A new P63 GPU offering featuring Nvidia’s GB200 and GB300 chips promises major performance and efficiency gains. Garman revealed that OpenAI is already using this infrastructure to train and improve its models.
AI Factories, Trainium, and the Future of Infrastructure
AWS also unveiled AWS AI Factories, which function like private AWS regions deployed directly within customer data centers. These allow organizations to run highly customized AI infrastructure while maintaining control over sensitive data.
Custom silicon remains central to AWS’s strategy. Garman highlighted increasing adoption of Trainium chips for inference workloads, including use by AI services like Perplexity’s Claude. He also detailed Project Rainier, where more than 500,000 Trainium 2 chips are being used to scale Claude’s development.
Next up is Trainium 3, which offers 4.4x more compute than the previous generation, with ultra servers now available. AWS is already working on Trainium 4, promising even greater performance, bandwidth, and memory—though no release date was announced.
