AWS Machine Learning Landscape

AWS offers machine learning at three levels of abstraction. Choosing the right one is mostly a question of who is building and how much control the problem needs.

1. AI services

Pre-trained, API-driven services that solve a specific task.

• Aimed at application developers
• No machine learning knowledge required
• Opinionated: they solve the task their way, and the model is not yours to change
• Pay per use, with no infrastructure to manage
• Reached through the AWS SDKs like any other service

This is the right level whenever the task is a standard one — recognising objects in a photo, transcribing audio, extracting form fields, detecting sentiment. See AWS AI services for the catalogue and for the services that have since closed to new customers.

2. Amazon SageMaker AI

Managed infrastructure and tooling for building, training and deploying your own models.

• Aimed at machine learning engineers and data scientists
• Provisions managed compute for training and hosting without the caller running servers
• Supplies built-in algorithms, and accepts custom algorithms and containers
• Covers the lifecycle: data preparation, training, hyperparameter tuning, deployment, monitoring

This is the right level when the task is specific to your data and no pre-trained service covers it. See SageMaker AI.

Note the naming: since 3 December 2024 the model-building service is Amazon SageMaker AI, and “Amazon SageMaker” now names a larger platform spanning data, analytics and AI, of which SageMaker AI is one component.

3. Frameworks and infrastructure

Raw compute with the machine learning stack installed, for teams that want to control the framework and the training loop themselves.

• Aimed at researchers and teams with existing framework-level code
• AWS Deep Learning AMIs and Deep Learning Containers — EC2 images and container images preloaded with PyTorch, TensorFlow and their GPU dependencies
• Accelerated instance families — GPU instances, and AWS Trainium and Inferentia for training and inference on AWS silicon
• AWS IoT Greengrass — running inference at the edge, on devices, where latency or connectivity rules out a round trip to the cloud

Characteristics shared by the AI services tier

• Simple API integration, no model training
• Scalable and highly available by default
• Pay-per-use
• Confidence scores returned with predictions, so the caller can set its own thresholds
• Console demo environments for evaluating a service before writing code

Choosing between the tiers

Work down, not up. Start at the AI services tier and move to SageMaker AI only when a pre-trained service demonstrably cannot do the job — the cost of a custom model is not the training run, it is owning the data pipeline, the retraining cadence and the evaluation for as long as the model is in production. Drop to frameworks and infrastructure only when SageMaker AI’s abstractions get in the way.