Amazon Web Services (AWS) has just announced a major step forward in its self-diseased AI chip development strategy when launching Trainium3 and the Trainium3 Ultra Server system at the AWS re:Invent 2025 event.
This is the company's most powerful 3 nanometric chip generation, optimized for both training and large-scale AI model reasoning.
According to AWS, Trainium3 is four times more efficient than the previous generation, and has four times more memory, suitable for AI models that require complex processing.
The Ultra Server is an accompanying system, which can contain up to 144 chips and allow thousands of server clusters to be connected together, reaching a maximum of one million Trainium3 chips running in parallel, 10 times the expansion capability of the previous generation.
One highlight is the ability to save energy when Trainium3 reduces power consumption by up to 40% compared to its predecessor chip line.
As data centers consumed more and more electricity, AWS said that power optimization brings double benefits, helping to reduce Amazon's operating costs and helping customers use cloud AI services more economically.
Many businesses such as Anthropic, LLM Karakuri (Japan), Splashmusic or Decart have tested a new generation of chips and recorded a significant decrease in reasoning costs. This shows that the Trainium3 system is not only technically strong but also attractive in terms of cost.
AWS also revealed for the first time Trainium4, the next generation of chips in development. Most notably, Trainium4 will support NV Link Fusion, Nvidia's high-speed connection technology.
This means that the Trainium4 chip can work in parallel and expand performance with the Nvidia GPU, creating a great competitive advantage in the AI infrastructure market that is heavily dependent on CUDA.
Although the launch date of Trainium4 has not been announced, coaches are likely to share more information at next year's re:Invent event, continuing to strengthen Nvidia's position as a formidable competitor in the AI ecosystem.
