Meta Shifts AI Workloads to Google Chips, Challenging Nvidia's Hold

Meta's AI models can now run on Google's custom AI chips. This changes how the company builds and runs its AI services. The new capability is available through a version of the SGLang inference engine that is native to PyTorch. A similar version of the vLLM inference engine is also now available as open source. This setup lets Meta use Google's tensor processing units, or TPUs, without a major rewrite of its code. Previously, these tools ran on TPUs using Google's own JAX framework.

The numbers behind this shift show a direct comparison of processing speed. A test by the startup Inferact found that a single user receiving tokens from Moonshot AI's 2.8-trillion-parameter Kimi K3 model could get over 700 tokens per second. This was on Google's seventh-generation Ironwood TPU. The test used 16 Ironwood chips with a method called speculative decoding. In a similar setup with 16 Nvidia GB200 chips, the speed was 452 tokens per second. For other models, the Google chips delivered about 1.4 to 2 times the throughput of the Nvidia baseline at small batch sizes.

Several people from Meta and Google were involved in the project. Zongwei Zhou, Meta's PyTorch-TPU lead, said the new SGLang-TorchTPU is running in production at Meta. He called the work a "close three-way collaboration between Meta, Google, and the SGLang team." That team works for RadixArk, a startup founded by the creators of SGLang. Jana van Greunen, director of PyTorch engineering at Meta, said her company "cannot be constrained by a single hardware architecture." She added that 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 the first time Meta has tried to use Google's TPUs. Google's PyTorch/XLA is considered a "predecessor" to the new TorchTPU. A report by SemiAnalysis said Meta research teams used TPUs through this older method from 2020 to 2023. Meta then canceled those contracts. Google's effort to make PyTorch work with TPUs was aimed mainly at Meta, which did not want to move to JAX.

The location of Meta's production TPUs is not clear. Talks last year included renting Google Cloud TPUs in 2026, with purchases in 2027. Meta later reportedly signed a multi-year rental agreement for these services.

This announcement came one day after Amazon open-sourced TorchNeuron. That tool is Amazon's own native PyTorch backend for its Trainium2 and Trainium3 chips. The move by Amazon was announced at the same PyTorch Conference. Both actions are seen as ways to reduce reliance on Nvidia's software.

The SGLang-TorchTPU engine will add more models. The first to be added are DeepSeek V4.1 Flash and Reflection Beam. The team is also working toward supporting new open models on their release day. This capability is planned for the first quarter of 2027, according to RadixArk. Google, Meta, and RadixArk plan to share more details at SGLang Summit 2026, which is set for Nov. 12-13 in San Francisco.