Meta’s Production Shift: Google’s TPU Chips Enter Major AI Infrastructure
A new PyTorch-native software bridge allows Meta to deploy Google's custom silicon in live environments, challenging the software dominance of Nvidia.
By Jonathan Hoffmann ·
The artificial intelligence hardware landscape underwent a significant shift this week as Meta confirmed that Google’s Tensor Processing Units (TPUs) are now actively serving production workloads within its infrastructure. This development marks the first time Google’s custom AI chips have been sold to an external customer for large-scale, production-grade use, signaling a potential fracture in the market dominance long held by Nvidia. The announcement was made via Google’s developer blog, highlighting a new PyTorch-native version of the SGLang inference engine that has been integrated into Meta’s operations.
Previously, running PyTorch workloads on TPUs required complex bridging through Google’s proprietary JAX framework. The new SGLang-TorchTPU engine eliminates this friction, allowing the hardware to function as a native option for PyTorch without extensive code rewriting. Zongwei Zhou, Meta’s lead for PyTorch-TPU initiatives, stated that the system is "running in production at Meta today." He credited the achievement to a "close three-way collaboration between Meta, Google, and the SGLang team." Zhou’s team operates out of RadixArk, a startup founded by the creators of SGLang, which co-developed the engine alongside the tech giants.
Meta executives emphasized the strategic necessity of hardware diversity. Jana van Greunen, director of PyTorch engineering at Meta, noted in the official post that the company "cannot be constrained by a single hardware architecture." She explained that the TorchTPU solution "enables us to deploy our PyTorch workloads to Google TPUs." Google confirmed that it co-developed this specific backend with Meta’s PyTorch team, underscoring a direct partnership aimed at reducing dependency on any single vendor.
This is not the first time Meta has experimented with Google silicon. Google describes PyTorch/XLA, an older bridge connecting PyTorch to TPUs, as a "predecessor" to the new TorchTPU. Industry reports from SemiAnalysis indicated that Meta research teams utilized TPUs via this older bridge between 2020 and 2023 before those contracts were canceled. The current push by Google was largely targeted at Meta, which had previously resisted migrating its core infrastructure to JAX. While the exact physical location of these production TPUs remains private, previous reports suggested talks included Google Cloud TPU rentals for 2026 with potential purchases in 2027, followed by a signed multi-year rental agreement.
The timing of the announcement is particularly notable. It arrived just one day after Amazon revealed it was open-sourcing TorchNeuron, its own native PyTorch backend for its Trainium2 and Trainium3 chips, at the same PyTorch Conference. Both moves appear designed to loosen the software lock-in that has historically favored Nvidia, giving major AI labs more leverage in hardware negotiations.
Performance benchmarks suggest the new setup is competitive. A tuned vLLM configuration serving Moonshot AI’s open-weight, 2.8-trillion-parameter Kimi K3 model reportedly achieves over 700 tokens per second for a single user on the seventh-generation Ironwood TPU. These figures come from Inferact, a startup founded by vLLM’s creators that assisted Google in building vLLM-TorchTPU. The test utilized 16 Ironwood chips with speculative decoding, a technique that accelerates large language model inference using a small draft model. In comparison, vLLM’s published setup on 16 Nvidia GB200s reached 452 tokens per second.
Further testing showed that the kernels for Kimi K3 and Qwen 3.8 27B deliver roughly 1.4 to 2 times the throughput of the GB200 baseline at small batch sizes, even without speculative decoding. This performance gap highlights the maturation of the TPU ecosystem for general-purpose AI workloads, moving beyond just Google-internal use cases.
Looking ahead, SGLang-TorchTPU will continue to expand its model support, starting with DeepSeek V4.1 Flash and Reflection Beam. RadixArk stated it is working toward Day 0 TPU support for new open models, meaning compatibility on release day, by the first quarter of 2027. Google, Meta, and RadixArk are scheduled to share further details at the SGLang Summit 2026, taking place November 12-13 in San Francisco.
