Note: This page is being actively updated as AWS re:Invent 2025 unfolds. Check back often for the latest announcements, keynote coverage, and insights. Last updated: December 4, 2025 (Afternoon).

AWS re:Invent 2025 held in Las Vegas from December 1st to 5th, is AWS’s premier learning conference for the global cloud computing community. This year’s event marks a decisive shift toward agentic AI—with nearly every major announcement centered on building, deploying, and operating AI agents at scale.

Day 1 - 12/02/25 (CEO Keynote)

8:00AM CEO Keynote - Matt Garman, CEO (AWS)

Matt Garman, CEO of Amazon Web Services, delivered the opening keynote sharing how AWS is innovating across every aspect of the world’s leading cloud. The central theme: “Why can’t developers focus on building?”—and AWS’s answer is billions of AI agents.

Key Themes

• Resilient Infrastructure - More Availability Zones coming globally
• Developer Focus - Vision for billions of AI agents handling operational toil
• Four Pillars for AI Agents:
  1. AI Infrastructure (GPUs, custom silicon)
  2. Inference Platform (Bedrock)
  3. Your Data (Nova Forge)
  4. Tools to Build Agents (AgentCore)

Announcements

AI Infrastructure & Custom Silicon

AWS AI Factories

Dedicated customer-specific AI infrastructure that integrates with existing hardware, data center, and networking investments.

Trainium3 UltraServers - First 3nm AWS AI Chip

The 4th generation of AWS’s custom AI silicon delivers breakthrough performance:

Trn3 UltraServer Aggregate Specs:

• Up to 20.7 TB HBM3e memory
• 706 TB/s memory bandwidth
• 362 FP8 PFLOPs compute
• Supports up to 144 Trainium3 chips
• NeuronSwitch-v1 fabric (2TB/s per chip)
• Native PyTorch, JAX, Hugging Face support
• Advanced data types: FP32, BF16, MXFP8, MXFP4

Optimized for: agentic/reasoning workloads, video generation, reinforcement learning, Mixture-of-Experts models.

Trainium4 - Coming Soon

Next generation already announced, continuing AWS’s aggressive custom silicon roadmap.

Amazon Bedrock & Foundation Models

Amazon Bedrock - Now Powers 100,000+ Organizations

• Guardrails: Blocks up to 88% harmful content with 99% accuracy
• Model Distillation: Up to 500% faster, 75% less cost
• Intelligent Prompt Routing: Auto-routes queries to optimal model
• New Models: Mistral Large 3, Ministral 3
• Compliance: ISO, SOC, GDPR, FedRAMP High, HIPAA

Amazon Nova 2 Family

Understanding Models (Text, Image, Video → Text):

• Nova Micro - Entry-level
• Nova Lite - Fast, cost-effective reasoning
• Nova Pro - Higher capability
• Nova 2 Lite - Powers Nova Act browser automation

Creative Content Models (Text, Image → Image/Video):

• Nova Canvas - Image generation
• Nova Reel - Video generation
• Nova 2 Sonic - Speech-to-speech for conversational AI

Built on AI technologies from Amazon internal systems (Alexa+, Amazon Ads).

Amazon Nova Forge

Build custom frontier models that deeply embed domain expertise without the traditional barriers of cost, compute, and time.

Amazon Nova Act

Build agents that automate browser-based UI workflows. Powered by custom Nova 2 Lite model for reliable production UI automation.

Agent Platform

Amazon Bedrock AgentCore

An agentic platform for building, deploying, and operating AI agents securely at scale—no infrastructure management needed.

Build:

• Persistent memory that learns from interactions
• Secure browser runtime
• Code interpreter for complex tasks
• Framework agnostic—any framework, model, or tool
• Semantic tool discovery

Deploy:

• Session isolation
• Long-running workloads up to 8 hours
• Native identity provider integration
• Fine-grained access policies
• Serverless deployment

Monitor:

• Real-time CloudWatch metrics
• OpenTelemetry integration
• Agent quality evaluation (correctness, safety, goal success)

AWS Agents & Applications

Amazon Q - Agentic Business Intelligence

• Deep Research - AI-powered research capabilities
• Analyze and Visualize - Data analysis and visualization
• Q Flows - Automated workflows for business processes

Kiro Autonomous Agent

An autonomous development agent that:

• Plans and executes multi-step coding tasks across multiple repositories
• Learns from your reviews and maintains context over time
• Runs work in isolated sandboxes
• Opens pull requests for review
• Integrates with GitHub and Jira

AWS Security Agent (Preview)

A frontier agent that proactively secures applications throughout the development lifecycle:

• Automated application security reviews tailored to organizational requirements
• Context-aware penetration testing on demand
• Continuously validates security from design to deployment

AWS DevOps Agent (Preview)

An autonomous on-call engineer that:

• Analyzes data across CloudWatch, GitHub, ServiceNow, and other tools
• Identifies root causes
• Coordinates incident response

Compute

New EC2 Instance Classes:

• P6e - NVIDIA GB200 & GB300
• X81 instances
• X8aedz instances
• C8a & C8ine instances
• M8azn instances
• EC2 M3 and M4 instances

Storage

S3 Vectors (GA)

First cloud object store with native vector storage and querying (product page):

• Up to 2 billion vectors per index
• 10,000 indexes per bucket (up to 20 trillion vectors)
• 100ms warm query latency
• 90% cost reduction in vector storage/querying
• Native integration with Bedrock Knowledge Bases

Additional S3 Updates:

• Object size limit increased to 50TB
• Batch operations 10x faster
• S3 Tables Intelligent Tiering
• S3 Table replication across accounts
• S3 FSx for NetApp OnTap
• GPU index search in OpenSearch - Serverless GPU acceleration for vector index

Database

• Oracle and SQL storage capacity improvements in RDS
• Optimize CPUs for RDS SQL
• SQL Developer addition
• Database savings plans

Serverless

Lambda Durable Functions (GA)

Long-running workflows with built-in state persistence (docs):

• Execute for up to one year
• Checkpoint/replay mechanism—skips completed operations on restart
• Pay only for actual processing time (no charges during suspension)
• SDK support: JavaScript, TypeScript, Python

Use Cases: Payment workflows, order fulfillment, AI coordination, multi-step processes.

Security

• GuardDuty extended for ECS and EC2 (no additional cost)
• Security Hub GA
• Unified data store in CloudWatch - Unified log management across operational, security, and compliance use cases

Innovation Talks - Day 1

11:30 AM - INV211: Amazon’s AI Innovations

How Amazon leaders across Zoox, Prime Video, and Amazon Stores are leveraging AI to power their next-generation innovations with AWS.

Zoox - Autonomous Ride-Hailing

Zoox is an autonomous ride-hailing service focused on comfort, control, and safety:

• Booking: Through the app, a vehicle arrives with lounge-like “carriage” seating
• Customization: Set music, temperature, and lighting per seat
• Comfort: Zoned climate control, ample legroom, quiet space to work
• Support: Live support available in-app

Current Markets: Las Vegas, San Francisco Coming Soon: Austin, Miami

1:00 PM - INV201: Harnessing Analytics

Speaker: Mai-Lan Tomsen Bukovec, VP of AWS Technology, explores emerging trends from Open Table Formats (OTF) to agentic infrastructure, and how to future-proof your data foundation for analytics at scale.

Key Topics:

• SageMaker Notebooks - Launch fully managed JupyterLab from Amazon SageMaker Studio in seconds
• Open Table Formats evolution
• Agentic infrastructure for analytics

2:30 PM - INV215: AWS storage innovations

This session unveils breakthrough innovations like S3 Tables for analytics optimization, S3 Vectors for AI/ML acceleration, and seamless SAN migration pathways that eliminate traditional infrastructure constraints.

4:00 PM - INV202: AI agents in action

Speaker: Shaown Nandi - Director Technology - Amazon Web Services (AWS)

New AWS capabilities empower builders to design secure, reasoning-driven agents that orchestrate data, code, and tools at scale, with an emphasis on governance, reliability, and cost efficiency. Discover how AWS customers are deploying production-ready agents today, and learn best practices to help you architect agentic applications that autonomously adapt, optimize, and act in real-time.

Day 2 - 12/03/25 (Agentic AI Keynote)

8:30AM KEY002: Future of Agentic AI Keynote - Dr. Swami Sivasubramanian, VP AWS

Dr. Swami Sivasubramanian, VP of AI and Data at AWS, presents how Agentic AI is transforming how we work. Key theme: “From Idea to Impact at speed.”

Key Themes

• From Idea to Impact at speed - Central message of the keynote
• What makes an Agent: Model + Code + Tools
• Challenge: Getting agents into production is too hard
• Answer: Model customization for efficiency (latency, scalability, agility)

Announcements

AI/ML & Bedrock

Strands SDK - Open Source AI Agents SDK

An open source SDK for building multi-agent AI systems (announcement):

• Python-based with TypeScript support (preview)
• Model-driven orchestration - Agent behavior driven by LLM capabilities
• Provider agnostic - Bedrock, OpenAI, Anthropic, Gemini, local models (Ollama)
• Multi-agent primitives: Handoffs, swarms, graph workflows
• Edge device support for robotics
• Deployment targets: Lambda, Fargate, EKS, EC2, AgentCore

AgentCore Memory

Episodic functionality that enables agents to remember and learn from previous experiences.

Bedrock Reinforcement Fine Tuning (RFT)

Advanced model customization without deep ML expertise:

• 66% average accuracy gains over base models
• No large labeled datasets required
• Currently supports Amazon Nova 2 Lite (more models coming)
• Reward functions: Rule-based graders, AI judges, built-in templates
• Supports objective (code gen, math) and subjective (instruction following) tasks

Amazon Nova Act - Enhanced

Automate production UI workflows with parallel training capabilities.

SageMaker

SageMaker AI - Serverless Model Customization

Quickly customize models with proprietary data:

• Supported models: Amazon Nova, Llama, Qwen, DeepSeek, GPT-OSS
• Techniques: Supervised fine-tuning, reinforcement learning, direct preference optimization (DPO)
• AI agent-guided workflow (preview) - Uses natural language for data generation, quality analysis, training
• Available regions: Ireland, N. Virginia, Tokyo, Oregon

SageMaker HyperPod Checkpointless Training

Revolutionary approach to training recovery:

• Eliminates checkpoint-based job-level restart
• Reduces training recovery from hours to minutes
• Up to 95% training goodput on large clusters
• Peer-to-peer state transfer from healthy accelerators
• Zero code changes with HyperPod recipes for Llama, GPT OSS

Innovation Talks - Day 2

10:00 AM - INV214: Intelligent Security

Speaker: Shaown Nandi - Director Technology - Amazon Web Services (AWS)

New AWS capabilities empower builders to design secure, reasoning-driven agents that orchestrate data, code, and tools at scale, with an emphasis on governance, reliability, and cost efficiency. Discover how AWS customers are deploying production-ready agents today, and learn best practices to help you architect agentic applications that autonomously adapt, optimize, and act in real-time.

Protection at scale from development to production. Deep dive into AWS Active Defense systems:

How These Feed AWS Services:

• Automatic Protection: AWS Shield, Amazon S3, Amazon VPC, Route 53 Resolver DNS Firewall
• Managed Rulesets: AWS WAF, AWS Network Firewall
• Findings: Amazon GuardDuty, Amazon Inspector

Key Insight: AWS uses massive-scale internal threat intelligence to automatically protect customer workloads—the “secret sauce” behind their managed security services.

11:30 AM - INV213: Cloud Network Innovation

Speaker: Robert Kennedy - AWS networking innovations shaping the future of cloud infrastructure.

AWS networking innovations that are shaping the future of cloud infrastructure. From our global backbone to AI/ML-optimized data centers, discover breakthrough technologies in content delivery, security, VPC networking, and global connectivity.

Cross-Region Connectivity for AWS PrivateLink

Share and access VPC endpoint services across different Regions. Service providers can now offer SaaS solutions privately to a global audience from a single Region.

AWS Interconnect - Multicloud (Preview)

Simple, resilient, high-speed private connections to other CSPs:

• First partner: Google Cloud (launch partner)
• Coming 2026: Microsoft Azure
• Eliminates need for third-party connectivity providers

CloudFront Flat Rate Plans

New predictable monthly pricing with no overage charges:

Included Services: Amazon CloudFront, CloudWatch, WAF, Lambda, Route 53, S3

1:00 PM - INV207: AWS Compute

Speakers: Willem Visser, VP of Amazon EC2 & Barry Cooks, VP of Compute Abstractions

ECS Express Mode

Rapidly launch containerized applications:

• Auto-deploy with automatic domain name generation
• Auto-scales in response to traffic patterns
• Consolidates up to 25 services behind single ALB
• No additional charge - pay only for created resources
• Available via Console, SDK, CLI, CloudFormation, CDK, Terraform

EKS Capabilities

New managed add-ons:

• Managed ArgoCD - GitOps continuous delivery
• ACK Controllers - Manage AWS resources via Kubernetes APIs
• kro (Kube Resource Orchestrator) - Custom Kubernetes APIs for resource composition

EKS Ultra-Scale Clusters

Massive scale Kubernetes clusters for enterprise workloads.

Additional Compute Updates

• GuardDuty Extended Threat Detection - Now supports EC2 and ECS
• Lambda Managed EC2

Day 3 - 12/04/25 (Infrastructure Innovations & Closing)

9:00AM KEY004: Infrastructure Innovations Keynote

Speakers: Peter DeSantis, SVP AWS & Dave Brown, VP of Compute and MLS

Dive deep into the technology that powers AWS services. Get a closer look at how AWS’s unique approach and culture of innovation create leading-edge solutions, from silicon to services.

Announcements

Compute & Custom Silicon

AWS Graviton5 (Preview) - Next-Gen ARM Processor

AWS’s latest custom ARM processor delivers significant improvements:

Architecture:

• Built on AWS Nitro System with hardware and software innovations
• Custom designed specifically for cloud workloads
• Available in EC2 M9g instances (preview)

Ideal Workloads:

• Application servers and microservices
• Gaming servers
• Midsize data stores
• Caching fleets
• Machine learning inference

Trainium3 Updates

Additional details from the keynote:

• Trainium PyTorch native support - First-class framework integration
• 4.4x higher performance vs Trainium2
• 5x Higher Tokens/MegaWatt efficiency
• Apple Swift demo showcased performance on Graviton

Trainium4 - Coming Soon

Next generation announced, continuing AWS’s aggressive custom silicon roadmap.

Serverless Evolution

Lambda Managed Instances

A new deployment model that combines serverless simplicity with EC2 flexibility:

• How it works: Lambda runs on EC2 instances in your account
• Provisioning: Lambda handles all provisioning automatically
• Benefits:
  • Access specialized compute options (GPU, custom instance types)
  • Optimize costs for steady-state workloads
  • Maintain serverless development experience
  • No infrastructure management required

This bridges the gap between Lambda’s simplicity and EC2’s flexibility, ideal for workloads that need specialized compute but want serverless operations.

Project Mantle (Bedrock Redesign)

Internal AWS initiative revealed during the keynote:

• Key insight: “Inference needs a different architecture”
• Focus: Redesigning Bedrock’s architecture specifically for LLM inference workloads
• Goal: Optimized performance and efficiency for generative AI inference
• Signals AWS’s continued investment in purpose-built AI infrastructure

Vector Search & AI Infrastructure

Nova Multimodal Embeddings

New embedding capabilities for multimodal content—combining text, image, and video understanding.

S3 Vectors - Enhanced Performance

Building on the GA announcement, additional performance details:

• Pre-compute approximate nearest neighbor (ANN) search
• Sub-100ms queries over 2TB of vector data
• Optimized for AI agent memory and context enhancement

OpenSearch Vector Search

Continued performance improvements for vector database workloads.

TwelveLabs Demo

Customer showcase demonstrating AI video analysis:

• Turns hours of video into structured, searchable data
• Example of AI/ML infrastructure enabling new use cases

11:00 AM - INV205: Reinventing Software Development

The rise of AI coding and autonomous agents is transforming how developers work—enhancing productivity, improving flow state, and making development more intuitive.

Key Topics

Kiro - Autonomous Development Agent

Deep dive into Kiro’s capabilities:

• Spec-driven development - Define requirements in natural language, Kiro implements
• Property-based testing - Automated test generation based on specifications
• Plans and executes multi-step coding tasks
• Works across multiple repositories
• Learns from your reviews and maintains context

Kiro Powers - Dynamic AI Context System

A breakthrough in AI agent capabilities—dynamic, context-aware integration that provides instant access to specialized knowledge for any technology.

Core Innovation:

• Tools load only when relevant to current task
• Prevents context overflow by activating powers selectively
• Minimal baseline context usage until needed

Technical Structure:

power/
├── POWER.md           # Entry point steering file
├── mcp-server.json    # MCP server configuration
├── workflows/         # Workflow-specific steering files
└── keywords.yaml      # Activation keywords

How It Works:

1. Mention “database” → Supabase power activates
2. Loads specific MCP tools for database work
3. Switches context automatically when task changes (e.g., moving to deployment)

Launch Partners:

• Datadog, Dynatrace (observability)
• Figma (design)
• Neon, Supabase (databases)
• Netlify (deployment)
• Postman (APIs)
• Stripe (payments)
• Strands Agent (AI agents)

Vision: A model for continual learning where AI agents can expand capabilities dynamically, automatically update with framework changes, and learn specialized knowledge without overwhelming the system.

3:30PM Closing Keynote - Dr. Werner Vogels, VP & CTO (Amazon)

Dr. Werner Vogels, Amazon.com’s VP and CTO, joins us for his fourteenth re:Invent appearance. Werner’s keynotes typically cover architectural best practices, distributed systems insights, and the principles that guide building systems at massive scale.

Notes to be added after keynote

Key Takeaways

1. Agentic AI is the focus - Every major announcement centers on building, deploying, and operating AI agents
2. Custom silicon strategy - AWS continues investing in custom chips: Trainium3 (3nm), Graviton5 (192 cores), with Trainium4 announced
3. Framework agnostic - AgentCore, Strands SDK, and Kiro work with any model/framework
4. Enterprise modernization - AWS Transform provides specialized agents for Windows, mainframe, VMware migrations
5. Vector-native storage - S3 Vectors brings native vector capabilities to object storage (90% cost reduction, sub-100ms queries over 2TB)
6. Long-running serverless - Lambda Durable Functions enable year-long workflows with state persistence
7. Dynamic AI tooling - Kiro Powers enables on-demand loading of specialized AI capabilities through MCP
8. Inference optimization - Project Mantle redesigns Bedrock architecture specifically for LLM inference
9. Serverless flexibility - Lambda Managed Instances bridges serverless simplicity with EC2 compute options
10. Multicloud connectivity - AWS Interconnect enables private connections to Google Cloud (Azure coming 2026)

Resources

• Live Stream: https://reinvent.awsevents.com/livestream/
• Session Catalog: https://registration.awsevents.com/flow/awsevents/reinvent25/public/page/catalog
• Top Announcements: https://aws.amazon.com/blogs/aws/top-announcements-of-aws-reinvent-2025/
• AWS Blog: https://aws.amazon.com/blogs/
• Videos: https://www.youtube.com/@amazonwebservices

Product Pages

• Amazon Nova
• Amazon Bedrock
• Amazon Bedrock AgentCore
• AWS Trainium
• EC2 Trn3 Instances
• EC2 M9g Instances (Graviton5)
• S3 Vectors
• Lambda Durable Functions
• Amazon Q Developer
• AWS Transform
• SageMaker Notebooks
• Kiro
• Kiro Powers
• TwelveLabs
• Zoox

Blog Posts

• Lambda Managed Instances
• S3 Vectors GA
• S3 Tables Intelligent Tiering
• OpenSearch GPU Acceleration
• Lambda Durable Functions
• Database Savings Plans
• CloudWatch Unified Data Store
• Nova 2 Lite
• Nova 2 Sonic
• Nova Forge
• AWS Security Agent
• AWS DevOps Agent