Google AI Chips Now Run Major Social Media Workloads

For businesses relying on artificial intelligence, a new hardware option is now available for production use. Google has confirmed that its custom AI chips, known as tensor processing units or TPUs, are currently running live workloads for Meta. This development means that one of the world's largest technology companies is no longer limited to a single type of processor for its AI tasks, providing a practical alternative to the dominant industry standard.

The change is driven by new software versions of the SGLang and vLLM inference engines. Previously, these tools required Google’s proprietary JAX framework to function on TPUs. The updated releases allow these engines to run natively on PyTorch, the popular AI framework developed by Meta, without requiring major code rewrites. Zongwei Zhou, Meta’s PyTorch-TPU lead, stated that the new system is "running in production at Meta today" following a close collaboration between Meta, Google, and the SGLang team.

Meta emphasized that it "cannot be constrained by a single hardware architecture," according to Jana van Greunen, director of PyTorch engineering at Meta. The company noted that this new setup enables the deployment of PyTorch workloads directly onto Google TPUs. While Meta previously used an older bridge called PyTorch/XLA from 2020 to 2023, that arrangement ended when Meta canceled those contracts. The new TorchTPU solution was co-developed specifically to address Meta's needs without forcing a migration to JAX.

The move appears to be part of a broader strategy to reduce reliance on Nvidia, the current market leader in AI chips. This announcement came just one day after Amazon revealed it was open-sourcing its own PyTorch backend for its Trainium chips at the same industry conference. Both initiatives aim to loosen the software lock-in that has long favored Nvidia's ecosystem, giving other hardware providers a clearer path to compete in the enterprise market.

Performance data suggests the new setup is competitive. Tests by Inferact, a startup involved in building the vLLM integration, showed that a tuned setup serving the Kimi K3 model exceeded 700 tokens per second for a single user on the seventh-generation Ironwood TPU. This figure surpasses the 452 tokens per second achieved by Nvidia GB200 chips in comparable published setups. For smaller batch sizes, the new kernels also delivered roughly 1.4 to two times the throughput of the Nvidia baseline.

Looking ahead, the SGLang-TorchTPU project plans to add support for new models, starting with DeepSeek V4.1 Flash and Reflection Beam. RadixArk, the startup behind the engine, aims to provide day-one 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 on November 12 and 13.

This shift indicates that large tech firms are actively diversifying their hardware portfolios. By making Google's TPUs a viable native option for PyTorch, the industry sees a more balanced landscape where software compatibility no longer dictates hardware choice as strictly as before.