AWS Architecture Blog·November 19, 2025Architecting Generative AI Workloads on AWS: Well-Architected Framework Update
This article announces an update to the AWS Well-Architected Generative AI Lens, providing best practices and guidance for designing and operating generative AI workloads on AWS. It covers strategic considerations for data architecture, responsible AI, and agentic systems, offering scenarios for common business applications and emphasizing the application of the Well-Architected Framework's six pillars to AI solutions.
Read original on AWS Architecture BlogThe AWS Well-Architected Framework is a critical resource for system designers aiming to build robust, secure, efficient, and cost-effective cloud applications. Its Generative AI Lens provides specific architectural best practices tailored for generative AI workloads, helping to evaluate architectures leveraging large language models (LLMs) to achieve business objectives while adhering to core design principles.
Key Updates and Architectural Considerations
- <p><strong>Amazon SageMaker HyperPod Guidance:</strong> The updated lens includes guidance for using SageMaker HyperPod for resilient model training and high-scale inference, integrating it with existing services like Amazon Bedrock and Amazon SageMaker AI. This highlights the importance of choosing appropriate infrastructure for different stages of the AI lifecycle.</p>
- <p><strong>Responsible AI Preamble:</strong> Emphasizes the eight core dimensions of responsible AI, which are crucial for architecting ethical and compliant generative AI systems. This moves beyond pure technical implementation to consider societal and governance aspects.</p>
- <p><strong>Data Architecture Preamble:</strong> Focuses on strategic considerations for modern data architectures supporting generative AI workloads. This section is vital for understanding how to design data ingestion, storage, and processing pipelines that can feed and be influenced by LLMs effectively.</p>
- <p><strong>Agentic AI Preamble:</strong> Introduces architectural paradigms for agentic systems powered by foundation models. These systems, a subset of distributed computing, require careful design for their interactions, state management, and orchestration.</p>
Design Principle for Generative AI
When designing generative AI workloads, consider not only the model itself but also the surrounding data pipelines, responsible AI implications, and how agentic systems might orchestrate complex, multi-step tasks. Performance efficiency and cost optimization often hinge on smart choices for model selection, prompt engineering, and workload integration.
Generative AI Scenarios for Business Applications
The lens now includes eight architectural scenarios covering common generative AI business applications such as autonomous call centers, knowledge worker co-pilots, and multi-tenant generative AI service systems. These scenarios provide practical guidance on applying generative AI technologies to real-world problems, offering insights into integration patterns and operational considerations. System designers can leverage these examples to understand various deployment models and trade-offs.
This update reinforces the necessity of a holistic approach to building AI systems, extending the traditional Well-Architected pillars to the unique challenges and opportunities presented by generative AI. It's a valuable resource for architects designing new AI-powered applications or evaluating existing ones for robustness and efficiency.