Meta Deploys Google's AI Chips in Production

Meta has begun running Google's proprietary tensor processing units (TPUs) within its own infrastructure to power live AI services, marking the first time this silicon is deployed to external customers. The deployment relies on a newly released, PyTorch-native iteration of the SGLang inference engine, as detailed in a post on Google’s developer blog. In coordination with this release, a corresponding version of the vLLM inference engine has also been made available as open-source software.

This development is technically significant because it circumvents the need to rewrite existing code. Previously, running these specific tools on TPU hardware required the use of Google's JAX AI framework. The new iterations of SGLang and vLLM allow developers to utilize PyTorch on TPUs without such substantial modifications.

Zongwei Zhou, who leads PyTorch-TPU efforts at Meta, described the collaboration as a three-way effort between Meta, Google, and the SGLang team. He stated that the software, developed in conjunction with Google and the startup RadixArk, provides Meta with "a remarkably mature, PyTorch-native serving stack." Jana van Greunen, director of PyTorch engineering at Meta, noted in Google’s TorchTPU post that Meta "cannot be constrained by a single hardware architecture" and that TorchTPU "enables us to deploy our PyTorch workloads to Google TPUs."

This is not the first time Meta has engaged with TPU technology, though previous attempts were unsuccessful. Google identifies PyTorch/XLA, an older bridge between PyTorch and TPUs, as a "predecessor" to the current TorchTPU. According to SemiAnalysis, Meta research teams utilized TPUs through this bridge from 2020 to 2023 before canceling those contracts. Google indicated that its PyTorch-on-TPU push was directed primarily at Meta, which declined to migrate to JAX. Regarding current infrastructure, it is not yet confirmed where these production TPUs are located. Talks last year reportedly involved renting Google Cloud TPU capacity in 2026 with purchases in 2027, and Meta later reportedly signed a multi-year rental agreement.

The announcement arrived one day after Amazon disclosed that it would open-source TorchNeuron, its own native PyTorch backend for its Trainium2 and Trainium3 chips. Both announcements were made at the PyTorch Conference. These parallel releases appear designed to loosen Nvidia’s software lock-in.

In terms of performance metrics, a tuned vLLM setup serving Moonshot AI’s open-weight, 2.8-trillion-parameter Kimi K3 model reportedly exceeds 700 tokens per second for a single user on the seventh-generation Ironwood TPU. This figure originates from Inferact, the startup founded by vLLM’s creators that assisted Google in building vLLM-TorchTPU. Its test involved 16 Ironwood chips using speculative decoding, a technique to accelerate large language model inference via a small draft model, compared 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.

Future updates for the engine are scheduled. SGLang-TorchTPU will continue to expand its model support, with DeepSeek V4.1 Flash and Reflection Beam listed as upcoming additions. RadixArk stated it is working toward Day 0 TPU support for new open models, providing support on release day, by the first quarter of 2027. Google, Meta, and RadixArk plan to share further details at SGLang Summit 2026, Nov. 12-13 in San Francisco.