Article 326: Google and Amazon Are Rewriting PyTorch to Make Nvidia's Biggest Moat Irrelevant

Two announcements in the same week that deserve to be read together, because they are the same move executed twice.

What shipped

Google: TorchTPU. A version of PyTorch - the framework behind more than 85% of AI applications, originally built by Meta in 2016 explicitly around Nvidia's hardware - that runs natively on Tensor Processing Units. The claim that matters is the migration path: existing applications run on TPUs after a few configuration tweaks, without rewriting underlying code. Alongside it, Google shipped a TPU agent suite that migrates apps to Google Cloud and auto-fixes errors during the move, plus an update to LiteRT for running PyTorch-based apps on phones and laptops.

AWS: TorchNeuron. The same idea for Trainium. AWS went further than a fork: it contributed patches upstream to core PyTorch to decouple key features from Nvidia GPUs, and shipped a vLLM plugin.

Both were presented at the PyTorch conference, with Hugging Face co-presenting TorchNeuron.

Why this is the right target

Google and Amazon have had competitive silicon for years. TSMC has been shipping it, the price-per-token maths has often favoured it, and enterprises still buy Nvidia by the rack. The reason is not performance. It is that a team with a working PyTorch stack, tuned kernels, CUDA-specific optimisations and a vLLM deployment has a switching cost measured in engineer-quarters - and no incentive to spend them.

Rewriting the framework is the only move that attacks that cost directly. If the answer to "can we move this to TPUs?" is "change the config," Nvidia's advantage narrows to raw performance and supply, which are negotiable. If the answer remains "rewrite your stack," nothing else matters.

The honest caveat

Constellation Research's Holger Mueller put the counter-argument plainly: it is a win for developer choice, "but it remains to be seen how much they can chip away at Nvidia's strongly entrenched position." A compatibility layer that produces correct-but-slower code changes nothing, because the buyer's problem was never portability in the abstract - it was unit economics. TorchTPU and TorchNeuron will be judged on tokens per dollar after migration, not on how little code changed to get there.

Key Takeaway: Nvidia has spent two years diversifying into networking, systems and reference designs. Meanwhile its two biggest customers-by-necessity have gone after the actual lock-in, which lives in a Linux Foundation repo. Neither Google nor Amazon needs to win the accelerator war outright. They only need portability to become cheap enough that "we're a CUDA shop" stops being a technical argument and becomes a procurement preference.

Sources: SiliconANGLE, "Google unveils TorchTPU in a push to break Nvidia's stranglehold on AI compute," 20 October 2026; Google Cloud blog; AWS What's New; PyTorch Foundation