Google released TorchTPU on Tuesday, an open-source port of PyTorch built to run on its own Tensor Processing Units instead of Nvidia's GPUs. It's a long-gestating project, and the timing is pointed: Amazon shipped its own version, called TorchNeuron, just a day earlier, built for its Trainium chips.
PyTorch is the toolkit most AI developers reach for first. By some estimates cited by TechCrunch, it underpins roughly 85% of AI applications built today. Meta, then still Facebook, created it back in 2016, before anyone outside a few labs had heard of large language models, and because Nvidia's GPUs were already the default hardware for AI research, much of PyTorch's plumbing got built around them. Meta handed the project to the Linux Foundation in 2022, but the Nvidia-shaped bones never really went away.
That's the stronghold Google and Amazon are now both trying to crack. TorchTPU lets developers move an existing PyTorch application onto Google's TPUs, and therefore onto Google Cloud, with what Google describes as a simple configuration change rather than a rewrite. Amazon's TorchNeuron works the same way for Trainium and AWS. Amazon is also pushing further upstream: it's submitting code to the main PyTorch project itself that would strip out Nvidia-specific dependencies across the board, making the toolkit hardware-agnostic by default rather than Nvidia-first.
Google's pitch leans on partnerships rather than code alone. It says it worked directly with Meta, Hugging Face, vLLM maker Inferact and SGLang maker RadixArk to make sure TorchTPU's libraries behave the way developers already expect from standard PyTorch, rather than forcing them to learn a parallel set of quirks. Google has also launched a TPU agent suite meant to walk developers through migrating applications and catching errors as they go, and it's updated LiteRT, its tool for running AI on phones, laptops and embedded hardware, to work better with PyTorch-based apps.
None of this forces anyone off Nvidia hardware. Switching costs in AI infrastructure are real, and plenty of startups have built tooling, habits and performance tuning around Nvidia's chips for years. TechCrunch's reporting is blunt about that limit: easier portability doesn't mean developers will actually move.
What it does mean is that the two companies with the deepest pockets outside Nvidia itself are now both making the switch cheaper and less technically painful, on the same week, with overlapping playbooks. Google's bet is service and tooling. Amazon's bet is rewriting PyTorch's own rules from the inside.
Neither chipmaker's silicon is new. What's new is that the software excuse for sticking with Nvidia, that nothing else runs your code without a rewrite, just got measurably weaker. Whether that translates into fewer GPU orders is a question Nvidia's next earnings call will start to answer.
TBFiled by The Tech Business Desk
Big tech, startups, earnings, regulation and policy — the decisions that shape the products.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.