AWS re:Invent 2026, held from 30 November to 4 December, announced a programme of more than 140 AI and machine‑learning sessions covering AI chips, model training, inference, agent development and sustainability.
Sessions are delivered in a mix of breakouts, chalk talks, builder sessions, workshops, code talks and lightning talks. Breakout recordings are posted after the event, allowing on‑demand access.
AI Chips and Training Platforms
AWS Trainium, a purpose‑built AI chip family, is highlighted for high‑performance training and inference at lower cost than GPU alternatives. Trainium runs on the AWS Neuron stack, which supports integration with PyTorch and other familiar frameworks without requiring code rewrites. NVIDIA GPUs remain available via Amazon EC2 instances, with optimised scheduling and performance for AI workloads.
Model Lifecycle and Agent Services
Amazon SageMaker AI is promoted as a full‑lifecycle platform that handles data preparation, training, fine‑tuning and deployment on endpoints or HyperPod. Amazon Bedrock offers managed access to hundreds of leading AI models from multiple providers through a unified API, while Bedrock AgentCore enables building, deploying and managing AI agents that select the best model for each task based on a provider mix. Ready‑to‑use agentic solutions such as Kiro, AWS Quick, Amazon Connect and partner‑built AI solutions on AWS Marketplace are also showcased.
Hands‑on workshops cover training pipelines on SageMaker HyperPod, world‑model deployment, and physical AI model factories using SageMaker HyperPod. Specific sessions address deployment of world models, vision‑language‑action models and agentic workflows using PyTorch and Neuron workflows.
Additional tracks discuss data governance, cost optimisation, business impact of AI, ROI measurement and prompt caching strategies, providing a comprehensive view of the AI stack from infrastructure to business outcomes.