Meta Deploys Google TPUs in Production, Signaling Shift in AI Hardware Market

Google’s custom silicon is now actively processing production AI workloads for Meta, marking the first time the company’s Tensor Processing Units (TPUs) have been sold to an external customer for such a purpose. This development, confirmed on Wednesday, represents a tangible step away from the industry’s default reliance on Nvidia hardware, driven by a new software bridge that allows Meta’s models to run natively on PyTorch without extensive code rewrites.

The catalyst for this shift is the release of SGLang-TorchTPU, an inference engine optimized for Google’s chips. Meta’s Zongwei Zhou, who leads the PyTorch-TPU effort, stated that the system is currently running in production environments at Meta. Zhou highlighted that this capability emerged from a three-way collaboration involving Meta, Google, and the SGLang team, which is based out of RadixArk, a startup founded by the original creators of the SGLang engine.

From a software architecture perspective, this move eliminates the need for Meta to migrate its codebase to Google’s JAX framework, which was previously required to utilize TPUs. By integrating directly with PyTorch, the dominant deep learning framework, Google has removed a major barrier to entry for large-scale AI developers. Jana van Greunen, a director of PyTorch engineering at Meta, emphasized that the company cannot be constrained by a single hardware architecture and that this new stack enables the deployment of PyTorch workloads directly onto Google hardware.

This is not the first instance of Meta and Google intersecting on hardware, but the current implementation is significantly more mature. In previous years, Meta research teams utilized TPUs through a bridge called PyTorch/XLA between 2020 and 2023 before those contracts were canceled. The new TorchTPU initiative, co-developed with Meta’s engineering team, is described by Google as a successor to that older method, offering a more robust and native integration.

The commercial terms of the arrangement remain partially opaque, though industry reports from last year suggest a phased approach. Discussions reportedly included the rental of Google Cloud TPUs in the current year, with potential hardware purchases scheduled for 2027. Meta is also understood to have signed a multi-year rental agreement, suggesting a long-term commitment to diversifying its compute infrastructure beyond Nvidia’s GPUs.

Market analysts view this announcement as part of a broader trend to loosen Nvidia’s software lock-in. The news arrived just one day after Amazon announced the open-sourcing of TorchNeuron, its native PyTorch backend for its Trainium2 and Trainium3 chips. Both moves signal that major technology firms are prioritizing hardware diversity to mitigate supply chain risks and optimize costs for large-scale AI inference.

Performance benchmarks suggest the new setup is competitive with industry standards. According to Inferact, a startup that assisted in building the vLLM-TorchTPU engine, a tuned configuration serving the 2.8-trillion-parameter Kimi K3 model achieved over 700 tokens per second for a single user. This figure was generated using 16 seventh-generation Ironwood TPUs with speculative decoding, a technique that uses a smaller model to draft outputs. In comparison, Nvidia’s GB200 hardware achieved 452 tokens per second under similar conditions.

Further performance data indicates that the new kernels for Kimi K3 and Qwen 3.8 27B deliver throughput that is roughly 1.4 to two times higher than the GB200 baseline at small batch sizes, provided speculative decoding is not employed. This efficiency gain is critical for workloads where low latency and high single-user throughput are prioritized over massive parallel batch processing.

Looking ahead, the ecosystem supporting these chips is expanding rapidly. RadixArk plans to add support for new models, including DeepSeek V4.1 Flash and Reflection Beam, to the SGLang-TorchTPU stack. The team is also targeting "Day 0" support for new open-source models, meaning compatibility at the time of release, by the first quarter of 2027. Google, Meta, and RadixArk are scheduled to share further details about this collaboration at the SGLang Summit 2026, which is set to take place in San Francisco in November.

Ultimately, the entry of TPUs into Meta’s production environment signals a maturing market where hardware choices are increasingly dictated by software compatibility and performance efficiency rather than brand loyalty alone. As the major players continue to build out their respective ecosystems, the competitive dynamics of the AI chip market are poised for continued disruption.