Chips
AI programs built for Nvidia can now move to Google's chips
Google has opened up TorchTPU so the most widely used AI software can run on its chips without being rewritten. Amazon did something similar a day earlier. It remains to be seen whether developers will make the move.

The essentials
- Google released TorchTPU as open-source software on October 20, 2026. It was developed together with Meta's PyTorch team (MarkTechPost, TechCrunch).
- Claimed by one partyPyTorch, the tool used to build many AI models, is used by 85% of AI applications, according to some estimates cited by TechCrunch (TechCrunch).
- With TorchTPU, PyTorch code runs on Google's AI chips by changing the settings, without rewriting the application (Hugging Face Blog, TechCrunch).
- Claimed by one partyThe day before, Amazon introduced TorchNeuron, which does the same for its Trainium chips (TechCrunch).
Why it matters
This suggests that Google and Amazon are going after Nvidia's strong point, which is not just the chip but the software. According to TechCrunch, most AI applications are written with its chips in mind (TechCrunch). If switching providers no longer means a rewrite, companies will find it easier to choose other chips and other clouds.
The details
Until now, using PyTorch on Google's chips meant going through a go-between, PyTorch/XLA, which according to Hugging Face worked but was hard to debug. The other option was to learn JAX, another Google tool (Hugging Face Blog). Google says TorchTPU replaces PyTorch/XLA and recommends it for all new projects. It developed it with Meta's PyTorch team and is releasing it under an open license (MarkTechPost).
According to TechCrunch, Google has worked with Meta, Hugging Face and the creators of vLLM and SGLang, two tools for getting models up and running, so that everything works as it does in the original PyTorch (TechCrunch). It has also released a set of agents that help move applications over to its chips (TechCrunch). There is a catch: in one fine-tuning case, TorchTPU took 11.49 seconds per step without configuration and 2.78 when properly tuned, according to Google's figures (MarkTechPost).
What we don't know
- Whether developers will actually leave Nvidia's chips: TechCrunch warns that they won't do so en masse just because they can.
- Whether the performance figures Google gives will hold up in independent tests.
- Whether using Google's chips is cheaper than using Nvidia's.
Sources
- Hugging Face ❤️ TorchTPU: TPUs, Natively
- Google Open-Sources TorchTPU, a Native PyTorch Backend for Cloud TPUs
- Google has joined Amazon in attacking Nvidia’s stronghold with AI app developers
Was this story useful? Yes Not really