AWS re:Invent 2025, held in Las Vegas from November 30 –4 December, put agentic AI and end‑to‑end AI infrastructure at the centre of the cloud provider’s roadmap. Across keynotes, press releases and partner sessions, the through‑line was clear: help enterprises move from chatbot-style pilots to production‑grade agents, backed by vertically integrated silicon, models and operations tooling.
This concise roundup highlights the major announcements across agentic AI and models, infrastructure and silicon, data and databases, modernisation and serverless, and security and operations—designed to give technology leaders a fast but substantive view of what changed, and where AWS is steering its customers next.
Agentic AI & Models
Amazon Nova 2 Family – AWS expanded Nova into a multi-modal lineup: Nova 2 Lite (cost-effective reasoning), Nova 2 Pro (complex agentic workloads), Nova 2 Sonic (speech-to-speech), and Nova 2 Omni (long-context multimodal). Nova 2 Lite benchmarked against Claude Haiku 4.5, GPT-5 Mini and Gemini Flash 2.5 for cost-performance on everyday enterprise tasks.
Amazon Bedrock AgentCore – New capabilities include policy constraints for agent actions, built-in evaluations for testing agent behaviour, enhanced memory for learning from historical interactions, and AgentOps integrations for observability and third-party framework support.
Amazon SageMaker AI – Unified, low-code interface for customising popular models (Nova, Llama, Qwen, DeepSeek) using reinforcement learning, supervised fine-tuning and direct preference optimisation. Serverless MLflow integration and automatic scaling reduce iteration cycles from months to days.
Frontier Agents in Production – AWS showcased Kiro (autonomous developer agent), DevOps agents for incident triage, and Security agents integrated into CI/CD pipelines. Industry commentary framed this as agentic AI moving “out of the POC graveyard” into production workloads.
Amazon Nova Act (GA) – UI automation agent for browser-controlled tasks (form filling, data extraction, QA testing) with reported >90% reliability for production workflows.
Amazon Connect Agentic Self-Service – Voice and messaging agents with autonomous decision-making for customer service, plus AI-powered product recommendations from real-time clickstream data.
Infrastructure & Silicon
Trainium3 – Custom AI training chip (3nm) delivering up to 4x faster training and up to 50% lower operating costs than equivalent GPUs. Trainium2 is now a multi-billion-dollar business approaching $2 billion annual revenue.
Trn3 UltraServers – Systems packing up to 144 Trainium3 chips with NeuronSwitch-v1 fabric, offering 4.4x more compute, 4x energy efficiency, and 4x memory bandwidth versus Trainium2 UltraServers.
Graviton5 – Arm-based CPU (3nm) with up to 25% better performance than Graviton4, optimised for non-GPU workloads. Early customer benchmarks show 20–35% performance gains and cost reductions.
AWS AI Factories – Dedicated AWS infrastructure deployed inside customer data centres, run exclusively for that customer. Combines Trainium, NVIDIA accelerators, AWS services (Bedrock, SageMaker AI), and operates like a private AWS Region. Targets regulated industries with data residency requirements.
Data, Analytics & Databases
Zero-ETL Expansion – Near real-time replication across Salesforce, SAP, ServiceNow, Zendesk, and self-managed databases into analytics and AI stores. Customers report latency reduction from minutes to seconds and 66–80% operational overhead reduction.
Amazon OpenSearch GPU-Accelerated Vector Search – Up to 10x faster vector database build times at roughly 25% of previous cost, with auto-optimisation for quality, speed and resource usage.
Database Savings Plans – New pricing model offering up to 35% savings across Aurora, RDS, DynamoDB, ElastiCache, DocumentDB, Neptune, Keyspaces, Timestream, and DMS with flexibility to switch engines and regions while retaining discounts.
Modernisation & Serverless
AWS Lambda Durable Functions – Write multi-step workflows as sequential code with durability primitives; suspend execution for up to one year while waiting for callbacks, resuming from checkpoints without charging for idle compute. Removes need for external orchestrators in many cases.
AWS Transform – AI-powered code modernisation tool that automates large-scale refactoring across codebases. Transform custom lets organisations define their own transformation patterns; custom learns them and executes autonomously. Customers report up to 80% reduction in modernisation execution time. Pre-built paths for Windows/.NET, VMware, and mainframe workloads.
EKS & Multicloud Networking – Enhanced Amazon EKS for managed workload orchestration and new AWS Interconnect service for high-bandwidth multicloud connectivity with other cloud providers.
By The Numbers
- Trainium2 multi-billion-dollar business at ~$2B annual revenue.
- AWS Transform: 1.1 billion lines of code processed, 810,000+ hours of manual effort saved.
- Air Canada: Lambda modernisations completed in days vs. weeks of manual work.
- Nova Act: >90% reliability for production UI automation workflows.
- Database Savings Plans: up to 35% cost savings on managed databases.
- Zero-ETL: up to 80% operational overhead reduction.
ALSO READ: OpenAI–TCS Talks Advance on HyperVault AI Data Centre Partnership: Reports
