What happened
Google's TPUs (Tensor Processing Units), its custom chips for AI workloads, are now running in production at Meta's data centers. This is the first time Google's chips have been used in production by an external customer.
Native PyTorch support
The announcement came on Google's developer blog. A new PyTorch-native version of the SGLang inference engine and a matching version of vLLM were released as open source. Previously, both engines relied on Google's own JAX framework to run on TPUs. The new versions let PyTorch-based AI applications run on TPUs mostly without rewriting.
Who was involved
Meta's PyTorch-TPU lead Zongwei Zhou said SGLang-TorchTPU is running in production at Meta today. The engine was built through a close collaboration between Meta, Google and the SGLang team. RadixArk, a startup founded by SGLang's creators, also took part. Meta said it cannot be constrained by a single hardware architecture.
Why it matters
The move aims to loosen NVIDIA's CUDA software lock-in. Amazon also announced at the same PyTorch conference that it was open-sourcing its own PyTorch backend for Trainium chips. By selling TPUs to external customers for the first time, Google is challenging NVIDIA's market dominance.
