Amazon's AI chips can now run Nvidia's go-to software

For years, buying Nvidia's AI chips meant writing software Nvidia's way. On Monday, Amazon released a free tool called TorchNeuron that breaks that link. Developers keep their existing PyTorch code, the toolkit most AI builders already use. They run it on Amazon's own Trainium chips instead. Only one line of code has to change, and nothing else needs rewriting.

Nvidia's real advantage was never just hardware. It's CUDA, the software layer almost every AI team already builds on. That's what makes switching chips expensive and risky. TorchNeuron attacks that lock-in directly. It lets AI startups test Amazon's cheaper Trainium chips without rewriting their code. But Amazon hasn't published independent numbers yet. Those numbers would have to show Trainium truly matches Nvidia on real workloads, not just in theory.

Keep reading — what to watch

Will big AI labs actually switch? Reworking a production AI pipeline carries real risk. Even a small test run can go wrong if performance comes up short. The real signal to watch is which companies publicly move workloads onto Trainium using TorchNeuron. And whether they report real cost or speed gains, not just that it technically works.