Google has released TorchTPU as an open-source project, allowing developers to run existing PyTorch AI applications directly on its Tensor Processing Units (TPUs) with minimal code changes. Announced on October 20, 2026, the software makes it easier to move AI workloads from Nvidia GPUs to Google Cloud without rewriting applications for different hardware.
The official announcement follows Amazon’s introduction of TorchNeuron, which provides similar compatibility for its Trainium processors. Both companies are developing tools that reduce developers’ dependence on Nvidia hardware when building and operating AI applications.
TorchTPU Brings Native PyTorch Support to Google TPUs
PyTorch is an open-source machine learning framework used to develop, train, and deploy AI models. Originally developed by Meta in 2016 and transferred to the Linux Foundation in 2022, it is used by approximately 85% of AI researchers and developers, according to Amazon.
Historically, many PyTorch features were optimized for Nvidia GPUs, making alternative processors harder to adopt. TorchTPU allows developers to use familiar PyTorch interfaces on Google’s chips, with support for debugging, model training, and inference.
Google developed TorchTPU with Meta’s PyTorch engineering team and collaborated with Hugging Face, Inferact, and RadixArk to support existing AI development tools. These integrations include Hugging Face Transformers, vLLM, SGLang, TorchTitan, Ray, and other commonly used frameworks.
The company also introduced the TPU Agent Suite, which uses AI agents to help developers migrate applications, identify compatibility issues, and resolve errors. Google updated its LiteRT framework to improve compatibility with PyTorch applications running on smartphones, laptops, and embedded devices.
Amazon Introduces TorchNeuron for Trainium Chips
Amazon announced TorchNeuron on October 19, providing native PyTorch support for its custom Trainium AI processors. The open-source tool allows developers to move existing workloads to Amazon Web Services with minimal changes to their code.
Amazon is also contributing improvements to the main PyTorch project to reduce dependencies on Nvidia-specific features. These changes are intended to support different AI processors through standard PyTorch interfaces.
Both companies are targeting developers who want greater flexibility when selecting AI computing infrastructure. Nvidia currently benefits from extensive software support and established developer workflows, which can make switching hardware more complicated.
Google’s TorchTPU is now available on GitHub, with integrations supporting AI model training and inference on TPUs. Amazon’s TorchNeuron is available in public beta for Trainium2 and Trainium3 instances, with both companies continuing to develop compatibility with the wider PyTorch ecosystem.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.