Amazon TorchNeuron operates PyTorch code on Trainium chips

Amazon introduced TorchNeuron at the PyTorch Conference North America 2026. PyTorch is a software kit for AI models. TorchNeuron lets developers operate existing PyTorch code on Trainium, the AI chips of Amazon. Almost no code changes are necessary. Amazon says it checked the 100,000 most popular models. Google released almost the same tool for its own chips.

How it works

Developers change the device name in their code from cuda to neuron. The backend, the layer between the code and the hardware, also works with distributed training, where many chips share work. Companies can move capacity from inference (the use of a model to give answers) to training in off peak hours.

The numbers

Amazon says it checked the 100,000 most popular models and more than 800 model designs. Cartesia uses TorchNeuron for voice AI. The Lieber Institute for Brain Development explores uses of it in drug discovery. Amazon gave no speed or cost numbers.

What is not known

A user named Pratham says the primary test has three parts:

• speed for each workload
• cost for each token (a unit of AI text)
• results that are the same each time

The source has no data on these points.

Insight: Both big clouds make a move between chips easy

Google and Meta released TorchTPU on 21 October 2026. It also needs only a small code change, but the model must have tuning (set up for speed) to be fast. Amazon offers almost the same path. This looks like a race to remove the code cost of a move. The fight then goes to speed and cost for each token. Buyers get more control, and must compare correct results.