Google teams up with Meta to launch a surprise attack on Nvidia, open-sources TorchTPU to migrate PyTorch code

What happened: Google open-sourced TorchTPU, a native TPU software stack co-developed with Meta's PyTorch team, letting developers deploy existing PyTorch models to Google TPUs with few code changes. Key numbers: Training validated on up to 8000 Ironwood TPUs; up to 52% model FLOPs utilization in Qwen3-8B tests; Zenteiq.ai saw inference Why it matters: TechCrunch reported most AI applications run on Nvidia GPUs, and the open-sourcing may aim to break Nvidia's dominance. The article notes Nvidia's software ecosystem remains mature, so TorchTPU is unlikely to change the landscape short term.

Google open-sourced TorchTPU, a native TPU software stack co-developed with Meta's PyTorch team, letting developers deploy existing PyTorch models to Google TPUs with few code changes.

TechCrunch reported most AI applications run on Nvidia GPUs, and the open-sourcing may aim to break Nvidia's dominance. The article notes Nvidia's software ecosystem remains mature, so TorchTPU is unlikely to change the landscape short term.