Trainium
Amazon's own AI chip, built to run training and inference on AWS without paying Nvidia's margin.
Trainium is Amazon’s custom accelerator for AI workloads, designed in-house and available only through AWS. The logic is the same as Google’s with its TPUs: if you are buying enormous quantities of a component from one supplier at that supplier’s price, at some point it becomes cheaper to design your own, even accounting for the years and the engineers it takes.
For customers the trade-off is cost against convenience. Trainium can be meaningfully cheaper per unit of work, but the software ecosystem around Nvidia is deeper and most code assumes it. Anthropic is the best-known large user, running Claude across a mix of Nvidia, Google and Amazon silicon.