Amazon releases an impressive new AI chip and teases a Nvidia-friendly roadmap  

Amazon Web Services (AWS) has unveiled its latest AI training chip, Trainium3, along with the Trainium3 UltraServer system at its AWS re:Invent 2025 conference. Built on a 3-nanometer process, Trainium3 delivers significant improvements over its predecessor, offering more than four times the speed and memory capacity for AI training and inference. Each UltraServer can host 144 chips, and thousands of these servers can be linked to scale up to one million Trainium3 chips, representing a tenfold increase from the previous generation. Additionally, the new chips are 40% more energy efficient, aligning with AWS’s goal to reduce operational costs and energy consumption while providing cost savings to AI cloud customers. Early adopters such as Anthropic, Karakuri, Splashmusic, and Decart have already reported substantial reductions in inference costs using Trainium3. Looking ahead, AWS teased the development of Trainium4, which promises another major performance boost and will support Nvidia’s