Amazon Trainium cuts AI costs 30%

Amazon and Anthropic widened their alliance on April 20, with Anthropic reserving up to 5 gigawatts of AWS capacity built around Amazon’s Trainium chips. (anthropic.com) (aboutamazon.com) Anthropic said it will spend more than $100 billion on AWS technologies over the next 10 years and use AWS as its primary training and cloud provider for mission-critical workloads. (anthropic.com) The core hardware is Trainium, Amazon’s in-house artificial intelligence chip, which AWS sells as an alternative to graphics processors from Nvidia for training and running large models. (aws.amazon.com) AWS says its current Trainium2 systems offer 30% to 40% better price performance than GPU-based EC2 P5e and P5en instances for generative artificial intelligence workloads. (aws.amazon.com) (press.aboutamazon.com) That claim matters because cloud customers buy compute in bulk, and a double-digit cost gap can change where model developers train systems that run into the billions or trillions of parameters. (press.aboutamazon.com) Anthropic said new Trainium2 capacity is coming online in the first half of 2026, with nearly 1 gigawatt of combined Trainium2 and Trainium3 capacity due by the end of 2026. (anthropic.com) Amazon said Project Rainier, the Trainium cluster it built with Anthropic, ranks among the world’s largest artificial intelligence compute clusters. Anthropic said it currently uses more than 1 million Trainium2 chips to train and serve Claude. (anthropic.com) (aboutamazon.com) AWS is also pushing the next chip generation. Its Trainium page says Trainium3 is a 3-nanometer chip, and Amazon expects significant Trainium3 capacity to come online in 2026. (aws.amazon.com) (aboutamazon.com) Amazon is not replacing Nvidia across the board. AWS still sells Nvidia-based instances, and Trainium’s price claims are framed against specific AWS GPU products, not every training setup in the market. (press.aboutamazon.com) (aws.amazon.com) But the new Anthropic commitment gives Amazon a large customer willing to lock future model training to Trainium2, Trainium3, Trainium4, and later generations. (anthropic.com)