Meta's Production Shift to Google Chips Signals Major Challenge to Nvidia's AI Dominance
A new PyTorch-native engine allows Meta to run Google's AI chips in production, marking the first time TPUs have been sold to external customers and intensifying the challenge to Nvidia's market position.
By Christopher Reed ·
Meta’s artificial intelligence infrastructure has entered a new phase of diversification, as the tech giant confirmed that Google’s tensor processing units (TPUs) are now running production workloads. This development, announced via Google’s developer blog, marks a significant milestone in the integration of custom silicon outside of Google’s own ecosystem. The shift is facilitated by a new PyTorch-native version of the SGLang inference engine, which allows Meta to utilize the hardware without extensive code rewrites, a critical step for large-scale deployment.
The transition represents a departure from previous reliance on Google’s JAX framework. By making TPUs a native option for PyTorch, the software stack becomes accessible to a broader range of developers who have standardized on the PyTorch library. Zongwei Zhou, Meta’s PyTorch-TPU lead, noted that SGLang-TorchTPU is already in production at Meta, citing a close three-way collaboration between Meta, Google, and the SGLang team. This engine, developed by RadixArk, a startup founded by SGLang’s creators, provides what Zhou described as a remarkably mature serving stack.
Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic necessity of this move. She stated that Meta cannot be constrained by a single hardware architecture and that the new TorchTPU backend enables the deployment of PyTorch workloads to Google TPUs. Google confirmed that it co-developed this specific backend with Meta’s PyTorch team, highlighting the collaborative nature of the technical integration.
This is not the first instance of Meta engaging with TPUs, though previous efforts were more limited. Google identifies PyTorch/XLA as the predecessor to the current TorchTPU. Reports from SemiAnalysis indicated that Meta research teams utilized TPUs through this older bridge between 2020 and 2023 before canceling those contracts. The current push appears to be a more robust attempt to address Meta’s preference to avoid migrating entirely to JAX, with recent talks reportedly involving cloud rentals in 2026 and potential purchases in 2027.
The announcement arrived just one day after Amazon revealed it was open-sourcing TorchNeuron, a native PyTorch backend for its own Trainium chips. Both moves appear designed to loosen Nvidia’s software lock-in, which has long dominated the AI hardware market. Performance metrics underscore the viability of the new setup; a tuned vLLM configuration on Google’s seventh-generation Ironwood TPU reportedly exceeds 700 tokens per second for a single user when serving the Kimi K3 model. This compares favorably to 452 tokens per second achieved on Nvidia GB200s in similar tests.
Looking ahead, SGLang-TorchTPU plans to expand model support, starting with DeepSeek V4.1 Flash and Reflection Beam. RadixArk aims to achieve Day 0 TPU support for new open models by the first quarter of 2027. Further details regarding the collaboration and future roadmap are expected to be shared at the SGLang Summit 2026, scheduled for November 12–13 in San Francisco.
Source: Tom's Hardware
