Google's TPUs Find a Home in Meta's AI Infrastructure

For everyday users, the most immediate impact of this shift is likely to be faster and more reliable responses from the AI tools they use daily. By integrating Google’s custom silicon into Meta’s infrastructure, the company gains a powerful alternative to the dominant graphics processing units it has relied on for years. This move suggests that the computing power behind large language models is becoming less dependent on a single vendor, which could eventually translate to better performance and lower costs for the services that power our digital lives.

The announcement, released on October 22, details how Google’s Tensor Processing Units are now serving AI workloads in production at Meta. This is achieved through a new, PyTorch-native version of the SGLang inference engine. Previously, running these tasks on Google’s hardware required using JAX, a different AI framework. The new setup allows developers to use PyTorch, the standard framework for much of the industry, without needing to rewrite their existing code. This makes Google’s chips a native, viable option for a much wider range of software developers.

This is not the first time Meta and Google have explored this connection. In the past, Meta used a bridge called PyTorch/XLA to access Google’s hardware. However, reports indicated that Meta canceled those contracts after using the system from 2020 to 2023. The current approach is more direct. Zongwei Zhou, Meta’s lead for PyTorch and TPUs, noted that the new system is a result of a tight collaboration between Meta, Google, and the team behind SGLang. He described the outcome as a mature serving stack that fits naturally into Meta’s existing workflows.

Jana van Greunen, a director of PyTorch engineering at Meta, emphasized the strategic importance of this flexibility. She stated that Meta cannot allow itself to be limited by a single hardware architecture. By adopting this new TorchTPU software, Meta can deploy its PyTorch-based workloads onto Google’s TPUs with greater ease. Google confirmed that it co-developed this specific software layer with Meta’s engineering team to ensure compatibility.

The timing of this news is particularly notable. It was announced just one day after Amazon revealed it was open-sourcing its own native PyTorch backend for its custom AI chips. Both moves appear designed to challenge the software dominance held by Nvidia. By making their respective hardware more accessible to the standard PyTorch ecosystem, both Google and Amazon are trying to loosen the grip that Nvidia has had on the AI software stack.

Performance metrics suggest the new setup is competitive. According to Inferact, a startup involved in building the new software, a tuned version of the vLLM engine can process over 700 tokens per second for a single user on Google’s seventh-generation Ironwood TPU. This test used 16 chips and a technique called speculative decoding. In comparison, a similar setup using 16 Nvidia GB200 chips reached 452 tokens per second. For smaller batch sizes, the new Google setup delivers roughly 1.4 to 2 times the throughput of the Nvidia baseline.

The partnership plans to expand further. SGLang-TorchTPU will begin supporting additional models, starting with DeepSeek V4.1 Flash and Reflection Beam. The teams are also working toward providing support for new open models on the day they are released. This goal is targeted for the first quarter of 2027. More details are expected to be shared at the SGLang Summit 2026, which is scheduled for November 12 and 13 in San Francisco.

While the exact location of Meta’s production TPUs is not publicly confirmed, industry reports from last year suggested a phased approach. These reports indicated that Meta discussed renting TPUs from Google Cloud in 2026, with direct purchases planned for 2027. It was also reported that Meta signed a multi-year rental agreement. This gradual integration aligns with the broader trend of major tech companies diversifying their hardware sources to reduce dependency on a single supplier.

This development marks a significant step in the evolution of AI infrastructure. By making custom chips more accessible through popular software frameworks, Google and Meta are paving the way for a more competitive market. For developers, this means more options; for consumers, it hints at a future where the AI tools we rely on are powered by a wider variety of high-performance hardware.