AWS Unveils Trainium3 UltraServer, Teases More Powerful Trainium4 With Nvidia Support

(RTTNews) - At AWS re:Invent 2025, Amazon Web Services unveiled its newest AI training chip, the Trainium3, along with a powerful new machine called the Trainium3 UltraServer.

This system is built on AWS's advanced 3-nanometer Trainium3 chip and their own networking tech, which reportedly offers over four times the performance and memory of the last generation.

Each UltraServer can hold 144 chips, and AWS claims that you can connect thousands of these servers, allowing for up to a million Trainium3 chips to work together—ten times the previous capacity.

They also pointed out a significant energy efficiency improvement of around 40%, which is crucial as the demand for power in global data centers keeps rising.

Some customers, like Anthropic, Karakuri, SplashMusic, and Decart, are already using the Trainium3 to lower their inference costs. Additionally, AWS teased its next chip, the Trainium4, which is currently in the works.

This new chip will utilize Nvidia's NVLink Fusion interconnect, allowing Trainium4 systems to work alongside Nvidia GPUs while preserving Amazon's cost-effective server design.

This could make AWS a more appealing choice for developers looking to create AI models that fit into Nvidia's CUDA ecosystem.