The energy is electric in Las Vegas as AWS re:Invent 2025 is currently in full swing, and the announcements pouring out of the keynotes are immediately reshaping the future of cloud and enterprise technology. This year, the focus has crystallized around autonomous AI, next-generation infrastructure, and strategic partnerships.
Expanding the Amazon Nova model family and introducing Nova Forge
Amazon is significantly expanding its family of proprietary Nova AI models with the rollout of the Amazon Nova 2 models. These new models are designed to deliver industry-leading price-performance across various tasks and are available via Amazon Bedrock.
The new models include:
-
Nova 2 Lite: A fast, cost-effective reasoning model for everyday tasks like customer service chatbots and document processing. It supports a one-million-token context window and is available for customization via supervised fine-tuning (SFT).
-
Nova 2 Pro: AWS’s most intelligent model, available in preview, designed for highly complex, multi-step agentic tasks such as video reasoning and software migrations.
-
Nova 2 Sonic: A new speech-to-speech model for conversational AI.
-
Nova 2 Omni: A multimodal reasoning and generation model that processes and generates text, images, video, and speech.
Alongside the models, AWS introduced Nova Forge, a pioneering “open training” service. Nova Forge enables organizations to build their own proprietary, customized frontier models, called “Novellas,” by blending their own data with Nova’s original training data and checkpoints. This service is designed to mitigate catastrophic forgetting, a risk where models forget foundational capabilities after being fine-tuned with new data post-training.
AWS is introducing autonomous frontier agents
AWS is moving into a new class of powerful, autonomous AI agents designed to work for hours or even days without constant human intervention, positioning them as an extension of a customer’s software development team. These agents maintain persistent context across sessions, meaning they don’t run out of memory or forget past instructions.
Three initial “frontier agents” were introduced:
-
Kiro autonomous agent: Acts as a virtual developer that can be assigned a complex task from a backlog. It independently figures out how to complete the work, spanning multiple repositories, and creates verified pull requests. It learns how a team likes to work and continuously deepens its understanding of the codebase and standards over time. For example, AWS CEO Matt Garman described assigning Kiro to update a critical code library used by 15 microservices in a single prompt.
-
AWS Security Agent: Works independently to identify security problems as code is written, tests it, and offers suggested fixes, transforming penetration testing from a slow, manual process to an on-demand capability.
-
AWS DevOps Agent: Automates operational tasks, such as testing new code for performance issues or compatibility with other software and environments, helping pinpoint root causes of performance issues autonomously.
The launch of Trainium3 UltraServers for AI training
To support the massive scale required for training next-generation AI models, AWS rolled out the Amazon EC2 Trn3 UltraServers. These systems are powered by the new Trainium3 chip, built on AWS’s first 3nm AI silicon.
Key specifications for the Trn3 UltraServers:
-
They deliver up to 4.4x higher performance and over 4x better performance/watt compared to the previous Trn2 UltraServers.
-
A single UltraServer can host up to 144 Trainium3 chips, providing up to 362 MXFP8 PFLOPs of compute.
-
The system features up to 20.7 TB of HBM3e (High Bandwidth Memory) and 706 TB/s aggregate memory bandwidth.
-
The Trainium3 chip is claimed to be 40% more energy efficient than the previous generation.
-
Looking ahead, AWS also teased the roadmap for Trainium4, which will support Nvidia’s NVLink Fusion interconnect technology, allowing for interoperability with Nvidia GPUs.
New AI-powered database savings plans
AWS is introducing Database Savings Plans, a new consumption model intended to help customers maintain cost efficiency while retaining flexibility across their database services and deployment options.
Additionally, new capabilities for Amazon RDS for SQL Server and Oracle were rolled out, including Developer Edition support for SQL Server and support for M7i/R7i instances with optimized CPUs. Storage options for these services are also expanding to support up to 256 TiB.
Announcing AWS AI Factories for on-premises sovereign AI
AWS introduced a new offering called AWS AI Factories to cater to enterprises and governments with strict data sovereignty and regulatory requirements. This product allows these customers to run AWS’s full AI systems, including compute, storage, and AI services like Bedrock and SageMaker, within their own existing data centers.
AWS AI Factories are a collaboration with Nvidia and can be provisioned with either Nvidia’s latest Blackwell GPUs or Amazon’s new Trainium3 chips. The service is designed to operate like a private AWS Region, providing secure, low-latency access to AI infrastructure while ensuring data remains on-premises. AWS handles the deployment and management of the integrated infrastructure.
Enhancements to Amazon Bedrock AgentCore for agent development
Amazon Bedrock AgentCore, the platform for building and deploying generative AI applications, received major new capabilities aimed at helping developers build and scale production-ready AI agents with greater control.
-
Policy in AgentCore: Allows developers to set boundaries for agent actions using natural language, enabling stronger governance. For example, a boundary could be set so the agent can automatically issue refunds up to $100 but must bring a human into the loop for anything larger.
-
AgentCore Evaluations: A suite of 13 pre-built evaluation systems to monitor agent behavior for factors including correctness, safety, and tool selection accuracy.
-
AgentCore Memory: Introduces new episodic functionality, helping agents develop a log of information on users over time (like flight or hotel preferences) to inform future decisions.
AWS Interconnect: Multicloud in partnership with Google Cloud
In a significant multicloud development, AWS rolled out AWS Interconnect – multicloud in partnership with Google Cloud, leveraging Google’s Cross-Cloud Interconnect. This new networking service is designed to allow customers to build private, high-bandwidth connections between the two rival cloud platforms with high levels of automation.
The solution enables customers to provision dedicated bandwidth on demand and establish connectivity in minutes, rather than weeks. The partnership is based on a jointly developed open specification for network interoperability and is engineered for high resiliency by leveraging quad-redundancy and MACsec encryption for enhanced security.
Amazon starts testing “Ultra-Fast” 30-minute deliveries
Amazon announced the launch of a new “ultra-fast” delivery option in select areas of Seattle and Philadelphia, promising delivery in 30 minutes or less for a range of items including groceries, cosmetics, electronics, and essentials.
This new service, called Amazon Now, is integrated into the main Amazon app and website.
-
Prime members can access this option for a delivery fee starting at $3.99 per order.
-
Non-Prime members pay $13.99.
-
A $1.99 small-basket fee applies to orders under $15.
To achieve this speed, Amazon is utilizing smaller, specialized fulfillment facilities strategically placed close to where customers live and work, reducing the travel distance for delivery partners. This domestic pilot follows the October launch of a similar initiative in the United Arab Emirates, where the company offers 15-minute deliveries.
