AWS re:Invent 2025 clearly demonstrates that “freedom to invent” lies at the heart of everything AWS does. Matt Garman, CEO of AWS, opens by emphasizing the foundational belief that gave birth to AWS: 

Infrastructure is not the limit for developers.  
 
They should not be spending weeks setting up servers, ironing bottlenecks, or waiting for an approval cycle. Freedom should not be a luxury, while building should be immediate, secure, at scale, and without friction.  
 
Twenty years later, AWS’s core mission remains unchanged, yet it has evolved into a new frontier. We’re no longer talking about freeing developers from hardware. Now, we’re working towards freeing every team, in every industry, to invent with agents and systems that don’t just address queries but take action on our behalf.

This year’s keynote has made this evolution transparent. AWS is laying down the infrastructures, models, platforms, and guardrails to assist organizations in shifting AI experiments to AI outcomes. The message, adorned with custom silicon to autonomous agents, is direct, bold, and confident.  
 
We’re moving from assisting to acting. And AWS is building the stack that makes invention inevitable! 
 
Let’s dive into some of the groundbreaking announcements this year.

Compute: The Rise of Purpose-Built AI Infrastructure

“To unlock the next wave of business transformation, we need to push beyond incremental speed. We need to rethink the infrastructure stack that powers AI.” – Matt Garman  
 
Garman presented this as his core argument: agents can only be as powerful as the compute can fuel them. And AWS is adamant in building the world’s most scalable, reliable, and energy-efficient foundation for AI.  
 
The keynote was anchored as it started off with announcements in scale:  

• Amazon S3 stores 500+ trillion objects, sustaining 200M+ requests per second.  

• For the third consecutive year, over half of AWS’s added CPU capacity was supported by Graviton. 

• Amazon Bedrock now powers AI inference for 100,000+ companies. 

• The Amazon Bedrock AgentCore SDK has been downloaded 2M+ times since launch. 

And the star of the show is Amazon EC2 Tranium3 UltraServers, which is now generally available. 

Amazon EC2 Trainum3 Ultra Servers

AWS introduced its most advanced AI system yet:  

• First AWS chip designed on a 3nm processor 

• 4.4x more compute, 3.9x more memory bandwidth 

• 5x more tokens per megawatt 

• Each Ultra Server contains 144 Trainium3 chips, delivering 362 FP8 PFLOPS and 700 TB per second aggregate bandwidth 

AWS has also revealed that AWS Trainium4 is under development and pointed out that the cadence of custom silicon innovation is faster than ever. 

NVIDIA UltraServers & GPU Leadership

AWS expanded its NVIDIA-powered lineup with P6e-GB300 UltraServers running the latest GB300 NVL72 systems. And the message is, if it runs on NVIDIA, AWS will run it at the biggest scale with the highest reliability.

AWS AI Factories

“We asked ourselves, what if customers could deploy AWS-level AI infrastructure inside their own data centers?” – Matt Garman 
 
The answer: AWS AI Factories.  
 
These factories are dedicated to AI Infrastructures that operate in a private AWS region inside the customers’ facilities.

More Compute Options

AWS also introduced additional compute classes, including:  

• X8i and X8aedz: next-gen large-memory instances 

• C8a and C8ine: CPU + network-optimized 

• M8azn: highest single-thread performance in the cloud 

• EC2 Mac M3 Ultra and EC2 Mac M4 Max 

• AWS Lambda Durable Functions for long-lived, agentic workflows

Storage: Bigger, Faster, and More Intelligent

“Data isn’t just fuel for your business; it’s context for your agents.” – Matt Garman 
 
Garman focused that agents need a richer and more accessible, performant system to rationale and act effectively. Storage cannot be passive anymore, especially when it becomes a part of the agents’ workflow.  

• Amazon S3 Object Size Limit increases to 50 TB: 10x larger object for next generation media, simulation outputs and model artifacts.  

• Amazon S3 Batch Operations is 10x Faster: It is a significant upgrade for high-volume data orchestration. 

• Amazon S3 Tables Enhancements: Further enhancements have been made with respect to Intelligent-Tiering and automatic cross-region replication.  

• Amazon S3 Access Points for FSx: Now available for FSx for NetApp ONTAP, bringing the file and object ecosystem even closer. 

• Amazon S3 Vectors: There is native vector storage for AI and RAG workloads.  

• Amazon OpenSearch Service Vector Indexing with GPUs: You can experience 10x faster indexing with GPU acceleration. 

Scaling Databases without Friction

“Your agents will rely on data that must be fast, consistent, and always available, regardless of how large your systems grow.” – Matt Garman 

Garman made it transparent that database scalability is not just operational but an important AI prerequisite. 

• Amazon RDS Increases Storage Capacity Up to 256 TB: SQL Server & Oracle storage limit is quadrupled. 

• SQL Server License Optimization: vCPU configuration to reduce licensing footprint & SQL Server Developer Edition support. 

• Database Savings Plan: A long-awaited discount model is launched, which covers multiple database engines. 

The Rise of Amazon Nova

“Models are no longer standalone systems. They are the cognitive engines behind your agents.” – Matt Garman 

Garman emphasized that the model layer has to evolve in parallel with agentic infrastructure.

Amazon Nova 2 Model Family

• Nova 2 Lite for optimized efficiency 

• Nova 2 Pro for advanced reasoning 

• Nova 2 Sonic for real-time conversational performance 

• Nova 2 Omni for multimodal frontier model 

Amazon Nova Forge  

This service is specifically for training custom frontier-grade models that can deeply embed domain expertise. The best part is that all of this can be done without the traditional barriers of cost, compute, and time. 

The Execution Layer for Agents: Amazon Bedrock AgentCore

“AI assistants help you. AI agents act for you. Getting this right requires an entirely new kind of platform.” – Matt Garman 
 
Garman emphasized that agentic systems need their own secure, scalable layer to operate effectively. Key AgentCore Foundations include: 

• Full session isolation 

• Secure serverless runtime 

• Short- and long-term agent memory 

• Gateway to tools, APIs, and data 

• Identity and permissions 

• Real-time observability 

New capabilities include: 

• AgentCore Policy: Natural language rules are fed into Cedar, which enforces them outside agent code. 

• AgentCore Evaluations: Thirteen evaluators monitor helpfulness, correctness, harmfulness, and more in production as well as in development. 

New Agents from AWS

AWS is moving towards enabling every employee with the same power to invest as developers have with the right context, data, and guardrails.  

Amazon Quick Suite

This service brings an enterprise AI experience that integrates structured BI data, SaaS apps like Microsoft 365, Jira, ServiceNow, and unstructured enterprise content from SharePoint, drives, and more. It leverages AI agents to generate deep insights, research reports with citations, and personal automation workflows. Already in use by thousands of Amazon employees, Quick delivers measurable value across all departments every day.  

Amazon Connect

Amazon Connect now acts as an agent-driven contact center, scaling self-service automation, real-time agent coaching, and supporting large enterprise migrations. Powered by AI agents, it has reached a $1B annualized run rate while continuing to expand automation and intelligence across customer interactions. 

Developer Tools: Modernization, Velocity, and Structure

AWS helps developers with not only speed and efficiency, but also focuses on making the development cycle smarter and more resilient.  

• AWS Transform Custom: Custom modernized agents for Angular to React, VBA to Python, Bash to Rust, and Proprietary languages. QAD’s modernization time is from 2 weeks to 3 days. 

• AWS Transform for Windows: Modernizes complete Windows applications five times faster with AI-powered transformations across code, UI frameworks, databases, and deployment configurations. 

• AWS Transform for Mainframe: Reimagines mainframe applications into cloud-native architectures. Automated testing and intelligent analysis reduce timelines from years to months. 

• Kiro: Not new, but worth mentioning again. It uses a spec-driven workflow and, while not yet standardized internally at AWS, demonstrates powerful capabilities for complex modernization workflows. 

A New Class of Autonomous Systems: Frontier Agents

Garman put a groundbreaking emphasis on the next leap which is not just models but agents. These can pursue goals over hours or days while maintaining security and safety. This vision gave rise to Frontier Agents, built for long-term persistence, true autonomy, and massive parallelism. 

• Kiro Autonomous Agent: Designed to learn from feedback, retain context, create and test pull requests, implement features, and manage library upgrades across microservices. 

• AWS Security Agent: Provides automated remediation guidance, continuous penetration testing, pull request security scanning, and architectural reviews. 

• AWS DevOps Agent: An “always-on-call” engineer that understands full resource topology, recommends fixes and CI/CD guardrails, identifies root causes, and diagnoses incidents instantly. 

The 10 Minute Launch Sprint

Garman closed the keynote as a lightning round to match AWS’s pace of innovation and rattled through more than 25 launches, and some of them are:  

• Amazon EMR Serverless clusters require no storage provision  

• AWS Security Hub with near real-time analytics and risk prioritization 

• Amazon GuardDuty adds extended threat detection for Amazon EC2 and Amazon ECS 

• Amazon CloudWatch offers unified data management and analytics for operations, security, and compliance 

The Freedom to Invent is Agentic and Autonomous

Garman’s keynote at AWs re:invent 2025 outlined a future where agents transform how organizations build, operate, and secure their systems. The message is clear: 

Agents are the new applicants.  
 
And AWS is building every smaller to larger layer required to power them, whether it’s silicon or runtime to governance.  
 
At Cloudelligent, we share this very mission. We are working with organizations to turn these AWS capabilities into meaningful outcomes through agentic workflows, modernization, AI-driven operations, and secure cloud transformations.  
 
The freedom to invent has truly never been greater. It’s time to build what comes next.