AWS machine learning services now cover everything from calling a foundation model through an API to training your own models on custom AI chips. That range is useful, but it also makes the first question harder: which service should you actually use?

The answer has shifted a lot since early 2025. Amazon Bedrock and AI agents are now the starting point for most new projects. SageMaker has been reorganized and renamed. Some older services have closed to new customers, and AWS's own chips have become a serious alternative to GPUs. This guide sorts AWS's AI and ML services into four layers, explains what each service is for, and ends with a simple way to choose.

The AWS AI and ML stack at a glance

The original version of this article listed seven items side by side, mixing managed services, machine images and open-source frameworks. It's clearer to think in four layers, each answering a different question:

  1. Generative AI and agents: use foundation models and build agents without managing models or servers (Amazon Bedrock, Bedrock AgentCore).
  2. The ML platform: build, train and deploy your own models (Amazon SageMaker AI and the unified Amazon SageMaker).
  3. Pre-built AI services: call an API for a specific task such as image analysis or text extraction, with no ML expertise needed (Rekognition, Comprehend, Textract and others).
  4. Infrastructure: run your own frameworks on AWS chips and GPUs (Trainium, Inferentia, EC2 GPU instances, Deep Learning AMIs and Containers, EKS).

Pre-built services and Bedrock get you to production fastest; the platform and infrastructure layers give you more control and, at scale, lower unit costs. Most real systems combine layers: for example, a Bedrock-powered assistant that also calls Textract and a custom SageMaker model.

 

#1: Amazon SageMaker

SageMaker is Amazon’s fully managed ML platform that simplifies the entire ML development lifecycle. It simplifies interactive development environments for data exploration and model training and includes scalable training of ML models and the deployment of models as real-time or batch endpoints. SageMaker also offers centralized storage and management of ML models, as well as continuous tracking of model performance and drift detection. 

Some of the other features include: 

  • Built-in Algorithms: Access to a wide range of built-in algorithms for common ML tasks like classification, regression, and clustering.
  • Hyperparameter Tuning: Automatically optimize hyperparameters to improve model performance.
  • Distributed Training: Efficiently train large models across multiple machines.
  • Model Interpretability: Understand how models make decisions and identify potential biases.

By taking away the complexities of infrastructure management, SageMaker enables developers to focus on building and refining ML models.

Next, let’s take a look at some of the text and image processing utilities in AWS machine learning services. 

#2: Amazon Comprehend

Amazon Comprehend is a sophisticated natural language processing (NLP) service designed to extract meaningful insights from text data.

Here are some of the key tasks it performs, along with real-world examples:

  • Sentiment Analysis

Comprehend can determine whether the sentiment in a text is positive, negative, or neutral. 

Here’s the use case: Businesses use it to analyze customer reviews or social media posts to understand public perception of their products or services.

  • Topic Modeling

This helps identify key themes within large volumes of text. For instance, news agencies use Comprehend to automatically categorize articles into topics like politics, sports, or entertainment, making it easier to organize content.

  • Entity Recognition

Comprehend recognizes and classifies named entities like people, organizations, and locations in the text. 

Example: a customer service company might use this to extract relevant details from emails or support tickets, such as customer names and locations, to improve service efficiency.

  • Key Phrase Extraction

It extracts important keywords or phrases from text, helping to summarize documents or improve SEO. 

For example, an e-commerce site can use this feature to identify key product features in customer reviews to improve search results and product descriptions.

#3: Amazon Rekognition

Amazon Rekognition is a powerful computer vision service that uses advanced deep learning techniques to analyze images and videos. 

By applying sophisticated algorithms to visual data, Rekognition can perform a wide range of tasks, including:

  • Object and Scene Detection: It can identify and categorize objects or scenes in images and videos. 

Uses: It’s used in visual search engines or for content moderation to ensure inappropriate images are flagged.

  • Face Analysis: Rekognition detects and analyzes faces, including identifying features like age, gender, and emotions. 

Uses: This is commonly used in security systems for access control or surveillance to compare and recognize faces in real-time.

  • Text in Image: Rekognition extracts text from images, even in challenging scenarios like handwritten text or text on top of pictures. 

Uses: This feature is used in document scanning, converting handwritten notes into editable text, or extracting information from images for further processing.

So, ML applications like the ones listed above can take up a significant amount of compute power on your instances and can be tricky to configure. 

Why not use a virtual machine to help with some of the load? Learn about Amazon’s ML-oriented AMIs (Amazon Machine Images) below. 

#4: Amazon Deep Learning AMIs

Amazon Deep Learning AMIs provide pre-configured virtual machine images optimized for deep learning workloads. These AMIs come pre-installed with popular deep learning frameworks like TensorFlow, PyTorch, MXNet, and Apache MXNet, along with the necessary libraries and tools. 

The ultimate benefit? It eliminates the need for manual setup and configuration, allowing data scientists and ML engineers to quickly start building and training deep learning models.

A few key features of Deep Learning AMIs include:

  • Faster Time to Market: Pre-configured environments reduce setup time and accelerate the development process.
  • Enhanced Performance: Optimized hardware and software configurations deliver high performance for training and inference.
  • Simplified Management: Managed infrastructure and automatic updates to streamline operations.

This is a robust list of tools that are developed by AWS, but there are also some external ML utilities that come included with AWS. 

Let’s take a look at a few of these now! 

#5: Apache MXNet

Apache MXNet on AWS is a highly scalable and flexible deep learning framework that supports a wide range of deep learning models, including:

  • Convolutional neural networks (CNNs)
  • Recurrent neural networks (RNNs)
  • Long short-term memory (LSTM) networks. 

It is known for its efficiency, scalability, and ease of use.  Some features of Apache MXNet include:

  • Hybrid Programming Model: Supports both imperative and symbolic programming styles, providing flexibility for different use cases.
  • Efficient Distributed Training: Enables efficient training of large models on multiple GPUs and machines.
  • Strong Community Support: A large and active community provides extensive documentation, tutorials, and support.
  • Language Agnostic: Supports multiple programming languages, including Python, R, Julia, Scala, and C++.

#6: PyTorch on Amazon

PyTorch is a popular open-source ML framework primarily used for deep learning applications. It's notable for its flexibility, ease of use, and strong community support. PyTorch can be run on AWS to simplify the process of building, training, and deploying ML models. It provides a seamless integration with the AWS ecosystem, offering a range of features and benefits for ML practitioners.

This is especially beneficial for a few reasons: 

  • Managed Infrastructure: AWS SageMaker handles the underlying infrastructure for PyTorch, allowing you to focus on model development and training.
  • Scalable Training: Easily scale your training jobs to multiple GPUs or instances to accelerate the training process.
  • Distributed Training: Efficiently train large models across multiple machines.
  • Hyperparameter Tuning: Automatically tune hyperparameters to optimize model performance.
  • Model Deployment: Deploy trained models as real-time or batch endpoints.
  • Model Monitoring: Monitor model performance over time and detect potential issues.
  • Integrated Tools: Access a range of tools like Jupyter Notebooks, debugging tools, and model visualization.

To use these utilities mentioned above, consider using AWS Deep Learning Containers that make short work of your containerization. 

#7: AWS Deep Learning Containers

Similar to the Deep Learning AMI, AWS Deep Learning Containers provide pre-built Docker images with popular deep learning frameworks. These containers come pre-installed with the latest versions of frameworks, libraries, and tools, saving you time and effort in setting up development environments.

There are some advantages to using these: 

  • Optimized Performance

Containers are designed for high performance, often leveraging hardware acceleration with GPUs or specialized hardware like TPUs. 

For example, a healthcare company used AWS Deep Learning Containers to speed up their medical image processing, achieving faster training times for deep learning models that detect diseases from X-rays.

  • Consistent Environments

Deep Learning Containers ensure a consistent development and deployment environment, reducing the risk of compatibility issues. Teams can be confident that the code runs the same way in both development and production environments.

  • Portability

These containers can be easily deployed across multiple platforms, including cloud environments like AWS, GCP, and Azure. This flexibility allows you to quickly move applications between different cloud providers based on your needs or cost preferences.

The bottom line: AWS offers a wide range of machine learning services that cater to different use cases, from training models to deploying scalable applications. Let’s wrap it up.

Running ML in production: MLOps, security and cost

Choosing a service is the easy part. Most ML projects stall between a working prototype and a reliable production system. Four practices make the difference:

  1. MLOps pipelines: version data, code and models together; automate training, evaluation and deployment with SageMaker Pipelines or a GitOps workflow on EKS; and monitor for drift.
  2. Evaluation and guardrails for generative AI: test prompts and agents against a fixed set of examples before each change, and apply Bedrock Guardrails and AgentCore Policy in production.
  3. Security and data governance: keep traffic private with VPC endpoints, encrypt with KMS, control access with least-privilege IAM, and log model and agent activity.
  4. Cost control: GPU and token costs grow quickly. Use batch inference where latency allows, right-size endpoints, consider Trainium or Inferentia for steady workloads, and tag every resource so you can see cost per model or feature.

How SquareOps helps

SquareOps builds and runs the infrastructure behind AI products on AWS: Bedrock and AgentCore deployments, SageMaker AI pipelines, and GPU and Trainium workloads on Amazon EKS with Karpenter. Our MLOps and AI infrastructure team covers architecture, security, observability and cost, and our 24x7 SRE team keeps it running. [VERIFY: replace the original's "over 20 certifications" and "24/7 uptime during your migration" with current, checkable credentials.]

Conclusion

AWS's AI and ML services make the most sense as layers. Start as high up the stack as your problem allows: a pre-built API or a Bedrock model gets you to production fastest. Move down to SageMaker AI or your own infrastructure only when you need a custom model, more control or lower unit costs at scale. Whichever layer you choose, plan for MLOps, security and cost from the first prototype.

Planning an AI or ML project on AWS? Talk to SquareOps about the right architecture.