Got a privilege to attend reinvent 2025 and learnt a lot from Keynote. Wish to share the new annoucement as recap and learning to all.
AWS re:Invent 2025 delivered an array of innovations, fundamentally reshaping the future of AI, compute, and cloud operations. The announcements focused on deploying powerful AI agents, accelerating model training with specialized infrastructure, and revolutionizing how developers manage technical debt and complex systems.
Here are 50 of the most significant new capabilities and services unveiled at re:Invent 2025:
I. Frontier AI Models and Customization
Introducing Nova 2 Lite: A fast, cost-effective reasoning model suitable for broad workloads, excelling at instruction following, tool calling, and code generation.
Launching Nova 2 Pro: Amazon's most intelligent reasoning model, purpose-built for highly complex agentic workflows, frequently outperforming leading models in agentic tool use benchmarks.
Previewing Nova 2 Omni: The industry's first unified model for multimodal reasoning and image generation, supporting input across text, image, video, and audio, while generating both text and image output.
Debuting Nova 2 Sonic: The next-generation speech-to-speech model enabling real-time, human-like conversational AI for applications.
Pioneering Nova Forge: A new service introducing "open training" that gives organizations exclusive access to Nova training checkpoints to blend proprietary data, resulting in custom Novella models.
Reinforcement Fine Tuning (RFT) in Bedrock: A new model customization capability using feedback-driven training that delivers an average of 66% accuracy gains over base models.
18 New Open-Weight Models on Bedrock: Massive expansion including Mistral Large 3, Ministral 3, Google Gemma 3, MiniMax M2, and Nvidia Nemotron.
Nova Act General Availability (GA): A new service for building AI agents that automate web browser-based tasks (UI workflows) with breakthrough reliability of over 90%.
Serverless Model Customization in SageMaker AI: New capabilities that accelerate model customization and experimentation cycles from months to days.
AWS Clean Rooms Synthetic Dataset Generation: Supports training ML models on sensitive collaborative data by generating privacy-enhancing synthetic datasets.
II. Advanced Agents and AgentCore Platform
Kiro Autonomous Agent: A frontier agent that acts as a virtual developer, autonomously tackling complex tasks from the backlog across multiple repositories while maintaining persistent context.
AWS Security Agent (Preview): A frontier agent that proactively reviews design documents, scans pull requests against organizational policies, and runs on-demand penetration testing.
AWS DevOps Agent (Preview): A frontier agent functioning as an autonomous on-call engineer, resolving and preventing incidents by correlating telemetry across observability, code, and CI/CD pipelines.
Policy in AgentCore (Preview): Provides real-time deterministic controls over specific agent actions and tool access, ensuring agents adhere to defined boundaries.
AgentCore Evaluations: New service helping developers continuously inspect agent quality using 13 pre-built evaluators for criteria like correctness, helpfulness, and harmfulness.
AgentCore Memory Episodic Functionality: Introduces new long-term memory to help agents learn from past experience and maintain context.
Amazon Quick Suite: A consumer AI experience for corporate employees, unifying structured and unstructured enterprise data and enabling the creation of Quick Flows (mini personal agents).
Kiro Powers: Enables developers to give Kiro agents instant expertise in specialized workflows and tools (e.g., Datadog, Figma, Postman) via Model Context Protocol (MCP) servers.
Strands Agents SDK in TypeScript (Preview): Extends the open-source agent framework to the TypeScript programming language.
Strands Edge Device Support (GA): Allows autonomous AI agents to run on small-scale devices for automotive, gaming, and robotics use cases.
III. AI Infrastructure and Core Compute
Trainium3 UltraServers GA: Powered by AWS's first three-nanometer AI chip, delivering up to 4.4x more compute and 3.9 times the memory bandwidth compared to Trainium2 UltraServers.
Trainium4 Announced: Projected to deliver six times the FP4 compute performance and four times more memory bandwidth compared to Trainium3.
AWS AI Factories: Enables customers to deploy dedicated AWS AI infrastructure (Nvidia GPUs, Trainium chips) inside the customer's own data centers to meet compliance and sovereignty needs.
Graviton5 Processors: AWS’s most advanced custom CPU, powering new EC2 M9g instances, delivering up to 25% higher performance than the previous generation.
New Nvidia P6e-GB300 UltraServers: Featuring the Nvidia GB300 NVL72 systems for demanding AI workloads and ideal for inference at scale.
Checkpointless Training on SageMaker HyperPod: Enables automatic recovery from infrastructure faults in minutes, achieving training cluster efficiency of up to 95%.
C8ine Instances: New instances utilizing custom Intel Xeon 6 processors and the latest Nitro v6 cards, delivering 2.5 times higher packet performance per vCPU.
M8azn Instances: Offering the absolute fastest CPU clock frequency available anywhere in the cloud, ideal for high-frequency trading and real-time analytics.
EC2 M3 Ultra Mac and M4 Max Mac Instances: Two new Apple Mac-based instances for developers using the latest Apple hardware.
IV. Modernization, Serverless, and Development
AWS Transform Custom: New AI-powered service allowing creation of custom code transformation agents to modernize any code, API, framework, or proprietary language, achieving transformations up to 5x faster.
AWS Transform Windows Modernization: Accelerates full-stack Windows modernization (code, databases, UI) and eliminates up to 70% of maintenance and licensing costs.
Lambda Durable Functions: Allows functions to program wait times and manage state for reliable, long-running workloads (up to a year).
Lambda Managed Instances: Allows customers to run Lambda functions on the Amazon EC2 instance of their choice (accessing specialized hardware/cost optimization) while retaining serverless simplicity.
IAM Policy Autopilot (Open Source MCP Server): Generates IAM policies based on developer intent and least-privilege design to prevent privilege sprawl.
AWS Transform Mainframe Reimagine Capabilities: New AI-powered capabilities to transform legacy mainframe applications into cloud-native architectures.
V. Storage, Data, and Analytics
S3 Max Object Size Increase: Maximum object size increased 10x, from 5 TB to 50 terabytes.
S3 Vectors General Availability (GA): Now supporting up to 20 trillion vectors per bucket and reducing the cost of storing and querying them by 90%.
Intelligent-Tiering for S3 Tables: Automatic cost optimization for S3 Table data, offering up to 80% savings on storage costs.
S3 Tables Automatic Replication: Enables automatic replication of S3 tables across AWS regions and accounts for data consistency.
S3 Batch Operations 10x Faster: Improved performance for large batch jobs to run up to 10x faster.
EMR Serverless No Local Storage Provisioning: Eliminates the need to provision local storage for Apache Spark workloads, reducing processing costs by up to 20%.
S3 Access Points for FSx for NetApp ONTAP: Allows customers to access ONTAP file data as if it were in S3, integrating it with S3-compatible AI/ML services.
VI. Cloud Operations and Security
CloudWatch Generative AI Observability: Provides comprehensive observability for generative AI applications and agents, monitoring latency, token usage, and errors without custom instrumentation.
CloudWatch Investigations 5 Whys Analysis: Integrated AI-powered workflow implementing AWS’s Correction of Errors (COE) methodology to drive to root causes for incidents.
CloudWatch Unified Data Store for Logs: A new unified store for operational, security, and compliance data, automating collection from AWS/third-party sources and storing it in S3 Tables.
CloudWatch Cross-Account/Cross-Region Log Centralization: Consolidates logs into a single destination account, with the first copy incurring no additional ingestion charges.
GuardDuty Extended Threat Detection for EC2 and ECS: Expansion providing broader visibility into sophisticated, multi-stage attacks across container and virtual machine environments.
AWS Security Hub GA: General availability with new capabilities, including near real-time risk analytics, a trends dashboard, and automated risk prioritization.
VII. FinOps, Databases, and Networking
Database Savings Plans: New flexible pricing model offering commitment-based discounts, providing savings of up to 35% across eligible database services.
RDS Storage Capacity Increase: Maximum storage capacity for RDS for SQL Server and Oracle increased from 64 TiB to 256 TiB (a 4x improvement).
RDS for SQL Server CPU Optimization and Developer Edition: Allows customers to specify the number of vCPUs to reduce CPU licensing costs and introduces support for the Developer Edition (no licensing fees).
Cost Efficiency Metric: AWS introduced a standardized Cost Efficiency Metric available in the Cost Optimization Hub to tie optimization to cloud business.
Compute Optimizer Automation: Allows FinOps practitioners to automatically apply optimization recommendations (e.g., managing EBS volumes or volume types) on a recurring schedule.
AWS Interconnect - Multicloud (Preview): Engineered solution for private, high-bandwidth connections between AWS and other service providers, starting with Google Cloud.
Route 53 Global Resolver (Preview): Simplifies hybrid DNS management with secure, anycast DNS resolution.
