Amazon Web Services has introduced a new version of its AI training chip, called Trainium3, featuring impressive specifications. The cloud provider made the announcement at its AWS re:Invent 2025 conference, also teasing the next product on its roadmap: Trainium4, which is already in development and will be designed to work with Nvidia’s chips.
AWS formally launched the Trainium3 UltraServer, a system powered by the company’s state-of-the-art 3-nanometer Trainium3 chip and its own networking technology. According to AWS, this third-generation chip and system offer significant performance improvements for both AI training and inference compared to the second generation.
The company states the new systems are more than four times faster and come with four times more memory, not only for training but also for delivering AI applications during peak demand. Furthermore, thousands of these UltraServers can be linked together to provide an application with access to up to 1 million Trainium3 chips, which is ten times the scale of the previous generation. Each UltraServer can host 144 chips.
Perhaps more importantly, AWS claims the chips and systems are also 40% more energy efficient than their predecessors. As the world races to build larger data centers with enormous power demands, AWS is focusing on creating systems that consume less electricity. This is directly in AWS’s interest, and in keeping with Amazon’s cost-conscious approach, the company promises these systems will also save money for its AI cloud customers.
AWS customers, including Anthropic, Japan’s LLM Karakuri, Splashmusic, and Decart, have already been using the third-generation chip and system, reportedly achieving significant reductions in their inference costs.
AWS also provided a glimpse of its roadmap with Trainium4, which is already in development. The company promises the chip will deliver another major performance improvement and will support Nvidia’s NVLink Fusion high-speed chip interconnect technology. This means Trainium4-powered systems will be able to interoperate and extend their performance with Nvidia GPUs while still utilizing Amazon’s lower-cost server rack technology.
It is also noteworthy that Nvidia’s CUDA has become the de facto standard supported by all AI applications. Trainium4-powered systems could make it easier to attract major AI applications built for Nvidia GPUs to run on Amazon’s cloud. Amazon did not announce a timeline for Trainium4, but if it follows previous rollout patterns, more details will likely be shared at next year’s conference.