AWS re:Invent Day 2: Amazon Launches New AI Models, Trainium3 Chips, and On-Prem AI Factories

Published December 2, 2025

Written by

AWS re:Invent Day 2 delivered major AI updates, including new Nova models, Trainium3 servers, and Bedrock’s largest model expansion.

Image: Amazon

The second day of AWS re:Invent arrived with the same loud energy as opening day, as Amazon pushed deeper into AI hardware, new models, and tools that promise faster building, faster training, and faster modernization.

Here are the top takeaways from the latest round of announcements.

New Nova models and the “open training” approach

Amazon rolled out four new models in the Nova family, all tuned for different AI needs such as advanced reasoning, multimodal tasks, code generation, and agent-style automation.

The company also highlighted Nova Forge, which lets customers work directly with Amazon’s pre-trained checkpoints and combine them with their own data. The idea is to give organizations more control without forcing them to build models from scratch.

Amazon said early adopters are already logging gains. Nova Act has achieved 90% reliability in browser-based UI automation, while teams at Reddit are replacing many small models with a single Nova-powered workflow. Hertz, meanwhile, reported speeding up development by 5× using the Nova Act.

Amazon Bedrock adds its largest set of new models yet

Day two also brought the biggest single expansion of models ever added to Amazon Bedrock. Eighteen new open-weight models are now available. These include new releases from Mistral AI, as well as models from Google, MiniMax, NVIDIA, OpenAI, and others.

The company’s own Nova 2 models joined the lineup as well, offering improved price-performance across reasoning and conversational workloads. The company continues pushing Bedrock as the place for customers who want multiple model options without rewriting their applications every time a new model appears.

Trainium3 UltraServers aim to shrink training timelines

AWS introduced the EC2 Trn3 UltraServers, powered by its first 3nm AI chip and featuring up to 144 Trainium3 chips in one integrated system. The company said the new hardware offers 4.4× more compute performance, 4× better energy efficiency, and 3× higher throughput per chip than the previous generation, reducing training cycles from months to weeks.

Several customers are already using the hardware. Teams from Anthropic, Karakuri, Metagenomi, NetoAI, Ricoh, and Splash Music are reporting up to 50% reductions in training and inference costs. Decart is seeing 4× faster inference for real-time generative video while spending half the cost of comparable GPU setups. Amazon Bedrock is also running production workloads on Trainium3.

AWS AI Factories bring cloud-scale AI into customer data centers

AWS also expanded its strategy for bringing cloud-grade AI compute directly into customer environments. The new AI Factories offering lets enterprises and government agencies deploy AWS-managed AI infrastructure inside their own data centers, “combining NVIDIA GPUs, Trainium chips, AWS networking, and AI services like Bedrock and SageMaker AI.”

The company framed the move as a practical option for organizations with sovereignty requirements or large on-prem investments. AWS noted that its longtime collaboration with NVIDIA is deepening under the program. Early adopters include HUMAIN in Saudi Arabia, which is building an “AI Zone” equipped with up to 150,000 AI chips.

New NVIDIA-powered EC2 instances

In another hardware move, AWS introduced the P6e-GB300 UltraServers, built around the NVIDIA GB300 NVL72 architecture. These instances offer the highest memory and compute performance within an UltraServer on AWS and are aimed at customers running large-scale inference, including trillion-parameter models. The new systems also run on the AWS Nitro System and integrate with tools like Amazon EKS.

AWS Transform promises faster modernization and less tech debt

Modernization is often the silent tax inside big organizations, and Amazon says it wants to cut that burden dramatically. AWS Transform received new agentic AI upgrades that help rework entire stacks far faster than manual rewrites.

The company said Transform can now modernize Windows-based stacks, including .NET, SQL Server, UI frameworks, and deployment layers. Amazon claims some companies can eliminate up to 70% of maintenance and licensing costs. Air Canada was pointed out as an early success story after modernizing thousands of Lambda functions and achieving an 80% reduction in time and cost compared to manual work.

S3 Vectors becomes generally available with massive scale

Another major Day-2 announcement was the general availability of Amazon S3 Vectors, a new capability that lets AI systems store and query vectors natively inside S3. The service now scales to two billion vectors per index — a 40× jump from its preview limit — and supports 20 trillion vectors per bucket.

AWS said S3 Vectors delivers 2–3× faster performance while cutting costs by as much as 90%. It is integrated with Amazon Bedrock and OpenSearch and is powering retrieval and agent workloads for customers such as BMW and Twilio.

Adobe strengthens its AI partnership with AWS

Adobe CEO Shantanu Narayen announced an expanded collaboration with AWS, focusing on model training, AI agents, cloud scalability, and creator tools. The partnership aims to bring moregenerative AI capabilities into Adobe’s ecosystem for marketers, businesses, and creative professionals.

Sony taps AWS to power enterprise AI and fan engagement

Another major customer spotlight came from Sony, which is scaling its AI vision across both internal operations and consumer entertainment.

Sony’s enterprise platform, which supports 57,000 employees, uses Amazon Bedrock AgentCore to handle 150,000 AI requests per day for content drafting, trend analysis, and information retrieval — an activity Sony expects to grow by a factor of 300.

“Through our long-standing partnership with AWS, we are now able to fully harness the power of data and AI to forge new bonds between fans and creators,” said Tsuyoshi Kodera, chief digital officer, corporate executive officer of Sony Group Corporation.

AWS and Bonterra boost nonprofit giving with new donation experience

AWS teamed with Amazon Business Solution Donation Driver and Bonterra’s GiveGab platform to launch a new donation experience built for mobile devices and modern nonprofit engagement. Debuting in time for Giving Tuesday and showcased at re:Invent, the solution is designed to streamline the donor experience for nonprofits using GiveGab on AWS.

Zendesk and AWS sign a strategic collaboration agreement

Zendesk announced a new strategic collaboration agreement with AWS to accelerate AI-powered customer experience tools.

By combining Amazon Connect voice technology, conversational analytics, and Zendesk’s AI tools, which run on Amazon Bedrock, the companies aim to replace fragmented customer support stacks with a unified platform built around voice, digital channels, and automation.

AWS announces 2025 Imagine Grant winners

AWS paused to celebrate the impact of cloud technology beyond the enterprise and announced the winners of its 2025 Imagine Grant program. This public grant recognizes non-profit organizations using cloud technology to tackle complex global issues.

The 2025 winners, who spanned three categories, including children’s health and global outreach, include the American Association for Cancer Research, The Scripps Research Institute, and the Institute for Systems Biology.

“Through the Imagine Grant program, we’re seeing organizations embrace cloud technology in ways that fundamentally reshape how they deliver on their missions. From scaling their impact to reaching underserved communities, these nonprofits are showing us what’s possible when vision meets innovation,” said Rick Buettner, managing director of Global Nonprofits at AWS.

Missed the first day of festivities? Check out TechRepublic’s coverage of Day One of AWS re:Invent.

Aminu Abdullahi

Aminu Abdullahi is an experienced B2B technology and finance writer. He has written for various publications, including TechRepublic, eWEEK, Enterprise Networking Planet, eSecurity Planet, CIO Insight, Enterprise Storage Forum, IT Business Edge, Webopedia, Software Pundit, Geekflare and more.