The Trainium3 chip, set to become available next year, will deliver up to twice the performance of its predecessor while improving energy efficiency by 40%. It may be crucial as the company seeks to challenge Nvidia’s dominance of the chip industry.
Trainium3 marks AWS’s first chip using a three-nanometer semiconductor manufacturing process.
AWS CEO Matt Garman highlighted the chip’s capabilities, including its performance with Meta’s Llama open-source model. The company also introduced UltraServers, which connect 64 Trainium2 chips into a single powerful server designed to handle large language models with trillions of parameters.
Also at the event, Apple’s senior director of machine learning and AI Benoit Dupin delivered a surprise keynote highlighting Apple’s extensive use of AWS services across multiple product lines — a high-profile endorsement of AWS’s cloud and AI infrastructure capabilities. The technology giant is currently evaluating AWS’s Trainium 2 AI training chip, with early tests suggesting potential efficiency improvements up to 50% when pre-training AI models.
In addition, a preview of Project Rainier revealed plans for an UltraCluster that could potentially access hundreds of thousands of Trainium2 chips. The project is being developed in partnership with AI startup Anthropic.
AWS’s announcements at re:invent represents a significant step in custom AI chip development—challenging existing market leaders and offering more efficient alternatives for machine learning infrastructure.
