Google Open-Sources TorchTPU to Pull PyTorch Developers Away from Nvidia

Quick Facts

• Google open-sourced TorchTPU on October 20, 2026, giving PyTorch developers a native path to run AI workloads on Google's TPU chips without rewriting code.
• The project was co-developed with Meta's PyTorch team and replaces PyTorch/XLA, which will be archived on November 15, 2026.
• TorchTPU has been validated on up to 8,000 Ironwood TPUs and reached 52% Model Flops Utilization training Qwen 3 8B at scale.

Google announced TorchTPU as an open-source project on October 20, 2026, at PyTorch Conference North America in San Jose, according to the original report. The move is a direct challenge to Nvidia's grip on AI compute, where most AI applications default to running on Nvidia GPUs and the CUDA software stack that powers them.

Google engineering vice president Bill Jia presented the announcement. The company joins Amazon in targeting that dependency, pushing developers toward alternative silicon.

What TorchTPU Does

TorchTPU is a backend that translates PyTorch instructions into commands a TPU can execute. Developers can switch from GPU to TPU by changing a single line of code, replacing .to('cuda') with .to('tpu'), with no application rewrite required.

That simplicity sets it apart from earlier approaches. Previous tools like PyTorch/XLA forced developers to adapt their code to JAX's programming model. TorchTPU was built as a native PyTorch backend from the start, with eager execution as a core design priority, as detailed in the Google Developers Blog.

The torch-tpu 0.2.0 package shipped on PyPI on October 19, 2026, with Linux x86-64 wheels supporting CPython 3.11 through 3.14. The release spans four GitHub repositories under the Apache 2.0 license, which permits commercial use.

Why It Matters for AI Developers

PyTorch is the dominant toolkit for building, training, and running AI models. Some estimates put its use in roughly 85% of AI applications. Meta originally built it in 2016 with Nvidia GPUs as the default hardware. Meta transferred the project to the Linux Foundation in 2022.

Because PyTorch was designed around Nvidia hardware, most AI workloads are tightly coupled to Nvidia's ecosystem. Porting them to alternative chips has historically required significant engineering effort. TorchTPU removes much of that friction.

Google says thousands of Hugging Face models can run on TPUs with minimal changes. In Hugging Face Transformers v5, passing device_map='tpu' deploys models directly to TPU hardware.

Performance and Scale

Google validated TorchTPU on up to 8,192 Ironwood TPUs, its seventh-generation TPU chip. Training Llama 3 405B on 1,024 TPU v7x chips with an 8,000-token context window in BF16 achieved 500 tokens per second per chip and more than 50% Model Flops Utilization. The system shows near-linear scaling across that chip count.

The backend supports approximately 450 ATen operations, the low-level operations that PyTorch uses internally to execute tensor math.

Ecosystem Integrations

Google worked with Meta, Hugging Face, Inferact (the creator of vLLM), and RadixArk (the creator of SGLang) to build compatibility across the AI tooling stack. Frameworks including vLLM, SGLang, TorchTitan, VeRL, and Ray have already been integrated, as described in the joint announcement with Meta.

Ray now supports native TorchTPU orchestration for training, inference, and reinforcement learning across CPUs, GPUs, and TPUs in a single workflow. Google also updated LiteRT, its toolkit for running models on phones and embedded devices, to work alongside TorchTPU-trained workloads.

PyTorch/XLA Is Gone

TorchTPU officially replaces PyTorch/XLA, the previous method for running PyTorch on Google TPUs. The PyTorch/XLA repository will be archived on November 15, 2026. Google's FAQ states plainly: PyTorch/XLA is deprecated and replaced by TorchTPU.

Google is positioning TorchTPU as production-ready. Some Cloud customers have already adopted the backend. The company is signaling that TPUs are now a first-class option for PyTorch workloads, sitting alongside GPUs and CPUs as supported hardware in the framework's core interface.