Google has open-sourced TorchTPU, allowing AI applications built with PyTorch to run on TPU chips and Google Cloud by changing configuration, without rewriting the applications.
Amazon has also launched TorchNeuron to move applications to Trainium chips and AWS without rewriting them. Both companies are reducing Nvidia’s advantage in AI systems, but there is still no evidence that developers will abandon Nvidia hardware on a large scale.
From PyTorch to Google’s Chips
TorchTPU acts as a bridge from existing PyTorch applications to TPU and Google Cloud. Developers can simply change the configuration without rewriting the application.
Google also has a TPU agent suite to help migrate applications and fix errors along the way, making it easier to start experimenting with Google’s chips without rebuilding the entire existing system.
The Problem Developers Don’t Want to Solve from Scratch
Teams that already have PyTorch applications are often reluctant to move to a new chip or cloud because they fear having to rewrite code and modify nearly the entire existing system. This is especially true for applications originally built to run on Nvidia chips.
TorchTPU addresses this by allowing existing applications to run on TPU and Google Cloud through configuration changes, without requiring a rewrite. Its selling point is therefore not just providing an alternative chip, but also reducing the friction of migrating systems.
Where TorchTPU Fits into Google’s AI Strategy
TorchTPU is a bridge from PyTorch to TPU and Google Cloud, allowing developers to migrate existing applications by changing the configuration without rewriting them.
Google also has a TPU agent suite to help migrate applications and fix errors along the way. LiteRT has also been improved to work better with AI applications from PyTorch, supporting AI deployment on phones, laptops, and embedded devices.
Overall, this is not simply the launch of an alternative chip. It is an effort to make Google’s tools and cloud work more closely with developers’ existing workflows.
What Changes from Traditional PyTorch to TorchTPU
PyTorch is an open-source tool for building, training, and running AI models. Many of its features traditionally work primarily with Nvidia chips, while TorchTPU provides a path for existing applications to run on TPU and migrate more easily to Google Cloud.
| Factor | PyTorch | TorchTPU |
|---|---|---|
| Main purpose | Build, train, and run AI models | Enable PyTorch applications to run on TPU |
| Hardware | Primarily supports Nvidia | Supports Google TPU |
| Migration to Google Cloud | Requires adjustments for the environment | Change the configuration without rewriting the application |
Where Application Migration Happens in Practice
If an existing PyTorch application is already available, moving it to run on Google TPU can be done by changing the configuration without rewriting the application. This is suitable for teams that want to try Google Cloud without rebuilding their entire codebase.
During migration, the TPU agent suite helps move the application to Google’s chips and cloud while also helping fix errors. Integration with Hugging Face, vLLM, and SGLang allows developers’ existing tools to continue working.
If the application is intended for use on phones, laptops, or embedded devices, LiteRT has also been improved to work better with PyTorch applications.
How Google and Amazon Are Challenging Nvidia Differently
Google uses TorchTPU to move PyTorch applications to TPU and Google Cloud by changing the configuration without rewriting them. Amazon uses TorchNeuron to move applications to Trainium chips and AWS in a similar way.
| Factor | Google’s TorchTPU | Amazon’s TorchNeuron | Traditional Nvidia approach |
|---|---|---|---|
| Target chip | TPU | Trainium | Nvidia GPU |
| Related cloud | Google Cloud | AWS | Not specified |
| Approach to reducing reliance on Nvidia | Move applications through a configuration change without rewriting them | Move applications without rewriting them, while contributing code to PyTorch to support broader hardware | PyTorch was built primarily to work with Nvidia |
The difference is that Amazon is trying to make core PyTorch more open to a wider range of hardware, while Google is focusing on application-migration tools and compatibility with the existing ecosystem.
TorchTPU’s key advantage is that it helps move PyTorch applications to TPU and Google Cloud through a configuration change without requiring a rewrite. It also works with Meta, Hugging Face, vLLM, and SGLang, and includes a TPU agent suite to help migrate applications and fix errors.
The limitation is that there is still no evidence that developers will abandon Nvidia hardware in large numbers. This launch therefore cannot yet establish how successful the platform shift will be.
Pros
- +Move PyTorch applications to TPU and Google Cloud without rewriting them
- +Works with popular AI tools and includes tools to help fix errors
Cons
- −There is still no data showing that large numbers of developers will move away from Nvidia
- −The success of the platform shift cannot be determined from a single announcement
Costs and Burdens This News Has Not Yet Addressed
The announcement does not specify Google Cloud or AWS service costs, including the cost of using TPU and Trainium. It is therefore not yet possible to conclude which option offers better value than Nvidia.
Even if applications can be migrated without rewriting them, there are still no post-migration performance tests or data on the actual operational burden, such as troubleshooting and tuning workloads for each type of chip. These factors could become additional costs, but the source provides no figures for assessing them.
How Much Could the AI Chip Market Change the Game?
TorchTPU and TorchNeuron make it easier to experiment with moving PyTorch applications to alternative chips and cloud platforms because the applications do not need to be rewritten entirely. Developers also have tools to help migrate systems and fix errors.
However, whether the market will truly change depends on developer adoption and real-world usage. If developers do not move away from Nvidia on a broad scale, Nvidia’s advantage will remain. Google and Amazon have nevertheless already lowered the barriers to trying alternative hardware.