Meta Deploys Google AI Chips

Meta Platforms is now running Google's Tensor Processing Units (TPUs) in production environments, marking the first time the hardware has been sold to a major external customer for large-scale AI workloads. This move, announced via a new PyTorch-native version of the SGLang inference engine, represents a significant challenge to Nvidia's dominant position in the AI chip market.

The transition relies on a new open-source version of the SGLang inference engine, known as SGLang-TorchTPU. Previously, running these engines required Google's proprietary JAX framework, but the new version allows the hardware to function natively with PyTorch. Zongwei Zhou, Meta's PyTorch-TPU lead at startup RadixArk, confirmed that the system is live, describing the result as a "remarkably mature, PyTorch-native serving stack" developed through close collaboration between Meta, Google, and the SGLang team.

Meta’s engineering leadership emphasized the strategic necessity of this shift. Jana van Greunen, director of PyTorch engineering at Meta, stated that the company "cannot be constrained by a single hardware architecture." By adopting TorchTPU, Meta can deploy its PyTorch workloads directly to Google's hardware. This initiative was co-developed by Google and Meta’s PyTorch team, aiming to remove the need for code migration to JAX, a primary barrier in previous attempts to use TPUs.

While this is not Meta's first foray into TPU technology, the current deployment is distinct. Google identifies the older PyTorch/XLA bridge as a "predecessor" to the new TorchTPU standard. Reports from last year indicated that Meta research teams used TPUs via this older bridge from 2020 to 2023 before canceling contracts. The current push is specifically tailored to Meta's preference for PyTorch, and while the physical location of the production TPUs is not specified, industry reports suggest talks involved Cloud rentals in 2026 with hardware purchases planned for 2027.

The announcement arrives on the heels of a similar move by Amazon, which open-sourced TorchNeuron for its Trainium chips at the same PyTorch Conference. These parallel efforts appear designed to erode Nvidia's software lock-in by providing native PyTorch support for competing hardware. Performance benchmarks suggest the new infrastructure is competitive; a tuned vLLM setup on the seventh-generation Ironwood TPU reportedly exceeded 700 tokens per second for a single user, outpacing the 452 tokens per second achieved by Nvidia's GB200 in comparable tests.

Looking ahead, the ecosystem is poised for rapid expansion. SGLang-TorchTPU will soon support additional models, including DeepSeek V4.1 Flash and Reflection Beam. RadixArk aims to achieve "Day 0" support for new open models by the first quarter of 2027, ensuring immediate compatibility upon release. Further details on this collaboration are expected to be shared at the SGLang Summit 2026, scheduled for November 12–13 in San Francisco.