Amazon's Trainium wins over Anthropic, OpenAI, Apple
Against the backdrop of Amazon's $5 billion investment deal with OpenAI, TechCrunch reporters were invited to tour Amazon's Trainium chip R&D lab in Austin. This facility serves as AWS's critical base for countering NVIDIA's monopoly and reducing costs associated with AI training and inference. Currently, Trainium chips have earned high praise from Anthropic, OpenAI, and even Apple; notably, Anthropic's Claude models now run on over one million Trainium 2 chips. The core of this collaboration lies in Amazon providing up to two gigawatts of Trainium computing power to OpenAI. Despite tight production capacity, Trainium is rapidly capturing market share thanks to its significant advantages in performance and cost efficiency. The latest Trainium 3 chips, paired with next-generation Neutron switches, enable efficient inter-chip connectivity, cutting operating costs by approximately 50% compared to traditional cloud services under equivalent performance conditions. More critically, Trainium now supports mainstream open-source frameworks like PyTorch, allowing developers to migrate with minimal code adjustments, thereby substantially lowering switching barriers. Beyond the chips themselves, Amazon has designed complementary server components—including liquid cooling technologies and the Nitro virtual management platform—to further optimize energy efficiency and performance. During the "boot" phase of chip development, teams conduct intensive engineering validation through methods such as onsite soldering and testing to address challenging issues related to heat dissipation and connectivity. Meanwhile, Project Rainier, one of the world's largest AI compute clusters, currently deploys half a million Trainium 2 chips exclusively serving Anthropic. Although OpenAI's involvement has drawn considerable attention, the engineer team remains focused primarily on Anthropic and internal Amazon services while already developing Trainium 4. Amazon CEO Andy Jassy places great expectations on this self-developed chip business, viewing it as a future growth pillar for AWS. As Trainium expands across both training and inference domains, Amazon aims to reshape the landscape of artificial intelligence infrastructure through vertically integrated hardware solutions.
