Google said on 20 October 2026 that TorchTPU, “our native PyTorch stack for Google TPUs, is now available as open source on GitHub,” announcing it at PyTorch Conference 2026 in a Google Cloud blog post. “With TorchTPU, TPUs join GPUs and CPUs as a native option for PyTorch,” the post says. TechCrunch reported that with the release, “Google joins Amazon in trying to break Nvidia’s stranglehold on the AI industry, where most AI apps are written by default to run on Nvidia’s chips, or GPUs.”
Amazon released TorchNeuron, its PyTorch backend for its Trainium chips, the day before, according to TechCrunch. Below: what Google released, how it compares with Amazon’s tool, and what the sources do not say.
Business Pill · SWITCHING COSTS
A short explainer on switching costs, why it is hard to leave a supplier. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, Amazon and Google released PyTorch backends for their own chips a day apart, and both describe moving existing PyTorch code with little or no rewriting.
PyTorch is the free, open-source toolkit developers use to build, train and run AI models. Google’s post says TorchTPU was “co-developed with the PyTorch team at Meta” and “lets you run existing PyTorch workloads on TPUs with minimal code changes, using the familiar APIs.” TPUs are Google’s own AI chips.
Google says it also partnered with Hugging Face so that “popular open-weight models across the Hugging Face Hub run natively on TorchTPU,” with Transformers v5 deploying to TPUs through a single device setting. It lists native TPU backends for the vLLM and SGLang serving engines, built with Inferact and RadixArk, plus TPU support in Ray, TorchTitan and VeRL.
The post also announces llm-d 1.0 “with Red Hat, IBM, and a community of more than 500 contributors and 25 organizations,” a TPU agent suite “in preview on GitHub” to help move PyTorch models onto TPUs, and LiteRT, its on-device runtime, joining the PyTorch ecosystem.
Google’s 20 October blog post quotes Zenteiq.ai founder Sashikumaar Ganesan: daily generated tokens rose from 1.47 billion to 5.97 billion per 1,000 TPU v6e accelerators, over 4× higher throughput, using GKE Inference Gateway powered by llm-d. It is a customer’s own figure and describes routing, not TorchTPU itself. Circles are sized by area.
The Numbers Google Gives
Google says TorchTitan recipes have been “validated running with up to 8000 Ironwood TPUs (TPU v7x)” and that TorchTPU’s training paths “demonstrated up to 52% Model Flops Utilization (MFU) on Qwen 3 8B.” Model Flops Utilization measures how much of a chip’s theoretical compute a training run actually uses.
The post quotes Zenteiq.ai founder Sashikumaar Ganesan saying his company “achieved over 4× higher aggregate inference throughput, increasing daily generated token capacity from 1.47 billion to 5.97 billion tokens per 1,000 TPU v6e accelerators,” using GKE Inference Gateway powered by llm-d. That is a customer’s own figure, and it describes request routing rather than TorchTPU itself.
Meta’s director of PyTorch engineering, Jana van Greunen, is quoted in the post: “We cannot be constrained by a single hardware architecture.” She says TorchTPU “enables us to deploy our PyTorch workloads to Google TPUs.”
How It Lines Up With Amazon’s TorchNeuron
Amazon’s tool “lets AI apps written with PyTorch run on Trainium by changing just one line of configuration, according to Amazon,” TechCrunch reported on 19 October. AWS’s What’s New post says it “launched the public beta of TorchNeuron, an open source native PyTorch backend for AWS Trainium,” available for Trainium2 and Trainium3 instances.
TechCrunch wrote that Google “may be one-upping Amazon in a few ways,” citing its work with Meta, Hugging Face, Inferact and RadixArk and its TPU agent suite. Amazon, for its part, said it is “contributing changes upstream to PyTorch itself, working with the PyTorch community to remove code that ties the framework to specific chips,” according to the Amazon post TechCrunch quoted.
The same day, Intel’s newsroom carried an item headlined “AMD and Intel Collaborate to Simplify AI Development on x86 CPUs Using PyTorch,” dated 20 October 2026. We are reporting that headline only.
Connecting the Coverage
On 19 October we reported on Amazon’s free PyTorch tool for Trainium, when Amazon said it would open-source TorchNeuron at PyTorch Conference and Reuters framed it as a push against Nvidia’s advantage. That piece has since carried two Updates: the AMD and Intel headline and the TorchNeuron public beta.
Google’s chips have also featured in our financing coverage. On 17 October we reported that Anthropic offered Blackstone 10% on $9 billion of a $60 billion chip financing for about 1.5 gigawatts of customised Google chips, according to an Italian newspaper’s account of a Financial Times report. In July we covered Google’s Ironwood TPU playbook for running Alibaba’s Qwen 3.5 on its own silicon.
The Structural Read
As we read it, the target is the cost of switching rather than the chip price: Google says TorchTPU runs existing PyTorch workloads “with minimal code changes,” and TechCrunch reported Amazon’s tool needs “just one line of configuration.”
As we read it, Google is leaning on partners to make the switch credible: its post lists Meta, Hugging Face, Inferact and RadixArk, and quotes Meta’s PyTorch engineering director saying “We cannot be constrained by a single hardware architecture.”
As we read it, the open part is the software and the rented part is the silicon: Google’s TorchTPU and Amazon’s TorchNeuron are both open source, while TPUs and Trainium are sold as cloud capacity.
Google Cloud blog, 20 October 2026
“With TorchTPU, TPUs join GPUs and CPUs as a native option for PyTorch”
Three Implications
UP TO 8000 IRONWOOD TPUSGoogle says TorchTitan recipes on TorchTPU have been validated running on up to 8000 Ironwood TPUs (TPU v7x).
52% MFUGoogle says TorchTPU training paths demonstrated up to 52% Model Flops Utilization on Qwen 3 8B.
TRAINIUM2 AND TRAINIUM3AWS says TorchNeuron, in public beta, is open source for Amazon EC2 Trainium2 and Trainium3 instances.
The framework puts it this way: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize).”
As we read it, Google and Amazon are each opening the software layer that moves PyTorch code while keeping their chips, TPUs and Trainium, as the layer they rent out. The framework was published in March 2026 and does not discuss Google, Amazon, Nvidia, TorchTPU or TorchNeuron.
What Is Not Established
Neither Google nor Amazon gives a figure for how many developers have moved PyTorch code to their chips, and neither post gives a price comparison with Nvidia GPUs. TechCrunch wrote that “none of this means AI app developers will automatically ditch Nvidia’s hardware en masse.” The 52% MFU and 4× figures are Google’s and a customer’s own, and we did not test them. Nvidia did not comment in the sources we read.
Business Engineer Framework
The Open vs Closed Meta-Framework
A Business Engineer framework on which layer of a technology stack to open and which to keep closed.
Google says TorchTPU is now open source and makes TPUs a native option for PyTorch, a day after Amazon released TorchNeuron for Trainium, according to TechCrunch. As we read it, both companies are aiming at the same thing: the cost of moving existing PyTorch code off Nvidia chips, which TechCrunch says most AI apps are written to run on by default.
A note on sourcing. This piece is built on Google Cloud’s 20 October blog post, read in full, AWS’s What’s New post, the headline of an Intel newsroom item, two TechCrunch articles and our own earlier pieces. We did not test TorchTPU or TorchNeuron and did not contact Google, Amazon or Nvidia. The chart’s circles are sized from figures a Google customer gives in Google’s post. Nothing here is a product endorsement.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
