Google TPUs Run Meta Production AI Workloads Via PyTorch-Native SGLang Engine
Google’s custom Tensor Processing Units are running artificial intelligence workloads in production at Meta, utilizing a newly developed, PyTorch-native version of the SGLang inference engine. Tom’s Hardware reported that the deployment follows a close collaboration between Google, Meta, and the SGLang creators at the startup RadixArk. Previously, both SGLang and vLLM required Google’s JAX AI framework to operate on TPU hardware, but the new updates allow Meta to bypass framework rewrites and run its workloads directly on Google chips.
Meta’s PyTorch-TPU lead Zongwei Zhou confirmed via a Google developer blog post that the serving stack is operational in production. Jana van Greunen, director of PyTorch engineering at Meta, stated that the company cannot be constrained by a single hardware architecture and that the new TorchTPU framework enables the deployment of standard PyTorch workloads onto Google silicon. Google also published an open-source version of vLLM adapted for native TPU use alongside SGLang.
Meta resumes TPU hardware relationship with Google Cloud
This deployment represents a renewed hardware relationship between the two tech giants. SemiAnalysis reported that Meta research teams previously used TPUs via an older bridge known as PyTorch/XLA from 2020 until Meta canceled those contracts in 2023. Discussions regarding the current hardware deployment reportedly began with Google Cloud TPU rentals scheduled for 2026 and hardware purchases planned for 2027, culminating in a multi-year rental agreement.
The timing of the Google announcement coincided with competitive hardware maneuvers across the cloud industry. Amazon announced the open-sourcing of TorchNeuron, its own native PyTorch backend for Trainium2 and Trainium3 chips, at the same PyTorch Conference. Both tech companies are attempting to loosen Nvidia’s software lock-in by providing native software stacks for alternative accelerators.
RadixArk expands model compatibility for TPU serving stack
Development teams are expanding model compatibility for the new TPU serving stack. RadixArk indicated that SGLang-TorchTPU will incorporate support for DeepSeek V4.1 Flash and Reflection Beam next. The startup is targeting Day 0 TPU support for newly released open-source models by the first quarter of 2027.
Google, Meta, and RadixArk plan to present further technical details regarding the TPU integrations at the SGLang Summit 2026, scheduled for November 12–13 in San Francisco.
Frequently Asked Questions About Meta TPU Production Deployment
What changed to make Google TPUs work natively with Meta’s AI framework?
Google and Meta co-developed new PyTorch-native versions of the SGLang and vLLM inference engines. Previously, these inference engines required Google’s JAX AI framework to function on TPU hardware, forcing a framework migration that Meta wanted to avoid.
Where are Meta’s production TPUs physically located?
The exact physical location of the hardware remains undisclosed in public announcements. However, previous industry talks reportedly involved Google Cloud TPU rentals for 2026 with equipment purchases to follow in 2027, leading to a multi-year rental agreement.
What companies collaborated to build the new PyTorch-native TPU stack?
Engineers from Google, Meta, and RadixArk—a startup founded by the original creators of SGLang—co-developed the new serving stack. The vLLM creators’ startup, Inferact, also assisted Google in building the vLLM-TorchTPU integration.