The IT Home page is stamped 2026/10/21 7:14:01. The judgment first: TorchTPU is on GitHub, but “no need to rewrite the code” is Google’s claim. This page has no migration steps and no comparison test.

[1]
A square sealed package seen from above
A square package seen from above, flat charcoal lid, muted blue rim., AI-generated illustration, not a news photograph

The page says Google announced on Tuesday, local time, that TorchTPU, a native PyTorch framework for Google TPUs, is now open source on GitHub. PyTorch is the open library developers use to build, train, and deploy models. The page recalls that Meta first developed it in 2016 and moved it to the Linux Foundation in 2022. Nvidia GPUs were already common research hardware, so many functions were built for them. That is background, not a measurement of this release.

TorchTPU is meant to run applications on Google’s tensor processing units and to move them to Google Cloud. The page’s sentence is that developers need only a simple configuration change and can finish the move without rewriting application code. Which settings change, and whether the old training or inference still runs, are not on this page, and there is no before-and-after comparison. “No rewrite” stays a company claim. It is not a checked migration cost.

[1]

The 85 percent figure cannot be treated as a share. The page says some statistics claim that 85 percent of AI applications use PyTorch. It does not say who counted, what the sample was, or which year. The same passage says most AI applications today run by default on Nvidia chips. Neither sentence has a source. This piece does not treat 85 percent, or “most,” as a checked market share.

The partners are named. The page says Google worked directly with Meta, Hugging Face, Inferact (the maker of vLLM), and RadixArk (the maker of SGLang), so that TorchTPU’s functions and codebase work with those tools the way stock PyTorch does. How far that compatibility goes is not given as a test number.

[1]

Two other items are product descriptions, without evaluations. One is a TPU agent kit. The page says it helps move applications onto Google’s chips and cloud and helps find errors during the move. The other is an upgrade to LiteRT. LiteRT runs models on phones, laptops, and embedded devices. The page says the upgrade supports PyTorch applications better. “Better” has no comparison score.

Amazon is another announcement from the same week, not the subject of this piece. The page says Amazon on Monday, local time, released its own PyTorch variant, TorchNeuron, and said applications can move to its Trainium chips and to Amazon Web Services without a rewrite. This piece notes that comparison and does not expand the method.

The page states its own limit: these developments do not mean developers will abandon Nvidia hardware just because a move to other chips and clouds is easier. Open source and “no rewrite” are both on this page. The engineering time saved, the speed, and the price are not given as numbers here.

[1]