Amazon AWS Officially Launches its New Trainium3 AI Chip: A Comparatively Cheaper Alternative to NVIDIA GPUs

At the ongoing AWS re:Invent 2025 event (1st ~ 5th December 2025), Amazon’s AWS officially launched its Trainium3 AI chip. Notably, it has been introduced by positioning the latest AI chip as a much cheaper alternative when compared to the NVIDIA GPUs.

Here’s more about it.

Trainium3 AI Chip – Officially Launched

As mentioned, the Trainium3 AI chip has now been brought out by Amazon Web Services (AWS) and this state of the art 3nm chipset is used to power the Trainium3 UltraServers, combined with the brand’s networking tech. In terms of performance, up to 4.4x higher performance is being offered, as well as 4x more memory bandwidth. In comparison to the previous Trainium2 UltraServers, 4x better performance-per-watt is also being claimed on it. Also, the UltraServer (each) will be able to host 144 chips (362 FP8 PFLOPS) as per the brand, and over 1 million Trainium3 chips can be linked together in the UltraClusters – which is 10x times more than the previous generation.

Speaking more, the Trainium-based systems are said to cut down the training and inference costs by around 50% when compared to other GPU clusters that are used on large models. Furthermore, 2x more computing performance and 40% more energy efficiency are also notable aspects, which assure both faster training and peak AI app delivering demand. Thus, it is pointed out to be a better and cheaper alternative to the NVIDIA GPUs.

Already customers like Anthropic, LLM Karakuri, Decart, and SplashMusic have been making use of the latest Trainium3 AI chip, and Amazon states that these brands were able to lower their inference costs. The goal for Amazon is not just to make systems that consume less, but that also offers lower costs for its customers too.