Amazon and Google Launch Native PyTorch Backends to Challenge Nvidia's AI Chip Dominance

Amazon and Google have launched competing PyTorch backends within 24 hours, each enabling AI developers to run existing PyTorch applications on their custom chips with minimal code changes, directly challenging Nvidia's longstanding advantage in the GPU-dominated machine learning ecosystem. On October 19, Amazon released TorchNeuron, a native PyTorch backend for its Trainium chips that lets developers switch device targets with a one-line configuration change. The following day, Google and Meta open-sourced TorchTPU, a similar solution for Google's Tensor Processing Units, both unveiled at the PyTorch Conference North America in San Jose.