Amazon Web Services, having developed its own artificial intelligence training chips for several years, has recently unveiled a new iteration called Trainium3, featuring substantial performance characteristics.

The cloud services provider revealed this information on Tuesday during AWS re:Invent 2025 and also previewed its forthcoming AI training product, Trainium4, currently under development and designed for compatibility with Nvidia’s processors.

At its yearly technology conference, AWS officially launched Trainium3 UltraServer, a system utilizing the company’s advanced, 3-nanometer Trainium3 chip alongside its internally developed networking technology. According to AWS, this third-generation chip and system demonstrate considerable improvements in both AI training and inference capabilities compared to its predecessor.

AWS states that the system delivers over four times the speed and possesses four times the memory capacity, benefiting not only the training process but also the delivery of AI applications during periods of high demand. Furthermore, thousands of UltraServers can be interconnected to provide an application with access to up to 1 million Trainium3 chips—a tenfold increase over the prior generation. The company specifies that each UltraServer can accommodate 144 chips.

Significantly, AWS also reports that the chips and systems are 40% more energy efficient than the previous generation. As the industry focuses on constructing larger data centers with immense power requirements, AWS is prioritizing the creation of systems that reduce energy consumption.

This approach aligns with AWS’s core business objectives, and the company asserts that these systems will also result in cost savings for its AI cloud customers.

Several AWS customers, including Anthropic (in which Amazon has invested), Japan’s LLM Karakuri, SplashMusic, and Decart, have already utilized the third-generation chip and system, achieving substantial reductions in their inference expenses, as stated by Amazon.

AWS also shared a glimpse of its future chip development, Trainium4, which is currently in progress. AWS has committed to delivering another significant performance enhancement with this chip and incorporating support for Nvidia’s NVLink Fusion high-speed chip interconnect technology.

This integration will enable AWS Trainium4-powered systems to work in conjunction with and enhance performance alongside Nvidia GPUs, while still leveraging Amazon’s cost-effective, internally designed server rack infrastructure.

It is important to note that Nvidia’s CUDA (Compute Unified Device Architecture) has become the industry standard for major AI applications. The Trainium4-powered systems could potentially attract more AI applications originally built for Nvidia GPUs to Amazon’s cloud platform.

Amazon has not yet announced a release timeframe for Trainium4. Based on past launch schedules, further details regarding Trainium4 are anticipated at next year’s conference.

Stay updated with all of TechCrunch’s reporting on this annual enterprise technology event here.

Sponsored: View re:Invent event broadcasts

Explore the newest announcements covering topics like autonomous AI, cloud systems, data protection, and a wide range of other subjects presented at the conference