Google's AI Chips Enter Meta's Data Centers

Google's custom-built tensor processing units are now handling production AI workloads for Meta, marking the first time these specialized chips are being sold to outside customers. This move directly challenges Nvidia's widespread control over the market for artificial intelligence hardware. The integration is made possible through a new PyTorch-native version of the SGLang inference engine, which was announced on Google's developer blog. A corresponding open-source version of the vLLM inference engine was also released.

Previously, running these tools on Google's hardware required using Google's own JAX AI framework. The new software, however, is designed to allow Meta to use its existing PyTorch codebase to run on Google's chips with little to no modification. Meta's PyTorch-TPU lead, Zongwei Zhou, described the collaboration as a 'close three-way' effort between Meta, Google, and the SGLang team. He stated that SGLang-TorchTPU is 'running in production at Meta today.' The engine provides Meta with a 'remarkably mature, PyTorch-native serving stack.'

Jana van Greunen, Meta's director of PyTorch engineering, emphasized the strategic importance of this flexibility. In a post on Google's blog, she stated that Meta 'cannot be constrained by a single hardware architecture' and that the new TorchTPU 'enables us to deploy our PyTorch workloads to Google TPUs.' This development represents a renewed effort by Google to capture a piece of Meta's business. Meta research teams did use TPUs from 2020 to 2023 through an older bridge technology before those contracts were canceled. Google's push for PyTorch compatibility has been aimed specifically at Meta, which has been unwilling to migrate to the JAX framework.

This announcement is part of a broader trend of chipmakers attempting to break Nvidia's software lock-in. It came just one day after Amazon announced it was open-sourcing TorchNeuron, its own native PyTorch backend for its Trainium chips, at the same PyTorch Conference. Regarding performance, a tuned vLLM setup on Google's seventh-generation Ironwood TPU reportedly achieved speeds exceeding 700 tokens per second for a single user. That figure was compared against a similar setup using Nvidia GB200 chips, which reached 452 tokens per second in a test run by Inferact, the startup that helped build the vLLM-TorchTPU engine. The SGLang project plans to add support for more models, including DeepSeek V4.1 Flash, and aims to provide day-one support for new open models by the first quarter of 2027. Further details are planned for the SGLang Summit 2026, scheduled for November 12-13 in San Francisco.

While the exact location of Meta's production TPUs is not yet clear, discussions from last year reportedly involved renting Google Cloud TPUs in 2026 with potential purchases in 2027. Meta has also reportedly signed a multi-year agreement for rentals. The company's use of Google's hardware is intended to provide an alternative for running its AI models.