Amazon used re:Invent 2025 to focus on scaling its in-house hardware, lowering model costs with the Nova 2 lineup, and pushing its AI tools toward more autonomous behavior instead of simple assistant-style workflows.
To reduce reliance on Nvidia and lower AI workload costs, AWS introduced the third generation of its accelerators. The Trainium3 UltraServer, built on a 3-nanometer process, is designed to offer four times the speed and memory of the previous version with a 40 percent improvement in energy efficiency.
The larger shift is scale. AWS now supports clusters of up to one million Trainium3 chips, an approach already used by customers such as Anthropic.
Amazon also signaled a change for the next generation. The upcoming Trainium4 will support Nvidia's NVLink Fusion, enabling hybrid systems where Amazon's lower-cost hardware can work directly with Nvidia GPUs. This reflects Nvidia's continued dominance of the CUDA ecosystem and gives AWS more flexibility in mixed environments.
Price war with Nova 2
