Google Open-Sources TorchTPU and Amazon Releases TorchNeuron, Two Moves Against Nvidia's Software Hold on AI
Two rival clouds are making PyTorch code run on their own chips with minimal edits. Both launched at the PyTorch Conference, and both come with caveats worth reading.
Outspoken Digest Technology Desk
Wednesday, October 21, 2026/2 min read
Nvidia's grip on AI hardware has never rested only on its chips. It rests on the software that developers already know. This week Google and Amazon each tried to loosen that hold by making the most popular framework, PyTorch, run on their own processors without a rewrite.
Amazon went first on Monday with TorchNeuron, and Google followed on Tuesday with TorchTPU, TechCrunch reported, both timed to the PyTorch Conference.
What is TorchTPU?
According to Google's announcement, TorchTPU is a native PyTorch stack for its Tensor Processing Units, built with Meta's PyTorch team and now open source on GitHub. Google says existing workloads run with minimal code changes: debug in eager mode, then move to torch.compile for speed.
It lists integrations with vLLM, SGLang, Ray and Hugging Face libraries, plus a preview suite of coding-agent plug-ins that help move apps onto TPUs. The performance figures are Google's own. It cites up to 52 percent model FLOPs utilisation training Qwen 3 8B, scaling across as many as 8,192 of its Ironwood chips, and a customer, Zenteiq, reporting over four times the inference throughput on 1,000 TPU v6e accelerators.
What is Amazon's TorchNeuron?
TorchNeuron is an open-source PyTorch backend for Amazon's Trainium2 and Trainium3 chips, released under the Apache 2.0 licence, The Next Web reported. Amazon says it plugs into PyTorch's built-in mechanism for new hardware, so standard testing and scaling tools keep working, and that switching needs a single device setting.
The limits are stated openly. The first release skips about 22,000 PyTorch tests, according to the PyTorch blog post cited by The Next Web, and does not yet support some features, including dynamic input shapes. TechCrunch adds that Amazon is also sending code upstream to PyTorch to decouple features from Nvidia chips.
Why does PyTorch matter to Nvidia?
Meta created PyTorch in 2016 and gave it to the Linux Foundation in 2022. Many of its features were first built around Nvidia GPUs, which made Nvidia the default for most AI applications. TechCrunch notes that PyTorch is used by roughly 85 percent of AI apps, by some accounts.
That is the real target. If a team can keep its PyTorch code and still choose a cheaper or more available chip, the switching cost falls. If you want to see how the same fight looks from the hardware side, our report on Microsoft's Maia 300 covers another custom chip aimed at the same prize.
Will developers actually leave Nvidia?
Not overnight. TechCrunch stresses that developers will not abandon Nvidia en masse, and neither release has independent benchmarks yet. The test is whether production teams report that their unmodified code runs fast on the new back ends. Watch for third-party results in the coming months.
Published in The Outspoken Digest
Editorial desk
Outspoken Digest Technology DeskSoftware, hardware, artificial intelligence and what they change for everyone else.
Newsletter
The Digest, in your inbox
One edition, sent when it is ready. No noise, and your address is never passed on.
Read Next
More Technology →
Astronomers Say a Pulsar's Odd Slowdown Is Explained by a Hidden Sunlike Star, Making It a Rare Stellar Trio
Oct 21, 2026/2 min read
Interstellar Comet 3I/ATLAS May Have Lost Metres of Its Surface to the Sun, a New Estimate Finds
Oct 21, 2026/2 min read
A Power Event in Google Cloud's London Region Knocked Out Banking Apps and Delivery Services for Hours
Oct 21, 2026/2 min read
An Indian Certificate Authority's Private Keys Were Trivially Guessable, and Browsers Have Revoked Them
Oct 21, 2026/2 min read