Amazon is reportedly in negotiations to sell its custom-designed artificial intelligence chips directly to third-party data centers. This strategic shift marks a significant departure from its previous model of keeping proprietary hardware exclusive to its Amazon Web Services (AWS) ecosystem. By offering its silicon to external operators, the retail and cloud giant is positioning itself as a merchant chip vendor, taking a direct shot at Nvidia’s near-monopoly on the AI hardware market.
Peter DeSantis, AWS’s head of AI infrastructure, confirmed the discussions during an interview in Paris, emphasizing that the rapid evolution of AI infrastructure necessitates new ways to serve a broader customer base. While DeSantis declined to name specific potential buyers, the move signals that Amazon is confident enough in its hardware performance to compete in the open market. This pivot could disrupt the current supply chain, which has been plagued by long lead times and high premiums for Nvidia’s industry-standard H100 and Blackwell chips.
At the center of this strategy is Trainium, Amazon’s dedicated AI accelerator launched in 2020. The hardware has already gained significant traction within the AWS environment, powering workloads for heavyweights such as OpenAI, Anthropic, and Uber. Amazon recently disclosed that Trainium-related services have already generated more than $225 billion in revenue commitments, illustrating the massive scale of demand for alternatives to mainstream silicon.
The potential expansion into external sales drove Amazon’s shares up by 1.8% to a mid-day peak following the news. As data center operators scramble to build out capacity for generative AI, the availability of high-performance, cost-effective alternatives like Trainium could shift the power dynamics of the industry. For Amazon, this represents a transition from being the world's largest cloud landlord to becoming a foundational architect of the global AI hardware layer.