Google TPUs now serve AI in production at Meta

Google's TPUs now run production AI workloads in Meta's data centers. New open-source engines let PyTorch models run natively on the chips.

Why it matters: Meta, a major Nvidia customer and PyTorch's steward, running on TPUs is a significant crack in Nvidia's dominance. It fits a wider push, with Google's tooling to move PyTorch off Nvidia, Amazon's TorchNeuron and AMD/Intel PyTorch work, to make PyTorch hardware-portable. Lower switching costs should increase buyers' leverage on chip pricing.