Amazon Unveils New Trainium3 AI Chip and Announces Trainium4
Date: 2025-12-05 Author: Gabriel Deangelo Categories: BUSINESS
The Amazon Web Services (AWS) team unveiled the next generation of its Trainium3 AI chips and shared plans for Trainium4, which will be compatible with Nvidia GPUs. The announcement took place at the AWS re:Invent 2025 conference.
The main innovation is the Trainium3 UltraServer. The chip is manufactured using 3-nanometer technology and is complemented by an enhanced AWS network architecture. The new system demonstrates more than four times the training speed and memory capacity of the previous version. This makes it suitable not only for training AI models but also for handling peak workloads in real time.
UltraServer servers can be combined into large clusters of up to a thousand machines, creating configurations with up to a million Trainium3 chips—ten times more than the previous generation. Each server contains 144 chips, and power consumption has been reduced by 40%. AWS notes that this is especially important as data center workloads increase.
The company emphasizes that the new technologies are aimed at accelerating computing and reducing customer costs. Trainium3 is already being used by companies such as Anthropic, Karakuri, SplashMusic, and Decart, which report reduced costs for inferring AI models.
AWS also shared its plans for Trainium4. The new chip will support Nvidia NVLink Fusion interconnects, enabling the integration of AWS processors and Nvidia graphics cards into a single computing infrastructure. Compatibility with NVLink Fusion will enable the seamless migration of large models and applications developed for CUDA to the AWS Cloud while maintaining server cost-effectiveness.
The commercial launch date for Trainium4 has not yet been announced, but AWS typically provides more details at its annual conference. Support for Nvidia CUDA will allow AWS to offer hybrid training and inference solutions that meet industry standards.
Against the backdrop of these announcements, it's worth noting that Nvidia recently reported record revenue of $57 billion for Q3 2025, highlighting the relevance of integration with its technology for major cloud players.
Thus, AWS demonstrates its commitment to strengthening its position in the high-performance AI market while simultaneously reducing customer costs and improving the energy efficiency of its data centers.
