Amazon Web Services (AWS) unveiled its plans to build a massive AI supercomputer called “Ultracluster” powered by its proprietary Trainium chips during its annual re:Invent conference on Tuesday.
Named Project Rainier, the chip cluster will be ready in 2025 and used by AI startup Anthropic, in which Amazon has invested $8 billion to date. It will be five times more powerful than Anthropic’s current training cluster and rival some of the world’s largest AI model training systems.
AWS also rolled out new Trainium2 chips and Ultraserver, a new server that has 64 of its own interconnected chips.
By designing custom silicon in-house, AWS aims to lower AI costs for customers and reduce reliance on Nvidia, which commands ~95% of the AI chips market, per IDC’s December research. In 2024, the AI semiconductors market was an estimated $117.5 billion, according to International Data Corp’s December research. It is estimated to reach $193.3 billion by 2027.
Apple is already testing AWS’s Trainium2 chips and expects to save costs by roughly 50%. AI coding startup Poolside said it is witnessing 40% in cost savings after using AWS hardware compared to running its AI models on Nvidia’s GPUs.