Google Launches TorchTPU to Run PyTorch Models on Cloud TPUs
Google and Meta release an open-source backend that lets developers run PyTorch AI models on Google Cloud TPUs without rewriting code.
Key points
- Google launched TorchTPU on Tuesday to enable PyTorch users to run AI models on Tensor Processing Units.
- The new backend replaces the older PyTorch/XLA system and allows execution without rewriting underlying code.
- Amazon recently announced TorchNeuron for its Trainum chips, contributing code to decouple PyTorch from Nvidia GPUs.
Google LLC launched TorchTPU on Tuesday, an open-source backend designed to allow developers to run PyTorch artificial intelligence models on Google Cloud Tensor Processing Units. According to SiliconANGLE, the release aims to reduce the market dominance of Nvidia Corp. by making it easier to deploy AI workloads on Google’s specialized hardware.
The new tool enables users to execute existing applications and models on TPUs by making simple configuration tweaks, eliminating the need to rewrite underlying code. As Threat Times reported on October 20, 2026, Google and Meta released this native PyTorch backend to replace the older PyTorch/XLA system.
Context and Background
Nvidia maintains a strong position in the AI industry because most applications are written specifically for its graphics processing units. SiliconANGLE notes that more than 85 percent of AI applications are created using PyTorch, a toolkit originally developed by Meta Platforms Inc. in 2016.
Meta designed PyTorch to work with Nvidia hardware when those chips were the preferred choice for AI research. Even after Meta open-sourced the framework in 2022 and transferred stewardship to the Linux Foundation, the focus on Nvidia hardware continued.
Who Is Affected
Google is not the only company attempting to reduce Nvidia’s advantage through PyTorch adaptations. Amazon Web Services Inc. announced its own variant, TorchNeuron, earlier in the same week. This tool allows AI applications to run on Amazon’s Trainum chips.
Amazon went further by contributing code to the original PyTorch project to decouple key features from Nvidia GPUs, aiming for broader hardware compatibility. Google did not contribute to the main project in this manner but offers a direct path for TPU usage.
What should you do?
Developers using PyTorch should evaluate whether their current models can benefit from running on Google Cloud TPUs using the new TorchTPU backend. Organizations relying on Nvidia GPUs should monitor how these alternative backends evolve, as they may offer cost or performance advantages for specific workloads.
What to do and how to stay safe: Google
- Review current AI model dependencies to determine if they are tightly coupled to Nvidia-specific libraries or hardware features.
- Test small-scale PyTorch models using the new TorchTPU backend to assess performance and compatibility with Google Cloud TPUs.
- Monitor updates from Amazon regarding TorchNeuron and its contributions to the main PyTorch project for broader hardware decoupling.
- Consult with cloud infrastructure teams to understand the cost implications of migrating workloads from GPUs to TPUs or other accelerators.
Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is TorchTPU?
TorchTPU is an open-source backend released by Google and Meta that allows PyTorch models to run natively on Google Cloud TPUs.
Does TorchTPU require rewriting code?
No, TorchTPU allows developers to run existing apps and models on TPUs by making simple configuration tweaks without rewriting code.
How does TorchTPU differ from Amazon's TorchNeuron?
TorchTPU runs on Google’s TPUs, while TorchNeuron runs on Amazon’s Trainum chips; Amazon also contributed code to decouple PyTorch from Nvidia GPUs.
