Google and Amazon Open PyTorch Paths to Their Own AI Chips, Challenging Nvidia's Dominance

Google has unveiled TorchTPU, an open‑source port of PyTorch that runs natively on its Tensor Processing Units. The release follows Amazon’s Monday announcement of TorchNeuron, which brings the same framework to the Trainium accelerator. Both projects allow existing PyTorch workloads to migrate with a simple configuration switch, eliminating the need for extensive rewrites.

TorchTPU integrates tightly with popular libraries such as Hugging Face Transformers, vLLM, and SGLang, thanks to direct collaboration with Meta, Hugging Face, Inferact, and RadixArk. Google also ships a TPU Agent Suite to automate migration, error detection, and performance tuning, while updating LiteRT for smoother on‑device inference.

Amazon is taking a complementary approach by contributing upstream patches that decouple PyTorch internals from Nvidia‑specific kernels, making the framework hardware‑agnostic by default. This dual strategy attacks the software lock‑in that has kept most AI developers on Nvidia GPUs.

Why it matters for GPU/AI infrastructure

By providing first‑class PyTorch support on proprietary silicon, the two cloud giants lower the switching cost for enterprises and startups. The result is a more competitive hardware market, potential price pressure on Nvidia, and greater flexibility for workloads that demand specialized accelerators.