Meta Runs Google TPUs in Production, Marking First External Sale of the Custom AI Chips

For the first time, the custom AI chips designed by Google are running outside of Google's own infrastructure. Meta, the company behind Facebook and Instagram, is now using these tensor processing units in production environments. This shift means that large technology firms can now rely on Google's specialized hardware for their daily AI operations without building their own chips or strictly adhering to Google's internal software frameworks. The move signals a significant change in the market for artificial intelligence compute resources.

The transition was made possible by new versions of popular inference engines. Specifically, SGLang and vLLM, two widely used tools for running AI models, now have native support for PyTorch on Google TPUs. Previously, developers had to use Google's JAX framework to run these chips. The new updates allow PyTorch users to deploy their models on TPUs with minimal code changes. This reduces the technical friction for companies that have built their AI stacks around PyTorch, which is the dominant framework in the industry.

Zongwei Zhou, the PyTorch-TPU lead at Meta, confirmed that the SGLang-TorchTPU engine is live in production. He described the achievement as the result of a close three-way collaboration between Meta, Google, and the SGLang team. The SGLang team operates out of RadixArk, a startup founded by the creators of the engine. Zhou noted that the new tool provides Meta with a remarkably mature serving stack that is native to PyTorch. This allows Meta to manage its AI workloads without needing to rewrite existing code.

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. The new TorchTPU capability enables the company to deploy its PyTorch workloads directly to Google TPUs. Google has stated that it co-developed this technology with Meta's PyTorch team. This partnership highlights a growing trend of major tech firms collaborating to ensure flexibility in their AI infrastructure.

This is not the first time Meta has experimented with Google's hardware. An older bridge called PyTorch/XLA allowed some level of compatibility, but Google now considers it a predecessor to the new system. Reports from last year indicated that Meta research teams used TPUs through this older bridge between 2020 and 2023. However, those contracts were eventually canceled. The current push by Google appears specifically aimed at winning back Meta, which had previously resisted migrating its entire stack to Google's JAX framework.

The logistics of where Meta's production TPUs are located remain somewhat unclear. However, industry talks from last year suggested that Google Cloud TPU rentals would begin in 2026, with direct purchases potentially following in 2027. Meta is reportedly the signatory of a multi-year rental agreement. This arrangement allows Meta to access the hardware without the immediate capital expenditure of buying the chips outright, a common strategy for scaling AI capabilities rapidly.

The announcement arrived just one day after Amazon open-sourced TorchNeuron. TorchNeuron is a native PyTorch backend for Amazon's Trainium2 and Trainium3 chips. Both the Google and Amazon moves appear designed to loosen the software lock-in that has benefited Nvidia. By providing native PyTorch support for their respective hardware, both companies are lowering the barrier to entry for developers who want to use alternatives to Nvidia's GPUs.

Performance benchmarks suggest the new setup is competitive. A tuned vLLM configuration serving the Kimi K3 model, which has 2.8 trillion parameters, reportedly exceeds 700 tokens per second for a single user on the seventh-generation Ironwood TPU. This test was conducted by Inferact, a startup that helped build the vLLM-TorchTPU integration. The test used 16 Ironwood chips with speculative decoding. In comparison, a similar setup on 16 Nvidia GB200s reached 452 tokens per second.

For smaller batch sizes without speculative decoding, the new kernels for Kimi K3 and Qwen 3.8 27B deliver roughly 1.4 to 2 times the throughput of the GB200 baseline. The SGLang-TorchTPU project plans to add more models, starting with DeepSeek V4.1 Flash and Reflection Beam. The team is also working toward Day 0 support for new open models, meaning they aim to provide compatibility on release day by the first quarter of 2027. Further details are expected at the SGLang Summit in San Francisco next month.

This development marks a pivotal moment for the AI hardware market. As Google sells its TPUs to external customers for the first time, the dominance of Nvidia faces a more direct challenge. The availability of native PyTorch support for major alternative chips gives developers more choices. For ordinary users, this competition could lead to faster and more efficient AI services as companies optimize their infrastructure costs.