Meta Activates Google TPUs via New PyTorch Integration

Meta has commenced production-level utilization of Google’s tensor processing units (TPUs) through the deployment of a new PyTorch-native version of the SGLang inference engine, a milestone that marks the first time Google has sold its custom AI chips to an external customer at scale. This development, announced on Google’s developer blog, represents a critical departure from the industry’s heavy reliance on Nvidia’s ecosystem, as the new software stack allows Meta to run its workloads directly on TPU hardware without extensive code rewriting.

The integration of SGLang-TorchTPU into Meta’s infrastructure is the result of a collaborative effort involving Meta, Google, and the SGLang team, which is part of the startup RadixArk. Zongwei Zhou, Meta’s lead for PyTorch-TPU, confirmed that the system is currently live in production. He described the outcome as a remarkably mature, PyTorch-native serving stack, highlighting the efficiency of the three-way partnership in bridging the gap between Google’s hardware and Meta’s preferred software framework.

Strategic flexibility has been a primary driver for this transition. Jana van Greunen, director of PyTorch engineering at Meta, stated that the company cannot be constrained by a single hardware architecture. By adopting TorchTPU, Meta can now deploy its PyTorch workloads to Google’s TPUs, a capability previously hindered by the reliance on Google’s JAX framework. This new native option offers a seamless alternative for developers who prefer PyTorch, effectively lowering the barrier to entry for TPU-based AI operations.

This marks a renewed commitment to TPU adoption for Meta, building upon earlier, less successful attempts. While Meta had previously used TPUs through the older PyTorch/XLA bridge between 2020 and 2023, those contracts were eventually canceled. The current push by Google appears specifically tailored to accommodate Meta’s refusal to migrate to JAX, suggesting a more targeted approach to securing major cloud customers for its custom silicon.

The timing of this announcement is notable, as it follows Amazon’s decision to open-source TorchNeuron, a native PyTorch backend for its Trainium chips, by just one day. Both moves appear designed to challenge Nvidia’s dominant position in AI software and hardware. In terms of performance, a tuned vLLM setup on the seventh-generation Ironwood TPU reportedly exceeds 700 tokens per second for the Kimi K3 model, outperforming the 452 tokens per second achieved by Nvidia’s GB200s in comparable tests.

Looking ahead, the ecosystem is set for rapid expansion. SGLang-TorchTPU is scheduled to support additional models, including DeepSeek V4.1 Flash, with a goal of achieving Day 0 support for new open models by the first quarter of 2027. Further details regarding the collaboration will be presented at the SGLang Summit 2026 in San Francisco next month, underscoring the growing momentum behind alternative AI hardware architectures.

The broader implications of these developments suggest a gradual but steady diversification of the AI infrastructure market. As major players like Meta and Amazon invest in native PyTorch support for non-Nvidia chips, the software lock-in that has defined the industry may begin to loosen, offering more options for organizations seeking to optimize their AI deployment strategies.