Meta Runs Google Chips in Production

The rivalry between tech giants often feels like a closed club, but a recent development has cracked the door open. Google has announced that its custom Tensor Processing Units (TPUs) are now actively serving AI workloads in production at Meta, the company behind the widely used PyTorch framework.

This milestone is made possible by a new, PyTorch-native version of the SGLang inference engine, which was detailed on Google’s developer blog. Simultaneously, an open-source version of vLLM, another major inference engine, was released. Previously, running these engines on TPUs required using Google’s internal JAX framework. The new tools change that dynamic, allowing the hardware to function as a native option for PyTorch developers without the need for extensive code rewrites.

The numbers behind this transition are striking, particularly regarding speed. According to Inferact, a startup founded by the creators of vLLM that assisted Google in building the new software, a tuned vLLM setup on the seventh-generation Ironwood TPU can exceed 700 tokens per second for a single user. This test involved running the 2.8-trillion-parameter Kimi K3 model from Moonshot AI. For comparison, the same model ran at 452 tokens per second on Nvidia’s GB200 chips in a published setup. Furthermore, new kernels for the Kimi K3 and Qwen 3.8 27B models deliver roughly 1.4 to 2 times the throughput of the GB200 baseline at small batch sizes, even without speculative decoding.

Key figures from both companies have highlighted the strategic importance of this move. Zongwei Zhou, Meta’s PyTorch-TPU lead, noted that the new engine provides "a remarkably mature, PyTorch-native serving stack." He emphasized that SGLang-TorchTPU is "running in production at Meta today," a result of a close collaboration between Meta, Google, and the SGLang team, which operates out of RadixArk. Jana van Greunen, Meta’s director of PyTorch engineering, added that the company "cannot be constrained by a single hardware architecture." She stated that TorchTPU "enables us to deploy our PyTorch workloads to Google TPUs," a capability Google says it co-developed directly with Meta’s engineering team.

This is not the first time Meta has experimented with Google’s silicon. Google identifies PyTorch/XLA, an older bridge between the two ecosystems, as a predecessor to the current TorchTPU. Reports from SemiAnalysis indicated last year that Meta’s research teams utilized TPUs through this older interface between 2020 and 2023, before those contracts were canceled. The current push appears to be a more robust attempt to accommodate Meta, which had previously resisted migrating its core operations to JAX.

The location of these new production units remains somewhat private, though industry talks last year suggested Google Cloud TPU rentals would begin in 2026, with potential purchases following in 2027. Meta is reportedly signed on for a multi-year rental agreement. The timing of this announcement is also significant; it arrived just one day after Amazon revealed it was open-sourcing TorchNeuron, its native PyTorch backend for Trainium chips, at the same PyTorch Conference.

The stakes are high as both moves seem designed to loosen Nvidia’s dominant software lock-in. By making alternative hardware easier to integrate with the industry-standard PyTorch framework, Google and Meta are creating a viable path for other developers. RadixArk plans to continue adding models to the SGLang-TorchTPU stack, starting with DeepSeek V4.1 Flash and Reflection Beam. The team is also targeting "Day 0" support for new open models by the first quarter of 2027. More details are expected to be shared at the SGLang Summit 2026, scheduled for November 12–13 in San Francisco.