AWS Unveils Trainium3 AI Chip, Teases Nvidia-Friendly Trainium4

Amazon Web Services (AWS), a long-time developer of its own AI training chips, has introduced its latest version, Trainium3, featuring impressive specifications designed to enhance artificial intelligence workloads. The cloud provider made the announcement at its AWS re:Invent 2025 conference, where it also offered a glimpse into its future roadmap by teasing Trainium4, a forthcoming chip engineered for seamless interoperability with Nvidia's widely used GPUs.

Trainium3: Powering Next-Gen AI Workloads

AWS formally launched the Trainium3 UltraServer, a system powered by the company's state-of-the-art 3-nanometer Trainium3 chip and proprietary networking technology. As anticipated, this third-generation chip and system deliver significant performance enhancements for AI training and inference compared to its predecessor.

According to AWS, the new systems are more than four times faster and offer four times more memory, not only for training large language models but also for efficiently delivering AI applications during peak demand. For massive-scale operations, thousands of UltraServers can be interconnected, providing an application with access to up to 1 million Trainium3 chips—a tenfold increase over the previous generation. Each UltraServer unit is capable of hosting 144 chips.

Beyond raw power, AWS emphasizes the energy efficiency of its new hardware, stating that Trainium3 chips and systems are 40% more energy-efficient than the prior generation. In an era where the world races to build bigger data centers powered by astronomical gigawatts of electricity, AWS is focused on creating systems that consume less power. This move aligns directly with AWS's interests and its characteristic Amazon cost-conscious approach, promising substantial cost savings for its AI cloud customers.

Early Adoption and Cost Savings

Several prominent AWS customers have already begun utilizing the third-generation chip and system. These include Anthropic (in which Amazon is also an investor), Japan's LLM Karakuri, Splashmusic, and Decart. These companies have reported significant reductions in their AI inference costs by leveraging Trainium3.

Trainium4: A Future with Nvidia Integration

AWS also presented a roadmap for its next-generation chip, Trainium4, which is currently under development. AWS promises that Trainium4 will offer another substantial leap in performance and, crucially, will support Nvidia's NVLink Fusion high-speed chip interconnect technology.

This strategic move means that future AWS systems powered by Trainium4 will be able to interoperate and extend their performance capabilities with Nvidia GPUs, all while utilizing Amazon's cost-effective, homegrown server rack technology. This compatibility is particularly significant given that Nvidia's CUDA (Compute Unified Device Architecture) has become the de facto standard for most AI applications. By enabling easier integration, Trainium4-powered systems could help AWS attract more large-scale AI applications that were originally built with Nvidia GPUs in mind to Amazon's cloud platform.

While AWS did not provide a specific timeline for Trainium4's release, it is anticipated that more information will be shared at next year's re:Invent conference, following Amazon's typical product rollout schedule.