Google and Amazon Both Want Your PyTorch Code to Run on Their Chips, Not Nvidia's
Google's TorchTPU and Amazon's TorchNeuron promise to move PyTorch apps to their own AI chips with a simple configuration change. No benchmarks yet.
Google and Amazon have each released a tool aimed at one of Nvidia’s quieter advantages: the software developers already write. According to TechCrunch on October 20, Google announced TorchTPU, an open-source version of PyTorch that runs on its Tensor Processing Units (TPUs), and Amazon announced TorchNeuron a day earlier for its Trainium chips. Both say existing PyTorch programs can move over with “a simple configuration change” and no rewriting.
PyTorch is the programming toolkit most AI apps are built with. Meta created it in 2016 and handed it to the Linux Foundation in 2022. TechCrunch says some accounts put its use at 85% of AI apps, though that figure is not firmly sourced. Google says it worked with Meta, Hugging Face, Inferact (maker of vLLM), RadixArk (maker of SGLang) and others so TorchTPU works well with popular tools. It also released a “TPU agent suite” to help developers port apps and fix errors on the way, and updated LiteRT, its tool for running AI on phones and laptops. Amazon says it is submitting code to PyTorch itself that “decouples various features from Nvidia’s chips.”
The caveats are the important part. The coverage reports no adoption figures and no performance comparisons for either tool. It also stresses that this does not mean developers will drop Nvidia hardware automatically. “Works with a configuration change” is a promise; how well it works on real, messy projects is still to be shown.
What’s behind it: a chip is only useful if software can use it easily. Most AI code has been written and tuned with Nvidia’s GPUs in mind, because PyTorch’s features were built around them. That convenience is a big reason people keep buying Nvidia, even when other chips may be cheaper. If moving to a competitor takes a weekend of rewriting, most teams stay put. If it takes one setting, the switching cost drops and prices could come under pressure. Think of it like phone chargers: the connector, more than the battery, decides what you buy.
What this means for you: for most people, nothing changes today. Over time, more competition among chip suppliers could mean cheaper AI services and less dependence on one company. If you train or run models yourself, wait for independent benchmarks and test your own workload on a small scale before moving anything.
Sources
DeepSeek Releases DeepInfer, the Engine It Uses to Serve Its Cheapest Model
DeepSeek open-sourced the serving software behind V4.1-Flash under Apache 2.0. No speed numbers yet, but it could lower the cost of running the model yourself.
