Google and Amazon cut the cost of leaving Nvidia to a one-line change

At PyTorch Conference 2026 Google open-sourced TorchTPU, a native PyTorch backend that targets TPUs via a standard torch.device("tpu") call with no rewrite. A day earlier Amazon released TorchNeuron in public beta, running PyTorch code on AWS Trainium chips unmodified.

• TorchTPU was co-developed with Meta's PyTorch team
• TorchNeuron runs on AWS Trainium without code changes
• Meta's Prometheus cluster uses GB200 NVL72 racks toward ~1M GPUs
• DeepSeek open-sourced DeepInfer and is co-designing chips with Huawei