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

Amazon Web Services, which has been building its ain AI grooming chips for years now, conscionable introduced a caller mentation known arsenic Trainium3 that comes with immoderate awesome specs.

The unreality provider, which made the announcement Tuesday astatine its AWS re:Invent 2025, besides teased the adjacent merchandise successful connected its AI grooming merchandise roadmap: Trainium4, which is already successful the works and volition beryllium capable to enactment with Nvidia’s chips.

AWS utilized its yearly tech league to formally motorboat Trainium3 UltraServer, a strategy powered by the company’s state-of-the art, 3 nanometer Trainium3 chip, arsenic good arsenic its homegrown networking tech. As you mightiness expect, the third-generation spot and strategy connection large bumps successful show for AI grooming and inference implicit the second-generation, according to AWS.

AWS says the systems are much than 4 times faster, with 4 times much memory, not conscionable for training, but for delivering AI apps astatine highest demand. Additionally, thousands of UltraServers tin beryllium linked unneurotic to supply an app with up to 1 cardinal Trainium3 chips — 10 times the erstwhile generation. Each UltraServer tin big 144 chips, according to the company. 

Perhaps much importantly, AWS says the chips and systems are besides 40% much vigor businesslike than the erstwhile generation.  While the satellite races to physique bigger information centers powered by astronomical gigawatts of electricity, information halfway elephantine AWS is trying to marque systems that portion less, not more.

It is, obviously, successful AWS’s nonstop interests to bash so. But successful its classic, Amazon cost-conscience way, it promises that these systems prevention its AI unreality customers money, too.  

AWS customers similar Anthropic (of which Amazon is besides an investor), Japan’s LLM Karakuri, Splashmusic, and Decart person already been utilizing the third-gen spot and strategy and importantly chopped their inference costs, Amazon said. 

AWS besides presented a spot of a roadmap for the adjacent chip, Trainium4, which is already successful development. AWS promised the spot volition supply different large step-up successful show and enactment Nvidia’s NVLink Fusion high-speed spot interconnect technology.  

This means the AWS Trainium4-powered systems volition beryllium capable to interoperate and widen their show with Nvidia GPUs portion inactive utilizing Amazon’s homegrown, lower-cost server rack technology.  

It’s worthy noting, too, that Nvidia’s CUDA (Compute Unified Device Architecture), has go the de facto modular that each AI apps support. The Trainium4-powered systems whitethorn marque it easier to woo large AI apps built with Nvidia GPUs successful caput to Amazon’s cloud.

Amazon did not denote a timeline for Trainium4. If Amazon follows erstwhile rollout timelines, we’ll apt perceive much astir Trainium4 astatine adjacent year’s conference.