a code editor, experiment tracking and access to training and deployment.
• JumpStart: A hub of pretrained open and proprietary models and solution templates you can deploy or fine-tune in a few clicks.
• Training jobs: Managed compute that runs your training code on the instances you choose, writes the model to S3 and shuts down. Supports distributed training, Spot and warm pools.
• Automatic Model Tuning: Runs many training jobs across a search space to find the best hyperparameters.
• Inference endpoints: Four ways to serve predictions: real-time, serverless, asynchronous and batch transform (section IV).
• Pipelines: Workflow orchestration for ML: preprocess, train, evaluate, register and deploy as a repeatable, versioned flow.
• Feature Store: One place for ML features, with an online store for serving and an offline store for training, so both use the same definitions.
• Model Registry: Versioned models with approval status and lineage, so only approved versions are deployed.
• Ground Truth and Data Wrangler: Managed data labeling, and visual data preparation for tabular data.
• Clarify and Model Monitor: Bias and explainability checks before deployment, and drift monitoring after it.

IIIThe machine-learning lifecycle on SageMaker

The components map onto seven repeatable stages.

• 1. Prepare and label data: Data lands in S3. Ground Truth labels it, Data Wrangler cleans it and Feature Store records the features. Most real ML time is spent here.
• 2. Experiment in Studio: Try a JumpStart model or your own code on a sample, with each run tracked so results are comparable.
• 3. Train at scale: Move from the notebook to a managed training job on the right instance; tune hyperparameters automatically. You pay for the seconds the cluster exists.
• 4. Evaluate and check bias: Test against a hold-out set; Clarify measures bias. A pipeline can block models that miss the bar.
• 5. Register and govern: Version approved models in the Model Registry with lineage, and approve before deployment.
• 6. Deploy: Serve the model on the endpoint type that fits the traffic.
• 7. Monitor and retrain: Model Monitor watches for drift; a pipeline can retrain automatically. That loop is what MLOps means in practice.

IVThe four ways to serve a model

Inference is where most production cost and latency sit. Two questions decide the mode: how predictable is the traffic, and how fast must each answer come back?

VSageMaker vs Amazon Bedrock

Not either/or. They sit at different points between control and convenience, and many teams run both.

Amazon Bedrock is a managed API to existing foundation models: Claude, Llama, Mistral, Amazon Nova, Cohere and others, through one interface. You never see an instance; you pay per token. Bedrock adds Agents, Knowledge Bases (managed RAG), Guardrails and fine-tuning on top.

SageMaker AI gives you the full ML stack and full control: your own architectures, any framework, the instance type, the container, the training regime and the scaling. With that control comes more work: you own instance choice and operations.

The deciding question: does a foundation model that already does what you need exist on Bedrock? If yes (a chat assistant, a summarizer, RAG over your documents), start with Bedrock. If no (a proprietary model, tabular ML, deep fine-tuning, strict control of serving), use SageMaker.

A common setup uses both: Bedrock for generative-AI features and SageMaker for proprietary models such as a recommendation engine or a demand forecaster.

VIWhat drives the SageMaker bill

No licence fee. You pay for the compute, storage and features you use, with compute billed per second.

Two lines dominate: training compute (spikes, then stops) and inference compute (runs all the time on real-time endpoints). GPU choice is the biggest lever: at AWS list prices for us-east-1 (checked October 10, 2026), an ml.m5.large costs $0.115 an hour, an ml.g5.xlarge $1.408 and an ml.p5.48xlarge $63.30.

Endpoint mode is the other lever. One ml.g5.xlarge real-time endpoint left on for a month costs about $1,028; scoring the same model nightly with batch transform costs about $42.

SageMaker Savings Plans cut compute cost by up to 64% for a 1- or 3-year commitment, and managed Spot lowers training cost for interruptible jobs. For full rates, worked examples and the levers, see the SageMaker pricing breakdown.

VIIWho SageMaker is for

SageMaker is for teams that own models, not just teams that call them.

• A strong fit if you train or fine-tune your own models: Custom architectures, proprietary models, deep fine-tuning of open-weights models and tabular ML.
• A strong fit if you need MLOps at scale: Many models in production that need versioning, lineage, retraining, drift monitoring and bias checks.
• A strong fit if you need control over serving: Specific instances, custom containers, scaling rules or AWS silicon (Trainium, Inferentia) to cut cost.
• Probably overkill if you only call a foundation model: For a chat assistant, summarizer or RAG app over an existing model, Bedrock is simpler and faster.
• Not the right layer for a non-technical team: SageMaker assumes ML and engineering skills. A business team that wants a ready-made assistant over company data should look at Amazon Quick (Amazon Q Business is closed to new customers).

VIIIGetting started, and who files which AWS funding

Going from zero to a deployed model is a short, well-known path.

• 1. Set up a domain: Create a SageMaker domain and a user profile with an IAM role that can read your S3 data and write artifacts. One-time setup.
• 2. Open Studio and load data: Open a notebook and point it at your data in S3. JumpStart gives you a working model in a few clicks to see the full flow.
• 3. Run a training job: Start small (one GPU or even CPU), with the framework container and data location. The Free Tier covers some first-month usage.
• 4. Deploy an endpoint: For a first deployment, serverless inference is low-risk: it scales to zero, so a forgotten test endpoint does not run up cost.
• 5. Add governance as you grow: Wrap it in a pipeline, register the model and add Model Monitor.

Partner-filed funding (the bigger route for most startups). A vetted AWS consulting partner can file proof of concept (POC) funding to test an ML workload, migration (MAP) funding if you are moving ML work to AWS from another cloud, or incremental workload (IW) funding if you already run on AWS and are launching new work. For startups this is often $50K to $150K. Depends on the workload; AWS approves. The partner files it, not you.

AWS Activate credits (you apply yourself). Activate Founders is $1,000 to start, up to $5,000; you apply on the AWS site. Activate Portfolio is up to $200,000 with the Org ID from a VC, accelerator or angel group that is an Activate Provider. A consulting partner cannot file Activate for you.

CloudRoute matches you with one vetted AWS partner who can build on SageMaker (training, deployment, MLOps) and file the partner funding when the work qualifies. Startups pay CloudRoute nothing. Partners pay CloudRoute a monthly membership plus 10% of AWS-approved funding.