TorchTPU is now open source, Google announced on Tuesday, and it lets PyTorch apps run on Google’s TPUs after a configuration change instead of a rewrite. Amazon shipped TorchNeuron for its Trainium chips a day earlier, so two of the largest cloud providers spent consecutive days telling developers that leaving Nvidia no longer means leaving PyTorch.

PyTorch sits under roughly 85% of AI apps by some accounts, which is TechCrunch’s phrasing and not a figure I can audit, and it grew up on Nvidia GPUs because that is what researchers had in 2016. Meta handed it to the Linux Foundation in 2022. The wrapper is the visible part, but what I care about more is that Amazon says it is submitting code to mainline PyTorch that pulls features away from Nvidia-specific hardware. A port layer belongs to Amazon. Upstream patches belong to everyone, and they are still there after the marketing cycle is over.

Google’s pitch leans on who it talked to: Meta, Hugging Face, Inferact (the vLLM maker) and RadixArk (SGLang), so the serving stacks startups already run should behave the same on TPUs. It added an agent tool to help with migrations and fixes, and updated LiteRT so PyTorch models run better on phones and embedded hardware. Whether any of it holds up on a messy production codebase is the open question. “One configuration change” is the vendors’ phrase, and I haven’t seen an outside team move a real workload over.

In September I argued that the GPU was never Nvidia’s real moat, that the software and orchestration around it was, and these two releases are the first time I’ve seen two clouds go at that layer on back-to-back days. The hardware bets were already placed: Meta’s MTIA 450 was running DeepSeek’s models last month, Anthropic is reportedly lining up Google TPUs under the $60 billion Project North Star financing (unconfirmed by either company), and its $5 billion AMD bet was the hedge on the other side. What these tools do is make the next developer’s decision cheaper. TechCrunch’s own read is that nobody will drop Nvidia in bulk because of this, and on habit alone I agree. Habits give way when a rented TPU costs clearly less than the GPU alternative, and neither announcement says anything about price.

Sources