At AWS re:Invent 2025, Amazon Web Services introduced the Trainium3 AI training chip and the Trainium3 UltraServer. The system features a 3-nanometer chip with superior performance and memory. Each UltraServer can hold 144 chips, with the capability to connect thousands for up to a million Trainium3 chips working together. AWS noted a 40% energy efficiency improvement to meet rising data center power demands. Customers like Anthropic, Karakuri, and others are already benefiting from lower inference costs. AWS also teased the Trainium4 chip, which will work alongside Nvidia GPUs, making it attractive for developers in Nvidia’s CUDA ecosystem.

Read more at Nasdaq: AWS Unveils Trainium3 UltraServer, Teases More Powerful Trainium4 With Nvidia Support