Meta's AI Stack Runs on Google Chips, Shifting the Balance in Data Centers

For the first time, the massive AI infrastructure powering Meta's services is running in production on chips that were once strictly internal to Google. This shift means that the company behind Facebook and Instagram can now execute its core artificial intelligence tasks using Google's tensor processing units. It marks a significant departure from the status quo where Nvidia's hardware and software ecosystem held a near-monopoly over large scale model training and inference.

The change is driven by the release of new, PyTorch native versions of two major inference engines, SGLang and vLLM. Previously, these tools required developers to use Google's own JAX framework to operate on TPUs. The new software stacks allow the hardware to function natively within the PyTorch environment. This update largely removes the need for extensive code rewriting. Zongwei Zhou, who leads Meta's PyTorch TPU efforts, confirmed that SGLang TorchTPU is live in production. He described the result as a remarkably mature serving stack that emerged from a close three way collaboration between Meta, Google, and the SGLang team.

Meta officials emphasized that the company should not be limited by a single hardware architecture. Jana van Greunen, director of PyTorch engineering at Meta, stated that this new toolset enables the deployment of PyTorch workloads directly onto Google TPUs. While this is not Meta's first interaction with the hardware, it represents a more integrated approach. Older bridges like PyTorch XLA were used by research teams from 2020 to 2023 before contracts were canceled. The current push appears specifically designed to accommodate Meta's refusal to migrate entirely to JAX.

The physical location of these production units remains a question. Reports from last year suggested that negotiations included rentals through Google Cloud in 2026, with potential purchases in 2027. Meta was later reported to have signed a multi year rental agreement. The timing of this announcement is notable because it arrived just one day after Amazon revealed it was open sourcing its own native PyTorch backend for its Trainium chips. Both moves appear calculated to loosen the software lock in that has benefited Nvidia for years.

Performance metrics suggest the new setup is competitive. A tuned vLLM configuration serving the open weight Kimi K3 model reportedly achieves over 700 tokens per second for a single user on the seventh generation Ironwood TPU. This figure comes from Inferact, a startup that helped build the software. Their tests used 16 Ironwood chips with speculative decoding, a method that speeds up inference. This outperformed the 452 tokens per second recorded on 16 Nvidia GB200 chips. Without that specific acceleration method, the new kernels still deliver roughly 1.4 to 2 times the throughput of the Nvidia baseline at small batch sizes.

The ecosystem around this new compatibility is expanding quickly. SGLang TorchTPU plans to add support for models like DeepSeek V4.1 Flash and Reflection Beam. RadixArk, the startup behind the engine, aims to offer day zero TPU support for new open models by the first quarter of 2027. Google, Meta, and RadixArk are scheduled to share further details at the SGLang Summit in San Francisco in November 2026. For readers in the Southeast, this signals a more diverse and potentially cost effective landscape for enterprise AI deployment in the coming years.