Particle.news
Download on the App Store

Amazon in Talks to Sell Trainium Chips as Customers Shift from Nvidia

The discussions would let other data centers buy AWS-designed accelerators and could change who controls cost and capacity for large AI workloads.

Overview

  • Reports published Thursday say Amazon is negotiating to sell its Trainium processors to outside companies so those firms can run the chips in their own data centers.
  • AWS’s custom silicon business has grown rapidly and exceeded a $20 billion annualized run rate according to company disclosures earlier this year.
  • Customers using Inferentia for inference report steep cost cuts and some migrations claim savings as large as 80 to 90 percent versus Nvidia GPUs.
  • Amazon says Trainium3, released late in 2025, boosts performance over prior generations and can cut training and inference costs versus conventional GPUs, and major users including Anthropic and Uber already run large Trainium deployments.
  • Nvidia remains dominant because of its high-end GPU performance and the widespread CUDA software ecosystem, and analysts say manufacturing limits and planned 2027 capacity ramps will decide if AWS chips win lasting share.