Meta Activates Google's TPUs to Break Nvidia's Hold

In a significant shift for the AI hardware market, Google's custom Tensor Processing Units are now powering production AI workloads at Meta, marking the first time these chips are being sold to external customers for this purpose. This development, announced via Google's developer blog, directly challenges Nvidia's long-standing control over the AI accelerator market by providing a major tech company with a viable, high-performance alternative to Nvidia GPUs. The integration is made possible through a new, PyTorch-native version of the SGLang inference engine, alongside a matching open-source version of vLLM, another widely used tool. Previously, running such workloads on TPUs required using Google's own JAX framework, but these new releases allow the hardware to be used as a native option for PyTorch, mostly without rewriting code.

The announcement outlines a detailed set of performance metrics and a complex web of corporate collaboration. The development of the new engines, SGLang-TorchTPU and vLLM-TorchTPU, was a close three-way collaboration involving Meta, Google, and the SGLang team, which works at RadixArk, a startup founded by SGLang's creators. "SGLang-TorchTPU is running in production at Meta today," stated Meta's PyTorch-TPU lead, Zongwei Zhou, adding that the engine provides Meta "a remarkably mature, PyTorch-native serving stack." This sentiment was echoed by Jana van Greunen, director of PyTorch engineering at Meta, who noted in Google's TorchTPU post that Meta "cannot be constrained by a single hardware architecture" and that TorchTPU "enables us to deploy our PyTorch workloads to Google TPUs." Google also confirmed it co-developed TorchTPU with Meta's PyTorch team, solidifying the partnership behind the technology.

On performance, a tuned vLLM setup serving Moonshot AI's open-weight, 2.8-trillion-parameter Kimi K3 model reportedly tops 700 tokens per second for a single user on the seventh-generation Ironwood TPU. That speed figure comes from Inferact, the startup founded by vLLM's creators that helped Google build vLLM-TorchTPU. Its test ran 16 Ironwood chips with speculative decoding, a method of speeding up large language model inference with a small draft model, against vLLM's published setup on 16 Nvidia GB200s, which reached 452 tokens per second. Its kernels for Kimi K3 and Qwen 3.8 27B deliver roughly 1.4 to 2 times the GB200 baseline's throughput at small batch sizes without speculative decoding.

The partnership between Google and Meta on TPUs is not entirely new. Google describes the previous bridge from PyTorch to TPUs, called PyTorch/XLA, as a "predecessor" to the new TorchTPU. According to a report from SemiAnalysis, Meta research teams used TPUs through this older bridge from 2020 to 2023 before Meta canceled those contracts. Google's current push to enable PyTorch on TPUs was aimed mainly at Meta, as the company did not want to migrate its AI workloads to Google's JAX framework. The current agreement's structure is still being clarified; while it is unclear where Meta's production TPUs sit, talks last year reportedly included Google Cloud TPU rentals in 2026 with purchases in 2027, and Meta later reportedly signed a multi-year rental agreement. This arrangement provides Meta with the hardware it needs while offering Google a new, high-value customer for its chips, directly competing with Nvidia's market dominance.

The announcement came one day after Amazon said it would be open-sourcing TorchNeuron, its own native PyTorch backend for its Trainium2 and Trainium3 chips, at the same PyTorch Conference. This timing highlights a broader industry effort to loosen Nvidia's software lock-in by making it easier for developers to run AI models on other companies' hardware. For Meta, this means it is not tied to a single hardware architecture, which could provide it with more flexibility and leverage in its AI infrastructure planning. The release of SGLang-TorchTPU is also an ongoing project. RadixArk stated that it will keep adding models, starting with DeepSeek V4.1 Flash and Reflection Beam. The team is also working toward providing Day 0 support for new open models, meaning support on release day, by the first quarter of 2027. Google, Meta, and RadixArk plan to share more details at SGLang Summit 2026, Nov. 12-13 in San Francisco.

This development is a clear attempt to provide an alternative to Nvidia's market dominance by enabling Google's TPUs to run AI workloads for Meta. The new release of SGLang-TorchTPU and vLLM-TorchTPU allows for a direct comparison with Nvidia's GB200s, which reached 452 tokens per second in a similar test. The development of these tools was a close three-way collaboration between Meta, Google, and the SGLang team, which works at RadixArk, a startup founded by SGLang's creators. This work is being done as part of a broader effort to loosen Nvidia's software lock-in, as evidenced by Amazon's simultaneous announcement regarding TorchNeuron for its Trainium2 and Trainium3 chips. The availability of these tools is a significant step in providing a high-performance alternative to Nvidia's hardware for running AI models.