Skip to content
NewsProduct launch

Google open-sources TorchTPU a day after AWS shipped its own PyTorch-to-chip bridge

· by Pondero Newsdesk

The short version

Google released TorchTPU as open source on October 20, 2026, letting PyTorch code run on its TPUs with a config change, one day after Amazon's TorchNeuron did the same for Trainium chips.

Google open-sources TorchTPU a day after AWS shipped its own PyTorch-to-chip bridge

Google released TorchTPU as an open-source project on October 20, 2026, letting developers run existing PyTorch code on its Tensor Processing Units through a configuration change rather than a rewrite, according to Google Cloud's announcement. The release landed one day after Amazon shipped the same kind of bridge, TorchNeuron, for its own Trainium chips, and both tools attack the same target: Nvidia's hold on PyTorch, the framework most AI developers write in by default.

What Google shipped

TorchTPU is now on GitHub, co-developed with the PyTorch team at Meta, and lets existing PyTorch workloads run on TPUs "with minimal code changes, using the familiar APIs," per Google's post. The announcement came at PyTorch Conference 2026, where Google said it worked directly with Meta, Hugging Face, and the teams behind the vLLM and SGLang inference engines, referred to by TechCrunch as Inferact and RadixArk, to keep TorchTPU's libraries compatible with tooling developers already use. Google also updated LiteRT, its tool for running AI on phones and embedded devices, to work with PyTorch apps, and shipped a separate TPU agent suite to help developers port workloads and debug errors along the way, TechCrunch reported. In Transformers v5, Hugging Face says deploying a model to a TPU is now as simple as passing device_map="tpu", per Google's post.

Jana van Greunen, Meta's director of PyTorch engineering, called the move "a massive step forward" that lets Meta "deploy our PyTorch workloads to Google TPUs" and "build the ecosystem flexibility everyone needs," per Google's blog. Hugging Face VP of product Jeff Boudier said the collaboration "puts that commitment into practice," referring to keeping PyTorch open to multiple hardware vendors.

Two cloud rivals are now running the same playbook within 24 hours

AWS's TorchNeuron, which Pondero covered in detail on October 20, uses the identical pitch: a config-line swap instead of a rewrite, aimed at loosening Nvidia's CUDA-driven default status in PyTorch. Google's version goes a step further on ecosystem breadth, per TechCrunch, by lining up named partnerships with Meta and Hugging Face and a dedicated migration tool, where AWS leaned more on upstream code contributions to strip Nvidia-specific assumptions out of PyTorch itself. For a team already running PyTorch on Nvidia GPUs, the practical shift is that two separate cloud providers now offer a low-friction way to test a workload on non-Nvidia silicon without committing to a full migration, which only matters if the switch holds up once a production workload, not a demo, runs on it.

TechCrunch's Julie Bort noted that neither release means developers will abandon Nvidia hardware en masse just because switching got easier, but that Amazon and Google are "clearly working to end Nvidia's advantage." Nvidia had not issued a public response to TorchTPU as of this writing, mirroring its silence on TorchNeuron the day before.

What to watch next

Watch for independent benchmarks comparing TPU price-performance against Nvidia's GPUs under TorchTPU, separate from Google's own figures, and whether vLLM and SGLang maintainers confirm the compatibility Google described at their own pace. Also worth tracking: whether AWS or Google responds to the other's release with a follow-up feature, given both shipped within a single week of each other.

Sources