Meta Turns to Google TPUs to Run AI Workloads

Meta is now running production artificial intelligence workloads on Google's custom processors, a significant step that allows the social media giant's software to operate on hardware outside of the graphics processing units it has long relied on. This capability comes through a new version of the SGLang inference engine that is native to PyTorch, the artificial intelligence framework developed and maintained by Meta. Google announced the release of this new software, along with a matching version for the vLLM inference engine, on its developer blog.

The move represents a concrete step in a strategy to reduce reliance on a single type of hardware. According to Jana van Greunen, Meta's director of PyTorch engineering, the company cannot be constrained to one architecture. She explained in a post that the new TorchTPU software enables the deployment of PyTorch workloads onto Google's tensor processing units. This new software was developed collaboratively by Google and Meta's PyTorch team.

The development marks a reversal in Meta's approach to these specific chips. Zongwei Zhou, who leads PyTorch-TPU development at Meta, announced that the new system, SGLang-TorchTPU, is already running in production at the company. He described the result of a close collaboration between Meta, Google, and the SGLang team as a remarkably mature, PyTorch-native serving stack.

This production deployment follows an earlier, less integrated attempt to use Google's technology. In the past, Meta's research teams did use these processors from 2020 to 2023 through a software bridge called PyTorch/XLA, an older tool that Google now describes as a predecessor to the new TorchTPU. Those earlier contracts were eventually canceled. The current push for native PyTorch support on these chips was driven by Meta's desire to avoid migrating its codebase to the JAX framework.

The arrangement for this new production access to Google's silicon is a commercial one. Discussions last year reportedly covered the rental of Google Cloud processing capacity in 2026 and the purchase of such hardware in 2027. Meta subsequently signed a multi-year agreement for the rental of this processing power, though the physical location of these chips within Meta's data centers has not been disclosed.

The announcement arrived one day after a similar move by Amazon, which open-sourced its own native PyTorch backend for its custom chips at the same PyTorch Conference. Both efforts are aimed at loosening Nvidia's hold on the software ecosystem, which has been a key factor in its market leadership. By enabling other hardware to run the PyTorch framework, these companies are creating direct alternatives to Nvidia's graphics processing units for running artificial intelligence models.

Early performance numbers from a partnership involving Google suggest the new configuration can be competitive. A tuned setup of the vLLM engine running on a seventh-generation Ironwood TPU is reported to exceed 700 tokens per second for a single user on a large, open-weight model. This test compared 16 of those chips against 16 Nvidia GB200 units, which were reported to reach 452 tokens per second. The test was conducted by Inferact, a company founded by the creators of vLLM.

Work on the software is ongoing. The SGLang-TorchTPU project plans to add support for more models, beginning with DeepSeek V4.1 Flash and Reflection Beam. The team at RadixArk, the startup behind SGLang, also stated it is working toward providing support for new open models on their release day by the first quarter of 2027. Further details are planned for an upcoming SGLang Summit in San Francisco on Nov. 12-13.