Google has released TorchTPU, an open-source PyTorch project that lets developers run their PyTorch applications on the company's Tensor Processing Units without rewriting their code, according to reporting from TechCrunch.
For years, moving an AI workload off Nvidia graphics processors meant wrestling with CUDA-specific code, custom kernels and long compatibility checklists. TorchTPU is Google's attempt to remove that friction, positioning TPUs as a first-class PyTorch target sitting alongside GPUs and CPUs — notable because PyTorch workloads have historically run overwhelmingly on Nvidia hardware.
The timing is deliberate. The release arrived one day after Amazon shipped a similar tool of its own, and both cloud giants have built billion-dollar custom silicon programs — Google's TPUs and Amazon's Trainium and Inferentia chips — whose success depends on convincing developers it is safe to leave Nvidia behind.
Whether developers actually switch will depend on performance, price and how completely the open-source tooling covers real-world models, but the direction of travel is clear: the two largest challengers in AI infrastructure are now competing to make Nvidia's software moat optional.
