Week 12 - AWS re:Invent 2025 Announcements

Week: 2025-11-24 to 2025-11-28
Status: “Planned”

Week 12 Overview

Highlights from AWS re:Invent 2025: new Nova foundation models (speech-to-speech, multimodal, cost-effective reasoning), Bedrock expansions (open-weight models, reinforcement fine-tuning), vector and AI infra updates (S3 Vectors GA), and compute launches like Graviton5 CPUs and Trainium3 UltraServers. Focus on mapping announcements to practical adoption plans.

Key Topics

GenAI & Models

  • Nova 2 family: Sonic (speech-to-speech), Lite (fast/cost-efficient), Omni (multimodal), Forge program for custom frontier models
  • Nova Act for reliable UI agents; Bedrock AgentCore adds policy controls and quality evals
  • Bedrock adds open-weight models (Mistral Large 3, Ministral 3) and reinforcement fine-tuning

Vector & Data

  • Amazon S3 Vectors GA: up to 2B vectors/index, ~100ms queries, lower cost vs. specialty DBs
  • Clean Rooms synthetic data for privacy-preserving ML collaboration

AI Dev Platform

  • SageMaker AI serverless MLflow and new training features (checkpointless, elastic scaling)

Compute & Hardware

  • Graviton5 CPUs for better price/perf on EC2
  • Trainium3 UltraServers (3nm) for faster, cheaper training/inference

Learning Objectives

  • Identify which re:Invent AI/compute launches impact current workloads
  • Plan pilot use cases for Nova models and Bedrock new capabilities
  • Outline migration path to S3 Vectors for vector search storage
  • Evaluate Graviton5/Trainium3 fit for cost/performance gains

Daily Breakdown

Day Focus Topics
56 Model Announcements Nova 2 variants (Sonic, Lite, Omni), Forge program, Nova Act agents
57 Bedrock & Agents Open-weight additions, reinforcement fine-tuning, AgentCore policy/quality
58 Vector & Data S3 Vectors GA, Clean Rooms synthetic data, vector scale/cost planning
59 SageMaker Platform Serverless MLflow, checkpointless & elastic training on HyperPod
60 Compute Launches Graviton5 CPUs, Trainium3 UltraServers, workload fit & migration checklist

Prerequisites

  • Familiarity with Bedrock model catalog and agent capabilities
  • Basics of vector search architectures
  • Understanding of EC2 instance families and accelerator choices

Next Steps

  • Select one pilot use case for Nova (speech, multimodal, or reasoning) and outline eval plan
  • Draft migration/POC plan for S3 Vectors vs. current vector store
  • Benchmark targets for Graviton5/Trainium3 against existing instances
  • Define governance/policy needs before adopting AgentCore and Nova Act agents