Amazon.com Inc (NASDAQ:AMZN) is doubling down on its in-house AI chips, and chief executive Andy Jassy says the strategy is already delivering.
At AWS re:Invent he revealed that Trainium2, Amazon’s current accelerator, has reached a multibillion-dollar revenue run rate, with more than 1 million chips in production and over 100,000 customers accessing them through the Bedrock AI platform.
A major contributor is Anthropic, which now trains new Claude models on more than 500,000 Trainium2 chips under Project Rainier.
Amazon also introduced Trainium3, claiming four-times faster performance than its predecessor while using less power.
The pitch is aimed at cloud customers burdened by high GPU prices and short supply, with Jassy emphasising Trainium’s “price-performance advantages” over other options. OpenAI has begun using AWS too, though its workloads run on Nvidia hardware and do not add to Trainium sales.
Despite this momentum, Nvidia’s grip on the AI-chip market remains strong thanks to its scale and CUDA software ecosystem, which makes switching costly. Amazon’s longer-term answer may be Trainium4, designed to work alongside Nvidia GPUs in the same system.
That could help customers blend hardware without leaving CUDA, giving Amazon a path to more market share even if Nvidia stays dominant.