Meta Deploys Google TPU Chips in Production, Challenging Nvidia's Grip

The landscape for artificial intelligence hardware is shifting as major tech firms seek alternatives to the dominant graphics processing unit ecosystem. Google announced that its Tensor Processing Units, or TPUs, are now serving AI workloads in production at Meta. This milestone was revealed through the release of a PyTorch-native version of the SGLang inference engine, a move that signals a deeper integration between the two companies' AI infrastructures.

The development allows Meta to run its AI models on Google's custom chips without relying on Google's proprietary JAX framework. Instead, the new versions of SGLang and the vLLM inference engine enable native support for PyTorch. This compatibility reduces the need for extensive code rewriting, making the hardware a more practical option for organizations that have built their systems around the PyTorch ecosystem.

Zongwei Zhou, Meta's lead for PyTorch and TPU integration, confirmed that the SGLang-TorchTPU system is actively running in production. He described the achievement as the result of a tight collaboration between Meta, Google, and the SGLang team. This team operates at RadixArk, a startup founded by the creators of SGLang, which worked closely with Google and Meta to build the new engine. Zhou noted that the setup provides Meta with a mature and native serving stack for PyTorch.

Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic importance of hardware diversity. She stated that Meta cannot be constrained by a single hardware architecture. By adopting TorchTPU, the company can deploy its PyTorch workloads directly onto Google's infrastructure. Google confirmed that it co-developed this solution with Meta's PyTorch team, highlighting the joint effort to bridge software and hardware gaps.

This is not Meta's first foray into using Google's silicon. The older PyTorch/XLA bridge served as a predecessor to the current TorchTPU system. Reports from SemiAnalysis indicated that Meta research teams utilized TPUs through this older bridge between 2020 and 2023. However, Meta eventually canceled those contracts. The recent push for native PyTorch support on TPUs was largely aimed at Meta, which had previously resisted migrating to Google's JAX framework.

While the exact location of these production TPUs remains unclear, industry talks reportedly included plans for Google Cloud TPU rentals in 2026 with potential purchases in 2027. Meta is also believed to have signed a multi-year rental agreement. The timing of this announcement is notable, as it came one day after Amazon opened the source code for TorchNeuron. This is Amazon's native PyTorch backend for its Trainium chips, and both moves appear designed to loosen the software lock-in provided by Nvidia.

Performance metrics suggest the new setup is competitive. Inferact, a startup founded by vLLM creators, reported that a tuned vLLM setup on seventh-generation Ironwood TPUs achieves over 700 tokens per second for a single user. This test used speculative decoding to speed up inference for a large open-weight model. In comparison, the same setup on Nvidia GB200 chips reached 452 tokens per second. For smaller batch sizes without speculative decoding, the TPU kernels deliver roughly 1.4 to 2 times the throughput of the Nvidia baseline.

The SGLang-TorchTPU ecosystem will continue to expand with support for additional models, including DeepSeek V4.1 Flash and Reflection Beam. RadixArk aims to provide Day 0 TPU support for new open models by the first quarter of 2027. Google, Meta, and RadixArk plan to share further details at the SGLang Summit 2026 in San Francisco.

This collaboration marks a significant step in diversifying the AI hardware market. By making Google's TPUs a native option for PyTorch users, the industry sees a direct challenge to the status quo. As more companies seek alternatives to Nvidia's dominance, the integration of custom silicon with popular software frameworks becomes a critical battleground for future AI development.