AWS and Google Cloud Open-Source Trainium and TPU-Adapted Versions of PyTorch…

Google Ad Server Settings

Simple Analytics Placeholder event function

×

End of Header

End of superbanner /3002452/Pbk_Rect1_300x300

/3002452/Pbk_Rect2_300x250

AWS and Google Cloud open-source PyTorch versions for AWS Trainium and Google Cloud TPU, taking on the challenge of migrating away from NVIDIA — October 27, 2026

At the "PyTorch Conference 2026" event held in California, USA, on October 20 and 21, AWS and Google Cloud successively announced open-source versions of PyTorch supporting their own AI processors.

AWS released as a public beta "TorchNeuron," which supports AWS Trainium, while Google Cloud announced "TorchTPU," which supports Google Cloud TPU.

TorchNeuron is now open source. Our new @PyTorch backend translates code into operations executable on AWS Trainium with minimal or no code changes. Learn how the team used AI agents and hundreds of thousands of models to validate it: https://t.co/czPvqSPAcR— Amazon Science (@AmazonScience) October 23, 2026

Big day for the @PyTorch ecosystem at #PyTorchConf2026

We’re officially open-sourcing TorchTPU—enabling native, hardware-accelerated PyTorch execution across Cloud TPUs and custom infrastructure.

Here’s everything we launched to supercharge your training & inference workloads https://t.co/1X5A5bhG0z— Google Cloud Tech (@GoogleCloudTech) October 21, 2026

Both TorchNeuron and TorchTPU had already been announced and were provided to some users as private betas or private previews.

Now that they have been released as open source, anyone can use them.

PyTorch running on NVIDIA CUDA is the de facto standard

PyTorch is an open-source machine learning framework developed in Python. It is widely used in research and development of machine learning, AI, and large language models (LLMs).

According to a 2024 survey by the Linux Foundation, PyTorch is used in more than 70% of AI-related research and development, making it an important presence in this field.

PyTorch is a framework that runs on both CPUs and GPUs, but in practice, running it on "CUDA," NVIDIA's general-purpose computing platform software for NVIDIA GPUs, has become the de facto standard.

For this reason, PyTorch and much of its related software are built to deliver maximum performance on CUDA.

Enabling native execution of PyTorch on Trainium and TPU

Against this backdrop, AWS developed AWS Trainium and Google Cloud developed TPU as their own AI processors, and both have been working on supporting the various software that has effectively been built for CUDA.

As for PyTorch, they have improved and maintained compilation infrastructure such as "XLA" for converting it into a format executable on processors such as AWS Trainium and Google Cloud TPU, but they faced challenges in achieving sufficient compatibility and in performance.

TorchNeuron and TorchTPU, released this time, are an extension of these efforts. Both are described as enabling PyTorch to run natively on their respective proprietary AI processors—that is, in a form that fully draws out the processor's performance—and as making porting for that purpose easy, representing major improvements in both compatibility and speed compared with before.

For example, AWS's TorchNeuron is described as follows.

With TorchNeuron, developers can enable PyTorch code on Trainium simply by switching to .to('neuron'), while maintaining full compatibility with eager execution, torch.compile, and the standard PyTorch API. In addition, TorchNeuron is integrated with the Neuron Kernel Interface (NKI), enabling the use of custom operators and performance optimization for Trainium hardware.

Meanwhile, Google Cloud's description of TorchTPU is as follows.

It enables seamless out-of-the-box model execution and TPU migration while delivering industry-competitive inference and training performance through eager execution optimizations, torch.compile integration, and native hardware acceleration.

Through seamless migration to TPU that allows AI models to be executed immediately upon introduction, it achieves high-performance inference and training on par with competitors through aggressive runtime optimizations, torch.compile integration, hardware-native acceleration, and more.

With the arrival of TorchNeuron and TorchTPU, the barriers to using PyTorch on AWS Trainium and Google Cloud TPU are expected to become much lower than before.

In particular, if compatibility with PyTorch's standard APIs and execution model improves and migration becomes possible without extensively rewriting existing code, developers will find it easier to consider options other than NVIDIA GPUs.

However, the choice of AI processor is not determined solely by whether a model runs. It involves various factors, including compatibility with existing libraries and tools, actual training and inference performance, cost, and ease of use from development to operations.

How far will the open-sourcing of TorchNeuron and TorchTPU change the AI development ecosystem that has grown around CUDA? Going forward, attention will focus on the spread of supported models and software, and on how much compatibility and performance can be achieved in actual workloads.

Related articles

AWS Google Cloud Cloud Machine Learning/AI Python

≪Previous article: "DuckDB 2.0 LTS" officially released. Client/server functionality via the Quack protocol and speedups through asynchronous processing, etc.

After working as an editor at an IT magazine and a publisher of an online media outlet, became independent. Started Publickey in 2009.

(Detailed profile)

Would you like to check out the latest news from Publickey?

AWS and Google Cloud open-source PyTorch versions for AWS Trainium and Google Cloud TPU, taking on the challenge of migrating away from NVIDIA

"DuckDB 2.0 LTS" officially released. Client/server functionality via the Quack protocol and speedups through asynchronous processing, etc.

Cloudflare announces "tsrs," a port of TypeScript to Rust. Up to 8x faster with up to 70% less memory than TypeScript 7. Plans to support all TypeScript features including LSP in the future.

"Cloudflare OS" to be offered as a managed service. GitHub connectivity, enhanced Gmail integration, and more.

AWS announces "AWS Lambda Web Functions," which can deploy and run Node.js frameworks and apps such as Next.js and Hono on AWS Lambda as-is.

Microsoft releases preview of "dotnetup," a .NET toolchain manager. Can manage multiple versions of the .NET environment.

Severe Linux kernel vulnerability "PwnEscape" (CVE-2026-99999) may allow execution of malicious code from a guest virtual machine on the host Linux OS.

Anthropic's Bun ships a TypeScript type checker more than 3x faster than TypeScript 7. In Bun's new version 1.4.3.

"Microsoft Execution Containers" (MXC) officially released, safely isolating AI agent execution on Windows. Flexibly customize boundaries for files, networks, and more.

Our efforts to ensure readers can read articles with peace of mind.

Simple Analytics 100% privacy-first analytics

- script for simple analytics events --

Global site tag (gtag.js) - Google Analytics