AWS Machine Learning Services: A Complete Guide to Machine Learning on AWS
Amazon offers several ways to implement machine learning, from purpose-built AI services to managed model development and infrastructure that provides direct control over compute, containers, and accelerators. The challenge is understanding which layer fits the workload and how the services connect across the ML lifecycle.
This guide puts AWS machine learning services into context by showing how the ecosystem is structured, how Amazon supports each stage from data preparation to deploying models in production, and where services such as Amazon SageMaker AI and Amazon Bedrock fit. By the end, you should have a practical framework for choosing the right approach to machine learning on AWS based on the use case, required control, operational effort, and cost.
Quick summary
- Amazon supports machine learning at several abstraction levels: purpose-built AI services handle common capabilities, Amazon SageMaker AI supports custom models, Amazon Bedrock focuses on foundation-model applications, and Amazon EC2 or Amazon EKS provide lower-level infrastructure control.
- The right approach depends on how much of the lifecycle your team needs to own. Managed capabilities reduce engineering work, while custom ML makes more sense when proprietary data, model behavior, or performance requirements demand greater control.
- For custom ML, Amazon SageMaker AI is the main managed service, while Amazon Bedrock is generally the more direct starting point for foundation-model applications.
- Budget for more than model training. Inference, data processing, storage, transfer, observability, and retraining can all become significant parts of AWS machine learning pricing.
AWS machine learning services at a glance
Before comparing individual AWS ML services, start with the outcome.
| If you want to... | Start with... | Why |
| Build and train a custom ML model | Amazon SageMaker AI | Gives you managed training, experimentation, and model lifecycle tools without building the ML platform yourself |
| Deploy a custom ML model | Amazon SageMaker AI | Supports real-time, serverless, asynchronous, and batch inference for different traffic patterns |
| Build a generative AI application | Amazon Bedrock | Provides managed access to foundation models plus RAG, guardrails, evaluation, and other GenAI capabilities |
| Build or operate AI agents | Amazon Bedrock AgentCore | Provides managed infrastructure for deploying, connecting, and operating agentic applications |
| Analyze images or video | Amazon Rekognition | Handles common computer vision tasks without requiring you to train and operate a custom model |
| Extract structured document data | Amazon Textract | Extracts text and structure such as forms, tables, layout, and signatures from documents |
| Analyze or classify text | Amazon Comprehend | Handles common NLP tasks such as entity detection, sentiment analysis, and text classification |
| Convert speech to text | Amazon Transcribe | Provides managed batch and streaming speech recognition without running your own transcription models |
| Translate text | Amazon Translate | Provides managed machine translation for supported languages and application workflows |
| Build recommendations | Amazon Personalize | Builds managed recommendation and ranking models from your interaction and item data |
| Run custom ML infrastructure | Amazon EC2, Amazon EKS, AWS Trainium | Gives you greater control over runtimes, orchestration, accelerators, and infrastructure configuration |
This table is a starting point rather than a complete Amazon machine learning services list. In practice, AWS for machine learning usually means combining ML capabilities with familiar services for storage, identity and access management, networking, databases, monitoring, and application integration.
Understanding machine learning on AWS: three layers
There is no single Amazon service for every machine learning scenario. Instead, the ecosystem falls into three broad levels of abstraction.
1. Purpose-built AI services
Amazon Rekognition, Textract, Comprehend, Transcribe, Translate, and Personalize solve defined AI problems without requiring you to manage most of the underlying machine learning implementation.
Choose this layer when an existing capability already meets your accuracy, customization, security, and latency requirements. For teams evaluating Amazon AI and ML services, it is often the shortest route to production.
2. Platforms for custom ML and foundation models
Amazon SageMaker is the broader platform for data, analytics, and AI. Amazon SageMaker AI focuses on building, training, and deploying ML models.
For custom model development, SageMaker AI is the most directly relevant managed platform. Amazon Bedrock provides managed access to foundation models plus retrieval-augmented generation (RAG), evaluation, and guardrails.
As a starting point:
- Use Amazon SageMaker AI for custom predictive machine learning and model training.
- Use Amazon Bedrock for applications built around foundation models.
- Use Amazon Bedrock with AgentCore for agentic applications that need managed agent infrastructure.
Amazon SageMaker AI and Amazon Bedrock can also be combined in the same architecture. They are compared in AWS's official Bedrock or SageMaker AI decision guide.
3. Machine learning infrastructure
For greater control, teams can use Amazon EC2, Amazon EKS, Amazon ECS, GPUs, AWS Trainium, AWS Inferentia, and related storage. This layer suits specialized frameworks, custom model servers, Kubernetes-native ML platforms, or requirements that managed services cannot meet.
📖 If containers are part of the design, our guide to Amazon Elastic Container Service covers the main operational considerations.
The machine learning lifecycle on AWS
A useful way to see how Amazon services for machine learning fit together is to map them to the machine learning lifecycle.
| Stage | Objective | Common services | Typical users |
| Collect and store | Make reliable data available | Amazon S3, Amazon Kinesis, AWS Glue, Amazon Redshift | Data/platform engineers |
| Prepare data and features | Create model-ready inputs | SageMaker Processing, Data Wrangler in SageMaker Canvas, Feature Store | Data scientists, ML engineers |
| Build and train | Develop and train models | SageMaker AI, Canvas, Autopilot, JumpStart | Data scientists, ML engineers |
| Evaluate and manage | Track, version, and automate models | MLflow with SageMaker AI, Model Registry, Pipelines | ML/MLOps engineers |
| Deploy | Serve predictions | SageMaker AI, Bedrock, Amazon EC2, Amazon EKS | ML/backend engineers |
| Operate | Monitor and update production models | Amazon CloudWatch, EventBridge, CI/CD workflows | ML/platform/SRE teams |
Prepare data for training and model predictions
Amazon S3 commonly stores training data, model artifacts, and batch inputs and outputs. AWS Glue and SageMaker Processing transform data, while SageMaker Feature Store helps keep reusable online and offline features consistent. Data Wrangler and Autopilot are available through SageMaker Canvas.
📖 For datasets with changing access patterns, Amazon S3 Intelligent-Tiering can help reduce storage costs.
Build, train, and manage models
SageMaker AI training jobs provision compute, run training, store model artifacts, and release the infrastructure when finished. Teams can use common frameworks, built-in algorithms, custom containers, Autopilot, JumpStart, or HyperPod. MLflow integrations, Model Registry, and Pipelines add experiment tracking, versioning, and automation.
⚠️ Availability note: Ground Truth, Model Monitor, Clarify, Debugger, and several other SageMaker AI features are no longer open to new customers as of July 30, 2026. Existing customers can continue using them, while other SageMaker AI features are unaffected. See the current AWS service availability guidance.
Deploy with the right model serving method
Amazon SageMaker AI provides four main inference approaches:
- Real-time inference for sustained, low-latency traffic.
- Serverless Inference for intermittent or unpredictable traffic.
- Asynchronous Inference for queued, larger, or longer-running requests.
- Batch Transform for running predictions on stored datasets without keeping an endpoint running continuously.
Some SageMaker AI real-time endpoints can scale to zero when using inference components with the appropriate managed instance scaling configuration. Scaling back from zero introduces provisioning latency and may cause requests to fail while capacity is being restored.
Example: Store transaction data in Amazon S3, prepare the data with SageMaker Processing and Feature Store, train and register the model in SageMaker AI, then make it available through a real-time endpoint so the transaction application can request predictions as needed.
A nightly inventory model could instead use Batch Transform and avoid an always-on endpoint.
📖 For production resilience, see our guide to AWS high availability architecture.
Operate and improve
After deployment, monitor latency, throughput, errors, utilization, infrastructure health, and business performance. Also evaluate whether the model continues to meet quality and business requirements as data and usage patterns change. Changes in input data, prediction quality, or business outcomes can signal that the model needs review or retraining. Based on those results, teams may need to update features, retrain the model, validate a new version, and redeploy it. For new environments, build observability around Amazon CloudWatch, available evaluation tools, and application-specific metrics.
AWS machine learning services by category
Managed AI services
Purpose-built services are useful when the problem maps to an existing capability and custom model development would add little value.
| Service | What it does | Typical use cases | Use it when | Avoid it when | Related services |
| Amazon Rekognition | Image and video analysis | Detection, moderation, text in images | Standard computer vision meets the requirement | You need a specialized model or training control | Amazon S3, AWS Lambda, SageMaker AI |
| Amazon Textract | Extracts text and document structure | Invoices, claims, forms, onboarding | Document structure matters, not just OCR | The main task is open-ended reasoning | Amazon S3, Comprehend, Bedrock |
| Amazon Comprehend | Managed NLP and classification | Entity detection, sentiment analysis, text routing, and classification | You need to identify, classify, or analyze information in text without generating new content | You need open-ended text generation or reasoning | Amazon S3, Textract, Bedrock |
| Amazon Transcribe | Batch and streaming speech-to-text | Calls, subtitles, searchable audio | You need reliable speech-to-text without building or operating your own transcription models | You need specialized model control | Amazon S3, Comprehend, Bedrock |
| Amazon Translate | Managed machine translation | Application content, multilingual workflows | Supported languages and quality are sufficient | High-risk content requires specialist review | Amazon S3, AWS Lambda |
| Amazon Personalize | Managed recommendation capabilities | Product recommendations, personalized rankings, and content discovery | You want personalized recommendations without building and operating your own recommendation system | You need full control over how recommendations are ranked or generated | Amazon S3, Amazon Kinesis, SageMaker AI |
Example: Use Amazon Textract to extract document content, then Amazon Comprehend for structured NLP or Amazon Bedrock for summarization and open-ended reasoning before passing the result to the application.
AWS machine learning platforms: Amazon SageMaker AI vs. Amazon Bedrock
Amazon SageMaker AI for custom ML
Amazon SageMaker AI supports custom model development, training, deployment, and lifecycle workflows. Typical workloads include classification, regression, forecasting, anomaly detection, recommendations, computer vision, and deep learning.
Use SageMaker AI when your data and model behavior justify owning the development lifecycle.
Example: Store account data in Amazon S3, prepare it with SageMaker Processing, train and register the model in SageMaker AI, then deploy it to a real-time endpoint.
A SaaS application could use the endpoint to request a churn-risk score before triggering a retention workflow.
Amazon Bedrock for foundation-model applications
Amazon Bedrock provides managed access to foundation models together with RAG, evaluation, guardrails, and other generative AI capabilities. Typical uses include assistants, document summarization, semantic search, content generation, and RAG.
Example: Store internal documents in Amazon S3, connect them to Knowledge Bases for Amazon Bedrock, use a foundation model to generate grounded responses, and deliver them through the internal assistant.
💡 In practice: Stormit used Amazon Bedrock in a coaching insights solution built around structured data. The architecture avoided an unnecessary RAG layer, while optimizing Bedrock Guardrails helped bring processing costs down to approximately $0.014–$0.017 per analysis.
For new agentic applications, Amazon Bedrock AgentCore provides managed capabilities for building, connecting, and operating agents.
Machine learning infrastructure for greater control
Lower-level AWS machine learning tools include Amazon EC2, Amazon EKS, Amazon ECS, GPUs, AWS Trainium, AWS Inferentia, and specialized storage. Use this layer when you need a specific runtime, Kubernetes-native operations, specialized frameworks, accelerator configurations, or infrastructure controls unavailable through a managed platform. The trade-off is more responsibility for scaling, resilience, observability, upgrades, and cost.
How to choose the right AWS ML service
Service selection becomes much easier when you start with the workload rather than comparing feature lists one service at a time.
1. Can a managed service already solve the problem?
If Amazon Rekognition, Textract, Comprehend, Transcribe, Translate, or Personalize meets the requirement, a custom ML pipeline may add unnecessary work.
2. Is the workload predictive ML or generative AI?
For classification, regression, forecasting, ranking, anomaly detection, and similar custom workloads, evaluate Amazon SageMaker AI first.
For foundation-model applications, RAG, summarization, and generation, evaluate Amazon Bedrock first.
For agentic applications, also evaluate Amazon Bedrock AgentCore.
3. What are the data requirements?
Consider the type, volume, velocity, and location of the data. Structured tabular data, documents, images, audio, and streaming data may point to different services and processing patterns. Also consider whether the data needs labeling, frequent updates, or additional security and governance controls.
4. How much control do you need?
Think of the options as a spectrum: managed AI services for the least operational overhead, Amazon Bedrock or Amazon SageMaker AI for more control, and Amazon EC2 or container-based infrastructure for the most flexibility and responsibility.
5. What does production model serving require?
Define latency, throughput, model and payload size, traffic predictability, batch versus online processing, and cold-start tolerance. These requirements usually determine the serving method.
6. Can the team operate it reliably?
Factor in ML/MLOps expertise, security, observability, compliance, availability, support burden, and budget. Choose a solution the team can operate reliably in production.
Common machine learning use cases on AWS
| Business problem | Practical approach |
| Predictive maintenance | Store telemetry in Amazon S3, prepare it with SageMaker Processing, and train a model in SageMaker AI for the maintenance application. |
| Fraud detection | Use Amazon S3 and Feature Store for transaction data, then deploy a SageMaker AI model to a real-time endpoint for fraud scoring. |
| Demand forecasting | Store historical sales in Amazon S3, prepare the data with SageMaker Processing or Canvas, then generate forecasts with SageMaker AI and Batch Transform. |
| Document intelligence | Store documents in Amazon S3, extract data with Amazon Textract, then analyze it with Amazon Comprehend or Amazon Bedrock. |
| Recommendations | Feed interaction data into Amazon Personalize and return recommendations to the application. |
| Computer vision | Use Amazon Rekognition for standard computer vision tasks and SageMaker AI when a custom model is required. |
| Generative AI / RAG | Store source documents in Amazon S3, connect them to Knowledge Bases for Amazon Bedrock, and use a foundation model to answer application queries. |
Some older services should not be chosen for new architectures. Amazon Forecast is closed to new customers, while Amazon Fraud Detector and Amazon Lookout for Equipment reach end of support on October 7, 2026.
AWS machine learning pricing: what drives the bill
There is no single price for machine learning on AWS. Costs depend on the services, data flow, traffic pattern, and operating model.
Managed service usage
Purpose-built services charge according to service-specific units such as images, pages, audio duration, characters, or requests. Amazon Bedrock pricing varies by model, usage mode, input and output, and additional capabilities.
Managed services can work well for variable demand because there is no dedicated model infrastructure to keep running. However, each service has its own pricing and capacity model, so "managed" does not necessarily mean that all infrastructure-related costs disappear.
Training and model serving
SageMaker AI training cost depends mainly on instance type, count, duration, and experiment frequency. Managed Spot Training can reduce eligible training compute costs by up to 90%, while SageMaker AI Savings Plans can reduce eligible instance usage by up to 64% with a one- or three-year commitment.
Model serving can become the larger recurring cost, so match the deployment model to measured demand rather than defaulting to an always-on endpoint.
Storage, processing, and data transfer
Include datasets, features, model artifacts, logs, predictions, processing jobs, and supporting infrastructure.
Data transfer can become a hidden cost when services span Regions, Availability Zones, or external environments. Our guide to AWS data transfer pricing explains where these charges commonly appear.
Beyond transfer, the same FinOps principles apply to ML workloads, including rightsizing, usage review, and commitment planning. See our guide to AWS cloud cost optimization for broader ways to control cloud spend.
Data labeling and human review
Supervised ML may also require annotation, specialist review, and quality assurance, adding costs outside the core model infrastructure.
How to estimate AWS ML costs
- Map expected data volumes and movement.
- Estimate training frequency, compute, and duration.
- Benchmark serving capacity rather than guessing.
- Add storage, processing, logging, orchestration, and any labeling or human review.
- Model retraining and traffic growth, then apply commitments only after usage patterns are understood.
AWS machine learning services cheat sheet
| Need | Service | Best for | Complexity |
| Custom ML development | Amazon SageMaker AI | Custom model workflows | Medium |
| Foundation-model applications | Amazon Bedrock | Generative AI and RAG | Low–Medium |
| AI agents | Amazon Bedrock AgentCore | Agent deployment and operations | Medium |
| Image/video analysis | Amazon Rekognition | Managed computer vision | Low |
| Document extraction | Amazon Textract | Structured documents | Low |
| Text analytics | Amazon Comprehend | NLP and classification | Low |
| Speech recognition | Amazon Transcribe | Speech-to-text | Low |
| Translation | Amazon Translate | Machine translation | Low |
| Recommendations | Amazon Personalize | Managed personalization | Low–Medium |
| Custom ML compute | Amazon EC2 | Direct compute control | High |
| Containerized ML | Amazon EKS / Amazon ECS | Custom orchestration | High |
| Data and model artifacts | Amazon S3 | Training data and model storage | Low |
| Reusable model features | Amazon SageMaker Feature Store | Shared features | Medium |
FAQs
The main categories are purpose-built AI services; Amazon SageMaker AI for custom ML; Amazon Bedrock for foundation-model applications; Amazon Bedrock AgentCore for agents; and infrastructure such as Amazon EC2 and Amazon EKS.
Machine learning on AWS can cover the lifecycle from data preparation and training to deployment and operations. Teams can use managed services or combine SageMaker AI, Bedrock, storage, compute, networking, and monitoring.
Amazon SageMaker is the broader platform for data, analytics, and AI. Amazon SageMaker AI is the managed ML service for building, training, and deploying models.
Managed AI services solve defined problems while Amazon handles most of the underlying ML implementation. A platform such as SageMaker AI provides greater control over model development, training, deployment, and lifecycle workflows.
Costs can include managed-service consumption, training, model serving, processing, storage, data transfer, observability, and supporting infrastructure. Estimate the complete production architecture rather than training alone.
Conclusion: choose the ML layer before the service
Amazon provides several routes to production ML, so start with the highest-level capability that meets the requirement. Use purpose-built services for common AI tasks, SageMaker AI when custom models and proprietary data matter, Bedrock for foundation-model applications, and lower-level infrastructure only when you need additional control.
Above all, decide how much of the ML lifecycle your team needs to own and choose the architecture accordingly.
If you are evaluating AWS machine learning solutions for a new workload, Stormit can help you choose the right cloud approach and design an architecture around performance, security, scalability, and cost.
With extensive experience as a developer, Roman specializes in serverless technologies and containers. His AWS expertise extends to backend and frontend development, where he excels in designing and implementing scalable solutions that leverage cloud services efficiently. Roman also demonstrates a deep understanding of infrastructure as code (IaC), employing tools and practices to automate and manage infrastructure effectively.