Amazon Web Services has unveiled its latest AI chip, Trainium3, alongside a powerful new system called UltraServer, marking a significant advancement in its push to expand its AI infrastructure capabilities. The announcement came Tuesday at AWS re:Invent 2025, where AWS highlighted the chip’s cutting-edge 3-nanometer design and integration with proprietary networking technology. The Trainium3 UltraServer system delivers substantial improvements over its predecessor. According to AWS, it is more than four times faster and features four times the memory capacity, enhancing both AI training and real-time inference performance. This boost is particularly valuable during peak demand, enabling faster delivery of AI-powered applications. The system can scale dramatically—thousands of UltraServers can be interconnected to support up to 1 million Trainium3 chips, a tenfold increase over the previous generation. Each UltraServer hosts 144 chips, underscoring the system’s massive parallel processing potential. Energy efficiency is another key focus. AWS claims the new system is 40% more energy efficient than the prior generation, a critical advantage as data centers worldwide grapple with rising power demands. By reducing energy consumption, AWS aims to lower operational costs for both itself and its customers—aligning with Amazon’s long-standing cost-conscious strategy. Early adopters, including Anthropic (a company in which Amazon holds an investment), Japan’s LLM Karakuri, SplashMusic, and Decart, have already begun using the Trainium3 system and reported significant reductions in inference costs, AWS said. Looking ahead, AWS teased the next milestone in its AI chip roadmap: Trainium4, already in development. The upcoming chip is expected to deliver another major leap in performance and will support Nvidia’s NVLink Fusion interconnect technology. This compatibility means Trainium4-powered systems will be able to seamlessly integrate with Nvidia GPUs, combining Amazon’s cost-effective server architecture with the widespread software ecosystem built around Nvidia’s CUDA platform. This move could help AWS attract AI developers and applications that have traditionally relied on Nvidia hardware, making Amazon’s cloud a more viable alternative for large-scale AI workloads. While no official release date has been announced, if AWS follows its historical rollout pattern, Trainium4 is likely to be unveiled at next year’s re:Invent conference. The developments underscore Amazon’s growing ambition to compete directly with Nvidia in the AI hardware space—offering not just cloud services, but a full-stack solution that includes custom chips, networking, and cross-platform compatibility.