Google TorchTPU and Amazon TorchNeuron Challenge Nvidia's AI Grip
Google opens TorchTPU as Amazon launches TorchNeuron, letting PyTorch apps shift from Nvidia GPUs to TPUs and Trainium without rewrites across rival clouds.
Summary
On October 20, 2026, Google released the long-awaited TorchTPU as an open-source project, allowing existing PyTorch applications to run on Google Tensor Processing Units and Google Cloud through a configuration change, without rewriting code. PyTorch, which some estimates place in 85% of AI applications, helps developers and startups build, train and run models. Meta, then Facebook, created it in 2016 around Nvidia GPUs, already the AI research standard, before transferring it to the Linux Foundation in 2022.
Amazon released open-source TorchNeuron on October 19, enabling the same no-rewrite migration to its Trainium chips and AWS. Amazon is also submitting code to the main PyTorch project that separates features from Nvidia hardware and makes other chips easier to support. Google worked with Meta, Hugging Face, Inferact, maker of vLLM, RadixArk, maker of SGLang, and others on TorchTPU compatibility. Its new TPU agent suite assists migrations and error correction, while an updated LiteRT improves PyTorch support on phones, laptops and embedded devices. Easier portability weakens Nvidia's software advantage, but does not ensure developers will abandon its GPUs at scale.
Positives
- TorchTPU moves existing PyTorch applications to Google TPUs and Google Cloud through one configuration change, without code rewrites.
- TorchNeuron similarly ports PyTorch applications to Amazon Trainium and AWS without rewriting them.
- Amazon is contributing code that separates PyTorch features from Nvidia hardware and broadens default support for other chips.
- Google coordinated TorchTPU compatibility with Meta, Hugging Face, Inferact, RadixArk and other AI development participants.
- Google's TPU agent suite handles migration and errors, while LiteRT expands PyTorch support across phones, laptops and embedded devices.
Risks & concerns
- Most AI applications remain written by default for Nvidia GPUs, preserving the company's entrenched hardware advantage.
- PyTorch features have targeted Nvidia hardware since Meta created the toolkit in 2016, reinforcing longstanding technical inertia.
- Simpler portability does not guarantee developers will leave Nvidia GPUs in large numbers, leaving future adoption uncertain.