Google’s TPUs for AI workloads now run in Meta's data centers

Google's tensor processing units (TPUs), custom AI chips, are now providing AI support to Meta, the creator of the widely used PyTorch AI framework. A new PyTorch-native version of the SGLang inference engine, announced on Google's developer blog, enables this. A matching version of vLLM, another commonly used inference engine, was also made open source.

Prior to this, both engines relied on Google's own JAX AI framework to run on TPUs. The updated versions allow the hardware to be a native choice for PyTorch, largely without the need for extensive rewriting.

The collaboration between Meta, Google, and the SGLang team resulted in the creation of the SGLang-TorchTPU, which is already in use at Meta. Meta's PyTorch-TPU lead, Zongwei Zhou, revealed that the team that developed the engine is based at RadixArk, a startup founded by the creators of SGLang. According to Zhou, RadixArk, Meta, and Google worked together to produce an engine that provides Meta with a highly mature, PyTorch-native serving stack.

Meta director of PyTorch engineering, Jana van Greunen, emphasized in Google's TorchTPU post that Meta is not tied to a single hardware architecture and that TorchTPU enables them to deploy their PyTorch workloads to Google TPUs.

Though Meta has not attempted to use TPUs before, Google refers to PyTorch/XLA, the older bridge from PyTorch to TPUs, as a predecessor to the new TorchTPU. SemiAnalysis reported last year that Meta's research teams used TPUs through this bridge from 2020 to 2023 before Meta ended those contracts. Google's push for PyTorch-on-TPU was primarily aimed at Meta, as the latter did not desire to transition to JAX.

The announcement of this collaboration occurred one day after Amazon announced the open-sourcing of TorchNeuron, its native PyTorch backend for its Trainium2 and Trainium3 chips. Both moves aim to reduce Nvidia's software lock-in.

In terms of performance, a tuned vLLM setup on the seventh-generation Ironwood TPU reportedly handles Moonshot AI's open-weight, 2.8-trillion-parameter Kimi K3 model at a rate of over 700 tokens per second for a single user. This speed figure, shared by Inferact, the startup founded by vLLM's creators that assisted Google in building vLLM-TorchTPU, is achieved with 16 Ironwood chips and speculative decoding.

The same setup on 16 Nvidia GB200s reached 452 tokens per second. The kernels for Kimi K3 and Qwen 3.8 27B deliver roughly 1.4 to 2 times the throughput of the GB200 baseline at small batch sizes without speculative decoding. SGLang-TorchTPU will continue to support additional models, starting with DeepSeek V4.1 Flash and Reflection Beam.

It also aims to achieve Day 0 TPU support for new open models by the first quarter of 2027, according to RadixArk. Google, Meta, and RadixArk plan to share more details about the development and progress of this technology at the SGLang Summit 2026, scheduled for November 12–13 in San Francisco.