Meta Runs Google Chips in Production for First Time

700 tokens per second. That is the headline number for the seventh-generation Ironwood TPU when serving a massive AI model. This speed marks a significant shift in how major tech companies handle artificial intelligence workloads. It signals that Google’s custom chips are no longer just an internal tool. They are now a viable option for external giants like Meta. The move directly challenges the dominance of Nvidia in the AI hardware market.

For years, the path to using Google’s tensor processing units (TPUs) was difficult for companies built on other software stacks. The bridge existed, but it was clunky. Meta, the creator of the widely used PyTorch framework, had to use an older interface called PyTorch/XLA. Reports indicated that Meta research teams used this setup from 2020 to 2023. However, those early contracts were eventually canceled. The friction came from the fact that TPUs were natively designed for Google’s own JAX framework. Meta did not want to rewrite its entire codebase to fit that standard.

The situation changed with the recent announcement of SGLang-TorchTPU. This is a new version of the SGLang inference engine that runs natively on PyTorch. It removes the need for major code rewrites. A matching open-source version of vLLM, another popular inference engine, was also released. Both tools now allow TPUs to be a native option for PyTorch users. Zongwei Zhou, Meta’s lead for PyTorch-TPU, confirmed that the system is live. He stated that SGLang-TorchTPU is running in production at Meta today. He credited a close collaboration between Meta, Google, and the SGLang team.

The engineering work was led by RadixArk, a startup founded by the creators of SGLang. They built the engine in partnership with Google and Meta. Zhou described the result as a remarkably mature serving stack. Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic importance of this shift. She noted that Meta cannot be constrained by a single hardware architecture. The new TorchTPU tool allows the company to deploy its workloads to Google TPUs without being locked into one vendor.

This announcement follows a similar move by Amazon. Just one day prior, Amazon open-sourced TorchNeuron. This is a native PyTorch backend for its own Trainium2 and Trainium3 chips. Both companies are making similar moves at the same PyTorch Conference. The goal appears to be loosening Nvidia’s software lock-in. By providing native support for PyTorch, these companies make it easier for developers to run models on non-Nvidia hardware. This creates more competition in the AI chip market.

Performance benchmarks show why this matters. Inferact, a startup founded by the creators of vLLM, tested a tuned setup on the Ironwood TPU. They used the Kimi K3 model, which has 2.8 trillion parameters. The test used 16 chips with speculative decoding. This method speeds up inference using a smaller draft model. The result was over 700 tokens per second for a single user. By comparison, 16 Nvidia GB200 chips reached 452 tokens per second in similar tests. For smaller batch sizes without speculative decoding, the new kernels deliver roughly 1.4 to 2 times the throughput of the GB200 baseline.

The rollout is still in its early stages. SGLang-TorchTPU will continue to add support for more models. The next additions include DeepSeek V4.1 Flash and Reflection Beam. RadixArk is also working toward Day 0 support for new open models. This means the software will support new AI models on the day they are released. The target for this capability is the first quarter of 2027. Google, Meta, and RadixArk plan to share more details at the SGLang Summit 2026. The event is scheduled for November 12 and 13 in San Francisco.

The location of Meta’s production TPUs is not yet public. However, industry talks last year suggested a path forward. Reports indicated that Meta considered renting Google Cloud TPUs in 2026. Purchases were expected in 2027. Meta reportedly signed a multi-year rental agreement. This deal underscores the scale of the transition. It shows that Meta is willing to commit to Google’s hardware for its long-term AI needs. The partnership represents a significant step away from the exclusive reliance on Nvidia that has characterized the AI boom.