Amazon Web Services (NASDAQ:AMZN), which has spent several years designing its own silicon for AI training, rolled out the latest generation of its accelerator lineup on Tuesday: Trainium3, a major step forward in performance and efficiency for large-scale AI workloads.
The announcement came during AWS re:Invent 2025, where the cloud giant also previewed the next product on its roadmap — Trainium4, a future chip already under development that will be able to work alongside Nvidia hardware.
Trainium3 UltraServer officially launches
During its keynote, AWS introduced the Trainium3 UltraServer, a new system built around Amazon’s custom 3-nanometer Trainium3 processor and its proprietary networking architecture. As expected, the third-generation chip delivers substantial gains over Trainium2.
AWS said Trainium3-based systems offer:
- Over 4× more performance
- 4× the memory capacity
- Improvements for both model training and real-time inference
- Support for scaling to enormous cluster sizes
Clusters can now link thousands of UltraServers, supporting up to 1 million Trainium3 chips—ten times the scale of the prior generation. Each UltraServer packs 144 Trainium3 chips, giving customers access to large AI clusters at cloud scale.
A focus on efficiency as power demand explodes
One of the most striking improvements is power efficiency. AWS claims Trainium3 delivers 40% better energy efficiency than the previous version — an increasingly important metric as AI data centers strain global electricity grids.
Lower energy use also aligns with Amazon’s cost-conscious culture. The company emphasized that customers will see lower AI training and inference costs, not just reduced power consumption.
Early adopters — including Anthropic, Japan-based LLM provider Karakuri, Splashmusic, and Decart — have already deployed Trainium3 systems and reported meaningful reductions in inference spending, AWS said.
Trainium4 roadmap: Designed to work with Nvidia
AWS also revealed a high-level roadmap for Trainium4, confirming the new chip will:
- Deliver a “significant jump” in performance
- Support Nvidia’s NVLink Fusion interconnect technology
This is notable because it will allow AWS server racks to integrate Trainium4 accelerators with Nvidia GPUs, blending Nvidia’s dominant AI ecosystem with Amazon’s lower-cost custom infrastructure.
Given the industry’s reliance on Nvidia’s CUDA platform — effectively the standard for modern AI — AWS’s nod toward compatibility could help attract customers whose models were originally built for Nvidia hardware.
Amazon did not provide a release date for Trainium4, but based on prior launch patterns, more details are likely to surface at re:Invent next year.