Amazon Web Services, which has spent years developing its own AI training hardware, has introduced its latest chip, Trainium3, boasting major performance gains and efficiency improvements.
The announcement was made Tuesday at AWS re:Invent 2025, where the company also teased the next product on its AI roadmap: Trainium4, currently in development and designed to work seamlessly with Nvidia hardware.
AWS officially launched the Trainium3 UltraServer, a new system built around the company’s advanced 3-nanometer Trainium3 chip and proprietary high-speed networking technology. As expected, the third-generation chip — and the system built around it — delivers substantial boosts in both AI training and inference compared to its predecessor.
According to AWS, the new system is over four times faster, offers four times more memory, and significantly increases the compute available for AI applications during periods of peak demand. Thousands of UltraServers can be connected, allowing customers to scale up to one million Trainium3 chips — 10 times more than the previous generation. Each UltraServer accommodates 144 chips.
Equally important, AWS claims the new chips and systems are 40% more energy efficient than before. As global data centers consume increasing amounts of electricity, AWS is positioning its hardware as a way to deliver more compute while using less power.
This efficiency focus also aligns with Amazon’s core cost-cutting philosophy. The company says Trainium3 systems can reduce AI infrastructure costs for cloud customers. Early adopters — including Anthropic, Karakuri, SplashMusic, and Decart — have already leveraged the third-generation chip to significantly lower inference expenses.
AWS also outlined early details about its next-generation chip, Trainium4, which is already under development. The chip promises another major performance leap and, notably, will support Nvidia’s NVLink Fusion, a high-speed interconnect technology.
This compatibility will allow Trainium4-based systems to integrate with and enhance Nvidia GPU setups while still running on Amazon’s more affordable, in-house server rack designs.
The move could prove strategically important. Nvidia’s CUDA platform remains the dominant ecosystem for AI development, and Trainium4’s compatibility may make AWS a more attractive cloud destination for customers building AI applications around Nvidia hardware.
Amazon has not shared a release date for Trainium4. If past launches are any indication, more details are likely to surface at next year’s re:Invent conference.
