Amazon releases an impressive new AI chip and teases a Nvidia-friendly roadmap

Amazon Web Services (AWS) has officially launched its latest AI training chip, Trainium3, during the annual AWS re:Invent 2025 event. This new chip, developed over several years, boasts cutting-edge specifications that promise significant advancements in AI training and inference. The introduction of the Trainium3 UltraServer, which utilizes the state-of-the-art 3-nanometer Trainium3 chip, marks a notable upgrade over its predecessor. According to AWS, this third-generation chip and system deliver over four times the performance and memory capacity, essential for both training and deploying AI applications during peak demand. AWS emphasizes that its UltraServers can be interconnected, allowing applications to leverage up to 1 million Trainium3 chips—ten times the capacity of the previous generation. Each UltraServer supports 144 chips, enhancing scalability for AI workloads. In addition to performance enhancements, AWS highlights that the new chips are also designed with energy efficiency in mind, being 40% more efficient than earlier models. This initiative aligns with global efforts to reduce the energy consumption of expansive data centers, as AWS aims to provide cost-effective solutions for its AI cloud customers. Several companies, including Anthropic, Japan’s LLM Karakuri, Splashmusic, and Decart, are already benefiting from the Trainium3 chip, reporting substantial reductions in their inference costs. Looking ahead, AWS has revealed plans for its next chip, Trainium4, which is currently under development. This upcoming chip is expected to enhance performance further and support Nvidia’s NVLink Fusion technology, facilitating better interoperability with Nvidia GPUs while maintaining Amazon's cost-effective server infrastructure. While a specific release timeline for Trainium4 has not been disclosed, AWS is anticipated to provide more updates during next year's conference.