Amazon Web Services (AWS) has announced plans to build what it claims will be the world’s largest artificial intelligence (AI) supercomputer. Dubbed Project Rainier, the supercomputer is expected to be operational by 2025.
The announcement was made on 3 December at the AWS re:Invent conference in Las Vegas.
The computer will be powered by AWS’s “Trainium” chips, which are custom-designed AI accelerators intended to deliver high performance while reducing costs for AI training and inference. The second-generation Trainium2 chips are expected to quadruple the performance of their predecessors, and are said to provide a cost-effective alternative to Nvidia’s GPUs.
Ethical AI
The supercomputer will support the development of advanced AI models for Anthropic, an AI safety and research company in which Amazon has invested US$4bn, with another US$4bn of funding announced on 22 November.
Anthropic was founded in 2021 by siblings Dario and Daniela Amodei, both former senior employees of OpenAI. The company has developed the “Claude” family of large language models optimised for different applications, with tasks ranging from customer interaction and rapid data processing, to problem-solving, advanced coding and content creation.
Claude implements “Constitutional AI”, which is a methodology to align AI systems with human values and ethical principles through a predefined “constitution”. This set of rules explicitly defines ethical boundaries, ensuring that the AI generates responses that are helpful, honest and harmless.
This approach distinguishes Anthropic from OpenAI, which has faced scrutiny over its profit-oriented structure and transparency regarding safety measures.
Claude is primarily trained and deployed on AWS infrastructure (although available on other platforms, such as Google Cloud’s Vertex AI).
Amazon Nova
In addition to hardware advancements, AWS announced “Amazon Nova”, a new family of generative AI models for enterprise customers, along with a suite of tools that allow businesses to create smaller, cost-effective AI models, manage AI agents, and ensure output accuracy through logical validation.
The Nova models support over 200 languages, and the lineup includes Nova Canvas for image generation and Nova Reel for video generation.
Amazon claims the models are at least 75% less expensive than the best-performing models in their respective classes.
Industry dominance
Industry analysts view Amazon’s initiatives as a challenge to Nvidia’s dominance in the AI hardware market. Patrick Moorhead, CEO of Moore Insights & Strategy, remarked, “This innovation shows Nvidia is no longer the only game in town for AI training.”
Shares in Amazon.com are up 2.01% to US$217.74, and up 50.32% over the past year.