In a major expansion of their partnership, Amazon.com disclosed on Monday its plan to inject up to $25 billion into AI startup Anthropic. This includes an immediate infusion of $5 billion and a potential additional $20 billion contingent upon achieving commercial milestones.
This new investment builds on the already substantial $8 billion previously invested by Amazon in Anthropic, potentially increasing its total commitment to $33 billion. As part of this deal, Anthropic has pledged to spend over $100 billion using Amazon Web Services technologies through 2036.
Anthropic will be granted access to up to 5 gigawatts of computing power for the training and deployment of its Claude AI models on Amazon’s custom silicon. Presently utilizing 1 million AWS Trainium2 chips, Anthropic is set to increase this capacity with additional Trainium2 and Trainium3 resources as Amazon rolls out another gigawatt by late 2026.
“Our tailor-made AI silicon delivers high performance at significantly reduced costs for customers, driving its high demand,” stated Amazon CEO Andy Jassy. “Anthropic’s decade-long commitment to operating its large language models on AWS Trainium underscores our collaborative advancements in custom silicon, as we continue to supply the technology and infrastructure required for generative AI development.”
This latest investment follows Amazon’s $50 billion contribution to OpenAI’s recent $110 billion funding round, which valued OpenAI at a pre-money valuation of $730 billion. The e-commerce giant anticipates allocating $200 billion towards capital expenditures this year, focusing primarily on the growth of AI infrastructure.
Established in 2021 by ex-OpenAI researchers and executives, Anthropic has reported annualized revenues exceeding $30 billion. Its Claude AI models are direct competitors to OpenAI’s GPT series and Google’s Gemini.
The roadmap for Amazon’s custom chips includes Trainium3 processors anticipated for release in December 2025, alongside the forthcoming Trainium4 expected to provide 2 exaflops of performance for FP4 data processing, according to AWS.