Meta's Production Use of Google TPUs Signals a Shift in AI Chip Dynamics

Google's custom AI chips are now running in production at Meta, marking the first time the company's tensor processing units have been sold to an external customer for large-scale AI workloads. This development, announced on Google's developer blog, represents a significant shift in the competitive landscape for artificial intelligence hardware. For years, Nvidia has dominated this space, but Meta's move to integrate Google's silicon into its core infrastructure suggests that the monopoly may finally be facing a credible technical challenge.

The shift hinges on a new version of the SGLang inference engine, which now supports PyTorch natively. Previously, running AI models on Google's TPUs required using the company's proprietary JAX framework. That requirement created a significant barrier for teams like Meta, which are heavily invested in PyTorch. The new SGLang-TorchTPU engine removes this friction, allowing developers to use their existing PyTorch code with minimal rewriting. A matching open-source version of vLLM was also released, further expanding the software options available for this hardware.

Meta has been using Google's technology in a limited capacity for several years, but this marks the first time it is described as being in full production use. The transition was the result of a close collaboration between Meta, Google, and SGLang's creators. Zongwei Zhou, Meta's PyTorch-TPU lead at the startup RadixArk, confirmed that the new engine is running in production today. He described the tool as a remarkably mature serving stack that fits seamlessly into Meta's existing workflow.

This is not the first time Meta has experimented with Google's chips. From 2020 to 2023, Meta's research teams used TPUs via an older bridge called PyTorch/XLA. However, those contracts were eventually canceled. Jana van Greunen, director of PyTorch engineering at Meta, explained that the company cannot be constrained by a single hardware architecture. She noted that the new TorchTPU backend enables Meta to deploy its PyTorch workloads directly to Google TPUs, providing the flexibility the company requires.

The timing of this announcement is notable. It came just one day after Amazon announced it would open-source TorchNeuron, a similar backend for its own Trainium chips. Both moves appear designed to loosen Nvidia's grip on the software stack. By making it easier to run standard PyTorch models on non-Nvidia hardware, these companies are reducing the switching costs for large AI labs.

Performance benchmarks suggest the new setup is competitive with Nvidia's offerings. According to Inferact, a startup that helped build the new tools, a tuned vLLM setup on the seventh-generation Ironwood TPU can achieve over 700 tokens per second for a single user. This figure compares favorably with Nvidia's GB200 chips, which reached 452 tokens per second in similar tests. The new kernels also show significant throughput improvements at small batch sizes for certain open-weight models.

While the specific location of Meta's production TPUs remains unclear, previous reports indicated that Meta and Google had discussions about cloud rentals in 2026 and potential purchases in 2027. Meta is reportedly part of a multi-year rental agreement. The companies plan to share more details at the SGLang Summit in San Francisco in November.

Looking ahead, SGLang-TorchTPU is expected to add support for more models, starting with DeepSeek V4.1 Flash. The team aims to provide day-zero support for new open models by the first quarter of 2027. If this timeline holds, Google's TPUs could become a standard option for AI developers, forcing Nvidia to compete not just on raw hardware speed, but on the entire software ecosystem.

The integration of Google's hardware into Meta's production environment is a clear signal that the AI chip market is becoming less monolithic. By lowering the software barriers to entry, Google and its partners are positioning their chips as a viable alternative to the industry standard. Whether this leads to a lasting shift in market share will depend on how quickly other major AI labs follow suit.