Google TPUs Launch Production Workloads at Meta

Google's custom AI accelerators, known as Tensor Processing Units (TPUs), are now powering live AI workloads inside Meta's data centers. The deployment runs through a new version of the SGLang inference engine built specifically for PyTorch, the software framework Meta uses to develop its artificial intelligence models. This marks the first time Google's TPUs are being sold to and utilized by external customers for production tasks. The move directly challenges Nvidia's stronghold on the AI hardware market by offering a viable alternative for large-scale model deployment.

The rollout centers on a new software stack developed jointly by Google and Meta. The new SGLang-TorchTPU engine, along with a matching version of the vLLM engine, allows Meta to run its AI models on Google's hardware without needing to rewrite its code or abandon its existing development tools. Previously, running Google's TPUs required using Google's own JAX framework. By enabling native support for PyTorch, Google has removed a major technical barrier for companies like Meta that are built entirely on PyTorch. This partnership highlights a concerted effort to make Google's silicon a direct competitor to Nvidia's graphics processing units (GPUs), which currently dominate the industry for running AI models.

The integration into Meta's infrastructure is already active. Zongwei Zhou, Meta's PyTorch-TPU lead, confirmed that the new engine is running in production at Meta today. He described the resulting setup as a remarkably mature, PyTorch-native serving stack. Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic importance of this flexibility. She stated that Meta cannot be constrained by a single hardware architecture and that the new software enables the company to deploy its PyTorch workloads to Google TPUs. This development follows years of previous attempts to bridge PyTorch and TPUs, which Meta abandoned between 2020 and 2023. Google has now co-developed this new solution directly with Meta's team to address those past failures.

Performance data released alongside the announcement suggests Google's hardware can compete on speed. Tests conducted by Inferact, the startup behind vLLM's creation, compared the two architectures directly. Using 16 Google Ironwood TPUs to serve the open-weight Kimi K3 model, the system achieved speeds over 700 tokens per second for a single user. In comparison, a configuration using 16 Nvidia GB200 GPUs reached 452 tokens per second for the same task. The tests also showed that for certain models, the Google hardware delivered roughly 1.4 to 2 times the throughput of the Nvidia baseline at smaller batch sizes. These figures provide concrete evidence that Google's TPUs can outperform Nvidia's high-end chips in specific inference scenarios.

This announcement arrives amid increasing competition among major cloud providers to reduce reliance on Nvidia. Just one day prior to Google's post, Amazon released open-source software for its own Trainium chips at the same PyTorch Conference. Both companies are pushing to loosen Nvidia's software lock-in by offering native support for their respective hardware within the PyTorch framework. Google, Meta, and RadixArk plan to share further details regarding this collaboration at the SGLang Summit in San Francisco, scheduled for November 12 and 13, 2026. The expansion of Google's TPU support to additional models, including DeepSeek V4.1 Flash, is planned for the first quarter of 2027.