# Google Open-Sources TorchTPU to Run PyTorch Code on TPUs

- URL: https://usa-24.com/torchtpu-open-source-pytorch-tpu/
- Publisher: USA 24
- Section: Artificial Intelligence (AI)
- Published: 2026-10-22T18:49:26.582Z
- Updated: 2026-10-22T18:49:26.592Z
- Topics: TorchTPU, PyTorch, Google TPU, Nvidia CUDA, Open Source AI

> Google open-sourced TorchTPU, a native PyTorch backend that lets existing code run on its TPU chips with minimal changes. Here is what works, what is still unverified, and how to test it.

## Key takeaways

- Google open-sourced TorchTPU on October 20, 2026, as a native PyTorch backend for its TPU chips under an Apache 2.0 license.
- The typical migration is a device-call swap, from .to("cuda") to .to("tpu"), according to reporting on the release.
- Google's 52% model FLOPs utilization figure for training Qwen 3 8B on Ironwood is self-reported, and the sources reviewed contain no independent benchmark of it.
- Amazon released a similar PyTorch backend, TorchNeuron, for its Trainium chips on the same day, according to AI Weekly.
- Developers should run a small training or serving job on Cloud TPU before moving any production workload.

Google put TorchTPU on GitHub as open source on Tuesday, October 20, and announced it at the PyTorch Conference 2026. The pitch is direct: take the PyTorch code you already have, change the line that picks the device, and run it on Google's Tensor Processing Units instead of an Nvidia GPU. In a field built around CUDA, that is a bigger claim than it sounds.

## What Google released with TorchTPU and who helped build it

TorchTPU is a native PyTorch stack for TPUs. In [Google's Cloud blog post announcing the release](https://cloud.google.com/blog/topics/ai-infrastructure/torchtpu-is-now-open-source-run-native-pytorch-on-tpus), the company describes it as a way to make TPUs a drop-in device alongside GPUs and CPUs. The current package is v0.2.0 on PyPI, it requires Python 3.11 or newer, and it is licensed under Apache 2.0.

The project was co-developed with the PyTorch team at Meta. That matters because PyTorch originated at Meta and is developed with support from the Linux Foundation, which gives the Google work a direct line into the framework's core engineers. Jana van Greunen, Director of PyTorch Engineering at Meta, said in Google's blog post that TorchTPU "lets us deploy our PyTorch workloads to Google TPUs." Meta has its own large AI infrastructure bets, and USA 24 has covered [Meta's move to lift its 2026 AI capex floor](https://usa-24.com/meta-ai-capex-class-c-stock/).

Work on the project reportedly began in mid-2025, according to CryptoBriefing. MarkTechPost reports that the backend replaces the older PyTorch/XLA path and that Google recommends it for all new TPU workloads. That recommendation is the clearest sign that Google wants TorchTPU to become the default path, not a side project.

## How to run PyTorch on TPU without a rewrite

Most teams will hit one change first. According to CryptoBriefing and Daily.dev's community summary, the typical migration is a device-call swap, from `.to("cuda")` to `.to("tpu")`. If the rest of the code uses familiar PyTorch APIs, the change can stop there.

Google says developers can debug in eager mode and then move to torch.compile for production performance. It also says TorchTPU works with vLLM, SGLang, Ray, TorchTitan and Hugging Face, among other tools. For teams that already built training and serving pipelines around those projects, that list matters more than the headline.

The key facts from the release, as reported across the sources:

- Package: TorchTPU v0.2.0 on PyPI
- Requirements: Python 3.11 or newer
- License: Apache 2.0
- Scale tested: TPU v5e up to 8,192 Ironwood (TPU7x) chips
- Replaces: the older PyTorch/XLA path

## The performance numbers, and which ones are unverified

Google's headline benchmark is up to 52% model FLOPs utilization (MFU) when training Qwen 3 8B with TorchTitan on Ironwood. Google describes that figure as self-reported. MFU measures how much of a chip's theoretical peak compute goes into useful model math. A higher number is better, but it does not translate directly into cost per token or time to train, and the sources reviewed include no independent benchmark of it.

Two community figures from the Daily.dev summary are unverified. One reports that 4,821 of 4,917 tested Hugging Face models passed, about 98%. The other puts TorchTPU at 1,828 tokens per second against 1,937 for JAX/MaxText on a 64-chip TPU v7 pod, a gap of roughly 6%. Treat both as leads to check, not results to plan around.

| Claim | Where it comes from | Status |
|---|---|---|
| Up to 52% MFU training Qwen 3 8B on Ironwood | Google | Self-reported |
| 4,821 of 4,917 Hugging Face models pass (about 98%) | Daily.dev community summary | Unverified |
| 1,828 vs. 1,937 tokens per second vs. JAX/MaxText, 64-chip TPU v7 | Daily.dev community summary | Unverified, roughly 6% gap |
| Tested from TPU v5e up to 8,192 Ironwood chips | MarkTechPost | Not independently verified in sources reviewed |

## What it means for Nvidia and Google TPU buyers

Nvidia's grip on AI training rests heavily on CUDA, the software layer most PyTorch code was written against. TorchTPU attacks that from the framework side. If a PyTorch model can move to TPUs with a one-line change, the switching cost that has kept many teams on GPUs gets smaller. That is the real pressure point. The near-term effect is unproven, though. No source quantified any shift in Nvidia's market share, and open questions about custom CUDA kernels and niche operators limit how far a move can go today.

Google is not acting alone. AI Weekly's daily roundup reports that Amazon released a similar PyTorch backend, TorchNeuron, for its Trainium chips under Apache 2.0 on October 20. Two cloud providers shipping portability layers in the same week suggests the contest is shifting toward software. Whoever makes it easiest to move a model between accelerators gives customers more room to bargain on price.

Google Cloud gains the most if customers test and stay. Lower switching friction is a real selling point for a cloud trying to win workloads from entrenched GPU setups. Google says some Cloud customers have already begun adopting the backend, but it has not named them, so the adoption case rests on the company's word for now.

Cost is the other half of the equation. Buyers are not only comparing chip prices. Power draw and capacity have become a political question, and USA 24 recently reported that [Trump says data centers should make their own power and pay dividends](https://usa-24.com/trump-data-center-dividend-power-bills/). Any TPU versus GPU comparison should include the power a cluster draws, not just the hourly rate.

## Testing TorchTPU before you commit

Developers have a clear first step. Review the TorchTPU repositories on GitHub and read the release notes. Then:

1. Create a clean Python 3.11 or newer environment and install TorchTPU v0.2.0 from PyPI.
2. Change the device call in one small training or inference script, and run it on a Cloud TPU.
3. Debug in eager mode first, then benchmark torch.compile against your current GPU baseline on throughput and cost per run.
4. Search your code for custom CUDA kernels and niche operators, since the sources do not establish how far compatibility extends.
5. Pull current Cloud TPU rates from Google Cloud's pricing pages. The sources reviewed do not include TPU pricing, so this release alone does not support a cost comparison.

Business buyers should ask Google Cloud and other vendors for benchmarks on their own models, not only the published figures. Teams already comparing models on Google Cloud can run the same kind of test alongside other options. USA 24 reported that [OpenAI's GPT-6 models arrived on Google Cloud Model Garden in preview](https://usa-24.com/openai-google-cloud-model-garden/), so the cloud's model catalog is widening at the same time its chip software is.

## What to watch next

Three things will show whether TorchTPU matters beyond launch week. First, whether independent benchmarks reproduce Google's 52% MFU figure and the community speed tests. Second, whether Google names the Cloud customers it says are already adopting the backend. Third, how far compatibility reaches into custom CUDA kernels, which is where many production workloads live.

Amazon's TorchNeuron will face the same questions, and Nvidia's response, whether in software or pricing, will show how seriously it takes the threat. The sharpest signal over the next few months will come from engineering teams that publish migration write-ups with real costs attached, not from launch-day numbers.

## Frequently asked questions

### Is TorchTPU free to use?

TorchTPU is open source under the Apache 2.0 license, so the software itself carries no license fee. Running it on Cloud TPUs still means paying for TPU time, and the sources reviewed do not include current Cloud TPU pricing. Check Google Cloud's pricing pages for current rates before estimating costs.

### Do I need to rewrite my PyTorch code to run on TPUs?

Google describes the typical migration as a device-call swap, changing .to("cuda") to .to("tpu"), with the rest of the code unchanged if it uses familiar PyTorch APIs. Custom CUDA kernels and niche operators may need extra work. The sources reviewed do not establish how far compatibility extends.

### Is TorchTPU faster than Nvidia GPUs?

Google's headline figure is up to 52% model FLOPs utilization when training Qwen 3 8B on Ironwood, which the company self-reports. Community tests in a Daily.dev summary are unverified. The sources reviewed include no independent head-to-head benchmark against Nvidia GPUs.

### What does TorchTPU replace?

TorchTPU replaces the older PyTorch/XLA path for TPUs, according to MarkTechPost, and Google recommends it for all new TPU workloads. The package is v0.2.0 on PyPI and requires Python 3.11 or newer.

## Sources

1. [TorchTPU is now open source: Run native PyTorch on TPUs — Google Cloud Blog](https://cloud.google.com/blog/topics/ai-infrastructure/torchtpu-is-now-open-source-run-native-pytorch-on-tpus)
2. [Google Open-Sources TorchTPU, a Native PyTorch Backend for Cloud TPUs — MarkTechPost](https://www.marktechpost.com/2026/10/20/google-torchtpu-open-source-native-pytorch-tpu/)
3. [Google open-sources TorchTPU to bring native PyTorch support to its TPUs — CryptoBriefing](https://cryptobriefing.com/google-open-sources-torchtpu-pytorch-tpus/)
4. [Google and Meta open-source TorchTPU to run PyTorch on TPUs — AI Early Edition](https://www.aiearlyedition.com/tools/google-meta-torchtpu-open-source/)
5. [TorchTPU is open source: native PyTorch on Google TPUs — Daily.dev (community summary)](https://daily.dev/posts/torchtpu-is-open-source-native-pytorch-on-google-tpus-8c1x9zqzd)
6. [AI News for October 21, 2026 — AI Weekly](https://aiweekly.co/ai-news-today/edition/2026-10-21)