Amazon Web Services (AWS) announced a new class of autonomous and scalable artificial intelligence, along with Nova 2 models and the Trainium3 chip to power Trainium3 UltraServers. These developments were presented as part of AWS re:Invent 2025, which runs until December 5 in Las Vegas (United States). Additionally, during this same event, some of the collaborations that Amazon is bringing to major companies to take advantage of AI services were announced.
Top Amazon news related to AI and chips
THE new class of AI agents They function as an extension of the software development team. Baptized as border agentsthese innovations are autonomous and They can work for hours or days without intervention. Additionally, three are available: Kiro Autonomous Agent, AWS Security Agent and AWS DevOps Agent.
On the other hand, the family of Nova 2 models —consisting of Nova 2 Lite, Nova 2 Pro, Nova 2 Omni and Nova 2 Sonic— offers reasoning, multimodal processing and conversational AI. Nova Forge, for its part, allows you to create AI models of border using an open training approach, and Nova Act achieves 90% reliability for AI agents in browser-based UX automation tasks.
Additionally, regarding training and deployment of AI models, AWS has confirmed the availability of Trainium3 UltraServers, powered by the new Trainium3 chip Built with 3nm technology to deliver greater computing performance, power efficiency and memory bandwidth.
And finally the solution Amazon Connect offers businesses automated voice experiences using neural text-to-speech technology, while AWS Transformation It allows you to modernize organizations with agentic artificial intelligence to accelerate code and application updates.
AWS Launches New Bedrock Models
AWS also added 28 new open weight models to Amazon Bedrockincluding Mistral Large 3 and Ministral 3 from Mistral AI, Gemma 3 from Google, M2 from MiniMax, Nemotron from Nvidia and GPT OSS Safeguard from OpenAI, among others.
In addition, Amazon Bedrock AgentCore has published its policy in preview, so organizational teams can establish clear boundaries on agent actions using natural language. While with AgentCore Reviews Quality control is simplified with thirteen predefined evaluators for dimensions such as accuracy and security, continuously sampling interactions in real time to trigger alerts when performance drops. In addition, AgentCore memory It incorporates episodic features that help agents learn from past experiences.
AWS Marketplace launches new solutions with AI
The AWS Marketplace digital catalog introduced AI-driven innovations to simplify the discovery and deployment of cloud solutions. These include agent mode for conversational discovery, express private offers for automated price personalization or multi-product solutions consolidating offers from multiple providers, and new flexible pricing models for professional services and AI agents.
AWS simplifies template customization
AWS announced new features in Amazon Bedrock and Amazon SageMaker HASThis puts advanced template customization within reach of any developer.
Fine adjustment of reinforcements In Amazon Bedrock makes it simple to create faster, more cost-effective models tailored to customer needswhile customizing serverless templates Amazon SageMaker AI reduces development time from months to days through an agent experience that guides developers in an autonomous approach
Both features address the efficiency challenge organizations face when deploying AI agents at scale, allowing small, specialized models to handle routine tasks.