What we know

Google’s Tensor Processing Units (TPUs) are now running AI workloads in Meta’s data centers. This marks the first time Google has sold TPUs to an external customer outside its own operations and Google Cloud. The deployment is made possible by new open-source engines that enable PyTorch models to run natively on TPUs, aligning with Meta’s preferred AI framework. Google’s TPUs are actively used in production AI at Meta. This is the first external sale of TPUs by Google. Open-source software now supports native PyTorch model execution on TPUs.

Why it matters

Google’s TPUs have traditionally been reserved for internal use or Google Cloud customers. Meta’s use of TPUs signals a notable shift in how Google’s AI hardware is distributed and integrated beyond its ecosystem. The open-source support for PyTorch models bridges a key compatibility gap, allowing Meta to apply TPU acceleration within its existing AI workflows.

What is still unknown

Details about the range of supported AI frameworks beyond PyTorch, the accuracy and performance of TPUs in Meta’s environment, and plans for wider availability remain unclear. No independent tests of TPU accuracy or energy efficiency have been reported.