What Are Foundation Models in Amazon Bedrock?

Foundation models are the main building blocks behind generative AI. In AWS-based AI systems, Amazon Bedrock provides a managed way to consume them at scale. In the AWS Certified Generative AI Developer – Professional (AIP-C01) exam, foundation models are rarely tested as isolated concepts. Instead, they appear embedded within architectural scenarios that require careful reasoning about design decisions, trade-offs, and constraints, with clear expectations around cost, speed, accuracy, and reliability.

For the AWS Certified Generative AI Developer Professional exam, candidates should approach foundation models the same way they approach core AWS services, such as databases, compute, and networking. This means clearly understanding what these models are capable of, their limitations, how they perform, and the scenarios in which they are most effective to use.

Foundation models and their role in Amazon Bedrock

Foundation models are large, pretrained models that learn general patterns from massive datasets and can be adapted to many downstream tasks without retraining from scratch. These tasks include text generation, summarization, classification, question answering, and embedding creation. In practice, foundation models act as reusable cognitive engines that applications can invoke on demand, rather than assets that each team must train and maintain independently.

Amazon Bedrock abstracts access to these models through a fully managed service. AWS manages the underlying infrastructure and operations for the service, handling provisioning, scaling, availability, and security. This abstraction allows teams to focus on application design instead of model hosting. Teams, however, are still responsible for securing their application data and access.

Think of it like a secure, managed kitchen (AWS) versus the recipe and ingredients you bring (your team). While AWS maintains the kitchen itself, you are responsible for safeguarding your secret recipes (data) and deciding who can use them (access).

For the exam, this distinction is critical because model training and development are explicitly out of scope. The candidate is evaluated as a GenAI integrator who selects, configures, and orchestrates models within an AWS architecture, not as a machine learning researcher.

To understand our operational responsibilities when working with foundation models on AWS, it’s important to know how to choose an appropriate model for a given task, configure ...