Meta puts Google's AI chips to work in production

A significant shift in the artificial intelligence hardware market is underway as Meta, the creator of the PyTorch framework, begins using Google’s custom chips in live production environments. This move marks the first time Google’s tensor processing units, or TPUs, have been sold to an external customer for large-scale deployment, challenging the long-standing dominance of Nvidia in the AI sector. The integration is made possible by a new, PyTorch-native version of the SGLang inference engine, which was announced on Google’s developer blog. Simultaneously, an open-source version of vLLM, another popular inference engine, was released to support the same hardware.

The path to this point has been complicated by software incompatibilities. Historically, running AI models on Google’s TPUs required developers to use Google’s JAX framework. Meta, however, preferred to stick with PyTorch. For years, a tool called PyTorch/XLA served as a bridge, but it is now considered a predecessor to the new system. Meta research teams utilized TPUs via this older bridge between 2020 and 2023 before canceling those contracts. The current effort represents a more direct solution, co-developed by Google and Meta to avoid the need for major code rewrites.

Zongwei Zhou, Meta’s lead for PyTorch-TPU efforts, confirmed that the new SGLang-TorchTPU system is currently live. He described the achievement as the result of a tight three-way collaboration between Meta, Google, and the SGLang team. Zhou noted that this partnership provides Meta with a highly mature serving stack that works natively with PyTorch. His team operates out of RadixArk, a startup founded by the creators of SGLang, which played a key role in building the engine alongside the two tech giants.

Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic importance of the update. She stated that Meta cannot rely on a single hardware architecture to meet its growing demands. By adopting this new software layer, Meta can deploy its existing PyTorch workloads directly onto Google’s TPUs. This flexibility allows the company to diversify its hardware portfolio without forcing its engineers to learn a new programming ecosystem.

The timing of this announcement is notable, as it arrived just one day after Amazon revealed it was open-sourcing TorchNeuron, a similar backend for its own Trainium chips. Both moves appear designed to reduce the software lock-in that has favored Nvidia for years. While the exact location of Meta’s production TPUs is not public, reports from last year suggested that Meta was in talks to rent Google Cloud TPUs in 2026, with potential purchases planned for 2027. Meta has since reportedly signed a multi-year rental agreement.

Performance benchmarks suggest the new setup is competitive. Inferact, a startup that helped Google build the vLLM version of the new engine, reported that a tuned setup serving a massive 2.8-trillion-parameter model achieved over 700 tokens per second for a single user. This test used 16 seventh-generation Ironwood TPUs and a technique called speculative decoding. In comparison, a similar setup on 16 Nvidia GB200 chips reached 452 tokens per second. Without that speed-up technique, the TPU kernels still delivered roughly 1.4 to 2 times the throughput of the Nvidia baseline at small batch sizes.

Looking ahead, the development teams plan to expand model support. SGLang-TorchTPU will begin adding new models, starting with DeepSeek V4.1 Flash and Reflection Beam. RadixArk aims to provide day-one support for new open models by the first quarter of 2027. Google, Meta, and RadixArk are scheduled to share more details about these efforts at the SGLang Summit in San Francisco on November 12 and 13.

For California-based tech workers and investors, this development signals a more competitive landscape for AI infrastructure. The ability to run PyTorch models on non-Nvidia hardware could influence hiring trends and data center planning in the coming year. As these new software bridges mature, the barrier to entry for using alternative chips lowers, potentially shifting the balance of power in the AI supply chain.