AWS Expands Its In-House AI Silicon With Trainium3

Amazon Web Services (AMZN) has introduced Trainium3, its first 3nm AI training chip, marking a major leap in AWS’s in-house semiconductor roadmap. The chip now powers the new Trainium3 UltraServers, which are immediately available to customers.

AWS said Trainium3 delivers:

4× higher energy efficiency
Nearly 4× memory bandwidth
compared to the previous Trainium2 UltraServers.

Customers Report Up to 50% Lower Training Costs

Early Adopters Benefit From Lower Costs

Companies building advanced AI systems — including Anthropic, Karakuri, Metagenomi, NetoAI, Ricoh, and Splash Music — have already seen:

Up to 50% lower training costs versus competing GPU solutions
Improved performance for large-scale model training

AWS highlighted additional success stories, including Decart, an AI lab focused on real-time video and image generation. With Trainium3, Decart achieved:

4× faster frame generation
At half the GPU cost

AWS Already Developing Trainium4 With Nvidia NVLink Fusion

Next-Gen Partnership for Rack-Scale AI

AWS confirmed it is in progress on Trainium4, with a key feature being integration with Nvidia’s NVLink Fusion — the latest high-speed interconnect for multi-chip and multi-server communication.

The goal: create a seamless, rack-scale AI infrastructure where:

Trainium4,
Graviton CPUs, and
Elastic Fabric Adapter (EFA)

all work together inside common Nvidia MGX racks, supporting mixed GPU and Trainium deployments efficiently.

This marks a rare deep collaboration between AWS and Nvidia to integrate custom cloud silicon with Nvidia’s high-speed networking technology.

Strategic Implications

The launch strengthens AWS's position against hyperscaler rivals Google (TPU) and Microsoft/OpenAI (Azura AI accelerators), as cloud providers increasingly shift toward custom silicon to lower AI training costs, reduce GPU reliance, and improve vertical integration.