Meta now uses Google TPUs for AI workloads in production
Meta has deployed Google’s tensor processing units in production for its AI workloads.
By Tariq Mahmoud ·
Meta now runs its AI workloads in production using Google’s tensor processing units (TPUs), thanks to a new PyTorch-native version of the SGLang inference engine. This update, announced on Google’s developer blog, also includes a matching version of vLLM, another widely used inference engine, which was released as open source. Previously, both engines used Google’s JAX AI framework to run on TPUs, but the new versions allow TPUs to be a native option for PyTorch with minimal rewriting.
Zongwei Zhou, Meta’s PyTorch-TPU lead, said the new engine gives Meta “a remarkably mature, PyTorch-native serving stack.” Jana van Greunen, director of PyTorch engineering at Meta, added that Meta “cannot be constrained by a single hardware architecture” and that the new TorchTPU “enables us to deploy our PyTorch workloads to Google TPUs.” Google said it co-developed TorchTPU with Meta’s PyTorch team.
This is not Meta’s first attempt at using TPUs. Google calls PyTorch/XLA, the older bridge from PyTorch to TPUs, a “predecessor” to the new TorchTPU. Meta research teams used TPUs through this bridge from 2020 to 2023 before Meta canceled those contracts. Google’s PyTorch-on-TPU push was aimed mainly at Meta, which did not want to migrate to JAX.
Although it is unclear where Meta’s production TPUs sit, talks last year reportedly included Google Cloud TPU rentals in 2026 with purchases in 2027, and Meta later reportedly signed a multi-year rental agreement. The announcement came one day after Amazon said it was open-sourcing TorchNeuron, its own native PyTorch backend for its Trainium2 and Trainium3 chips, at the same PyTorch Conference. Both moves appear to be aimed at loosening Nvidia’s software lock-in.
On performance, a tuned vLLM setup serving Moonshot AI’s open-weight, 2.8-trillion-parameter Kimi K3 model reportedly tops 700 tokens per second for a single user on the seventh-generation Ironwood TPU. This speed figure comes from Inferact, the startup founded by vLLM’s creators that helped Google build vLLM-TorchTPU. Its test ran 16 Ironwood chips with speculative decoding, a method of speeding up large language model inference with a small draft model, against vLLM’s published setup on 16 Nvidia GB200s, which reached 452 tokens per second. Its kernels for Kimi K3 and Qwen 3.8 27B deliver roughly 1.4 to 2 times the GB200 baseline’s throughput at small batch sizes without speculative decoding.
SGLang-TorchTPU will keep adding models, starting with DeepSeek V4.1 Flash and Reflection Beam. It’s also working toward Day 0 TPU support for new open models, meaning support on release day, by the first quarter of 2027, according to RadixArk. Google, Meta and RadixArk plan to share more at SGLang Summit 2026, Nov. 12–13 in San Francisco.
Source: Tom's Hardware
