Amazon and Google ship PyTorch ports aimed at Nvidia lock-in

Amazon and Google released PyTorch tools so apps can run on Trainium and TPUs with a configuration change, not a rewrite.

Amazon Web Services released TorchNeuron, a PyTorch-based tool it says runs existing PyTorch apps on its Trainium chips after one configuration change and no code rewrite. Amazon says it will contribute upstream PyTorch changes meant to remove code that ties the framework to specific chips, including Nvidia’s default position via CUDA, and cites Deepgram processing 70% more audio per dollar after moving with native PyTorch. A day later, Google said TorchTPU, its PyTorch version for Tensor Processing Units, is open source and that a configuration change can move existing PyTorch apps to TPUs and Google Cloud. Google says it worked with Meta, Hugging Face, Inferact, and RadixArk, and also launched a TPU agent suite and updated LiteRT for PyTorch apps. Amazon says about 85% of AI apps are written with PyTorch; TechCrunch’s later report attributes that figure to “some accounts.”