Amazon Web Services has been building its own AI training chips for years, and has announced a new version known as Trainium3 with some impressive specs.
Announced at AWS re:Invent 2025 on Tuesday, the cloud provider also teased the next product on its AI training product roadmap: Trainium4. Trainium4 is already in development and will be able to work with Nvidia chips.
AWS used its annual technology conference to officially announce the Trainium3 UltraServer, a system powered by the company’s cutting-edge 3-nanometer Trainium3 chip and proprietary networking technology. As you might expect, third-generation chips and systems offer significant improvements in AI training and inference performance compared to second-generation chips, according to AWS.
AWS says the system is more than 4x faster and has 4x more memory, making it suitable for not only training but also delivering AI apps at peak times. Additionally, you can link thousands of UltraServers to provide up to 1 million Trainium3 chips to your app, 10x more than the previous generation. The company says each UltraServer can host 144 chips.
Perhaps more importantly, AWS says the chips and systems are also 40% more energy efficient than the previous generation. As the world races to build massive data centers that provide astronomical gigawatts of power, data center giant AWS is creating a system that lets you drink less, not more.
Obviously, it is in AWS’s direct interest to do so. But in classic cost-conscious fashion, Amazon promises these systems will also save AI Cloud customers money.
Amazon says AWS customers such as Anthropic (which Amazon also owns), Japan’s LLM Karakuri, SplashMusic, and Descart are already using third-generation chips and systems, significantly reducing inference costs.
tech crunch event
san francisco
|
October 13-15, 2026
AWS also provided a bit of a roadmap for its upcoming chip, Trainium4, which is already in development. AWS promised that the chip would deliver even more significant performance improvements and support Nvidia’s NVLink Fusion high-speed chip interconnect technology.
This means that systems powered by AWS Trainium4 can interoperate with Nvidia GPUs to scale performance while using Amazon’s own low-cost server rack technology.
It’s also worth noting that Nvidia’s CUDA (Compute Unified Device Architecture) has become the de facto standard that all MajorAI apps are built to support. Systems powered by Trainium4 could make it easier to bring large-scale AI apps built with Nvidia GPUs in mind to Amazon’s cloud.
Amazon has not announced a schedule for Trainium4. If Amazon follows its rollout schedule so far, we’ll likely hear more about Trainium4 at next year’s conference.
Check out all of our coverage of TechCrunch’s annual enterprise technology event here.
