Summary of "AWS re:Invent 2025 - Keynote with CEO Matt Garman"

Summary of AWS re:Invent 2025 Keynote with CEO Matt Garman

Key Technological Concepts & Product Features

1. AWS Growth & Infrastructure Highlights

AWS is now a $132 billion business, growing 20% year-over-year.
Amazon S3 stores over 500 trillion objects and handles 200 million requests per second.
Graviton CPUs have represented over half of new CPU capacity added for the past 3 years.
Amazon Bedrock powers AI inference for more than 100,000 companies.
Launch of Bedrock AgentCore SDK for enterprise-scale AI agents, downloaded over 2 million times.
Introduction of Ocelot, AWS’s first quantum computing chip prototype, reducing quantum error correction costs by 90%.
AWS operates 38 regions, 120 availability zones, and maintains 9 million km of private network cables.

2. AI and Agentic Systems

AI agents are emerging as the next major phase of AI, delivering real business value by automating tasks and workflows.
AWS’s AI infrastructure is optimized end-to-end for AI workloads, especially with Nvidia GPUs and custom AI chips.
New P6e-GB300 EC2 UltraServer instances powered by Nvidia GB300 GPUs offer 20x compute over the previous generation.
Partnership with Saudi Arabia’s Humain AI zone and launch of AWS AI Factories, enabling customers to deploy AWS AI infrastructure on-premises with full AWS AI services.

3. Custom AI Chips - Trainium Series

Trainium2 chips power the majority of Amazon Bedrock inference today.
Launch of Trainium3 UltraServers featuring the first 3-nm AI chip in the cloud, offering:
- 4.4x compute improvement
- 3.9x memory bandwidth increase
- 5x tokens per megawatt efficiency over Trainium2
Preview of Trainium4 with planned improvements:
- 6x FP4 compute
- 4x memory bandwidth
- 2x memory capacity
Trainium chips enable large-scale AI training and inference with integrated software and networking.

4. Amazon Bedrock & AI Models

Bedrock is a generative AI platform offering broad model choice, customization, and integration with data and guardrails.
Introduction of new open-weight models including:
- Google’s Gemma
- Nvidia’s Nemotron
- Mistral AI’s Large and Ministral 3 models
Expansion of Amazon’s own Nova family of foundation models:
- Nova 2 Lite: cost-effective, fast reasoning model
- Nova 2 Pro: high-end reasoning model outperforming competitors
- Nova 2 Sonic: real-time speech-to-speech conversational AI
- Nova 2 Omni: first unified multimodal reasoning model supporting text, image, video, audio inputs and text/image outputs
Launch of Amazon Nova Forge, a new service enabling customers to blend proprietary data into training checkpoints to create custom models (“Novellas”) that deeply understand their domain without forgetting foundational knowledge.

5. Agent Platform - Bedrock AgentCore

Modular, secure, serverless runtime for building, deploying, and operating AI agents at scale.
Features include:
- Session isolation
- Memory management
- Secure identity and authentication
- Tool discovery
- Observability
- Code interpreter
- Managed browser
Supports integration with popular frameworks (Crew AI, LangChain, LlamaIndex, Strands) and models (Bedrock, OpenAI, Gemini).
New capabilities announced:
- AgentCore Policy: real-time deterministic controls over agent actions and tool access via natural language policies compiled to Cedar language.
- AgentCore Evaluations: continuous quality inspection of agent behavior with pre-built and custom evaluators integrated into CloudWatch.

6. Customer & Partner Highlights

Startups showcased:
- AudioShake (sound separation technology)
- Lila (AI science factories)
- Gradial (marketing automation)
- Writer (enterprise agentic AI platform)
Large enterprises highlighted:
- Sony
- Adobe
- Nasdaq
- Bristol-Myers Squibb
- Mars
Use cases:
- Sony uses AWS AI and Bedrock AgentCore for enterprise AI transformation, fan engagement, and content IP management.
- Adobe integrates AWS AI infrastructure into Firefly, Acrobat Studio, Experience Platform, and Adobe GenStudio, leveraging Bedrock and SageMaker for generative AI and agentic workflows.
- Writer built a full-stack enterprise AI platform on AWS, integrating Bedrock Guardrails and models for compliance and agent governance.

7. Developer Tools & Modernization

AWS Transform helps modernize legacy codebases and applications with custom transformation agents supporting any language or framework.
Launch of Kiro, an agentic development environment for spec-driven AI coding, enabling developers to collaborate with AI agents to build complex software faster.
Introduction of Kiro Autonomous Agent, a frontier agent that autonomously manages complex coding tasks across multiple repositories, maintains context, learns team preferences, and scales parallel workflows.
Launch of AWS Security Agent, a frontier agent embedding security reviews, vulnerability scanning, and penetration testing into developer workflows with real-time feedback.
Launch of AWS DevOps Agent, a frontier agent that autonomously investigates incidents, diagnoses root causes, integrates with observability tools, and recommends CI/CD guardrails to prevent future issues.

8. Other Notable AWS Service Announcements

New EC2 instances:
- Large memory X family with Intel Xeon 6
- AMD EPYC with 3TB memory
- C8a (AMD EPYC)
- C8ine (Intel Xeon with Nitro v6)
- M8azn (fastest CPU clock frequency)
- New Apple Mac instances (M3 Ultra Mac, M4 Max Mac)
Lambda Durable Functions for long-running workloads with state management.
S3 max object size increased 10x to 50 TB; batch operations 10x faster.
Intelligent-Tiering for S3 Tables and cross-region replication.
General availability of S3 Vectors for storing/querying trillions of vector embeddings with 90% cost reduction.
GPU acceleration for vector indexing in Amazon OpenSearch (10x faster, 1/4 cost).
EMR Serverless no longer requires provisioning local storage.
GuardDuty extended threat detection now supports ECS containers.
GA of enhanced Security Hub with near real-time risk analytics.
New unified data store in CloudWatch for operational security and compliance logs.
Increased RDS storage for SQL Server and Oracle to 256 TB with 4x IOPS.
New ability to specify vCPU count for SQL Server licensing cost optimization.
Support for SQL Server Developer Edition for free development/testing.
Launch of Database Savings Plans for up to 35% savings across database usage.

Reviews, Guides, Tutorials, and Use Cases Highlighted

AudioShake: Sound separation technology used in media, healthcare, and accessibility.
Lila: AI science factories for autonomous scientific research.
Gradial: AI-powered marketing content orchestration using Bedrock and Nova.
Sony: Leveraging AWS for fan engagement, enterprise AI, and compliance automation.
Adobe: Creative AI workflows powered by AWS AI services; integration of Bedrock and SageMaker.
Writer: Full-stack enterprise AI platform with Bedrock integration and compliance guardrails.
Nasdaq & Bristol-Myers Squibb: AgentCore for financial and pharmaceutical AI workflows.
QAD: Code modernization with AWS Transform.
Thomson Reuters: Mainframe modernization using AWS Transform.
Developer Testimonials on Kiro: Significant productivity gains, large project delivery with fewer people and faster timelines.
Reddit: Custom model training with Nova Forge for content moderation.
Internal Amazon use of Kiro: Standardized AI development environment.

Main Speakers / Sources

Matt Garman – CEO of AWS, keynote presenter.
Jessica (AudioShake) – Startup founder presenting sound separation technology.
John Kodera – Chief Digital Officer, Sony Group Corporation.
Shantanu Narayen – CEO and Chair, Adobe.
May Habib – CEO, Writer.
Various AWS product and engineering leaders presenting new services and agent capabilities.

This keynote emphasized AWS’s leadership in AI infrastructure, agentic AI platforms, custom AI chips, developer productivity tools, and broad cloud service innovation, showcasing how customers from startups to global enterprises are leveraging AWS to innovate rapidly and securely.