Google’s TPU Inference Stack Is Running in Production at Meta

Google has announced that Meta is running its new SGLang-TorchTPU inference software in production, giving the social-media giant a more direct route to use Google’s Tensor Processing Units with PyTorch workloads.

The disclosure came on 20 October 2026 at PyTorch Conference, alongside Google’s launch of two PyTorch-native TPU inference backends: vLLM-TorchTPU and SGLang-TorchTPU. Google also released TorchTPU, its native PyTorch software stack for TPUs, as open source.

“SGLang-TorchTPU is running in production at Meta today,” said Zongwei Zhou, Meta’s PyTorch-TPU lead.

Meta’s Jana van Greunen, director of PyTorch engineering, said: “We cannot be constrained by a single hardware architecture.” Google says TorchTPU supports standard PyTorch APIs, eager-mode debugging, torch.compile, distributed workloads and tools including vLLM, SGLang and Hugging Face libraries.

What is confirmed — and what is not

The announcement confirms TPU-backed workloads are operating in Meta’s production environment, but Google and Meta have not identified the workloads, TPU model, number of chips or deployment location. It is also not confirmed whether Meta is using Google Cloud capacity or hardware installed in its own facilities.

Google’s current Ironwood TPU, also called TPU7x, is designed primarily for inference and supports PyTorch. Google Cloud lists Ironwood access in its London region from $13.20 per chip-hour on demand, although that public price does not show what Meta is paying or where its deployment is running.

Why this is different

Developers have previously been able to use PyTorch with TPUs through PyTorch/XLA, a bridge to Google’s XLA compiler stack. TorchTPU is intended to make TPUs behave more like a native PyTorch target, with standard tensors, compilation and distributed primitives.

For consumers, the immediate effect is unlikely to be visible. The significance is that one of the world’s largest AI operators is publicly using Google’s TPU software in production, potentially widening the hardware options behind the AI features built into apps and online services.