Google TPUs Break Into Meta's Production AI Stack

Google’s custom AI chips are now running production workloads at Meta, marking a significant shift in how major tech firms source their computing power. This development represents the first time Google’s Tensor Processing Units (TPUs) have been sold to and used by an external customer in this capacity. The integration is facilitated by a new PyTorch-native version of the SGLang inference engine, a move that bypasses the need for developers to rewrite their code for Google’s proprietary JAX framework.

The primary claim here is that Google has successfully loosened its hardware ecosystem to accommodate the industry-standard PyTorch framework, which is widely used by Meta and other organizations. Previously, running TPUs required a specific bridge through JAX. By releasing a native PyTorch version of SGLang, along with an open-source version of the vLLM inference engine, Google is making its hardware a direct, native option for the broader developer community. This reduces the friction of switching from Nvidia GPUs to Google TPUs.

Zongwei Zhou, the PyTorch-TPU lead at Meta, confirmed that the SGLang-TorchTPU system is currently live in Meta’s production environment. He described the result as a "remarkably mature, PyTorch-native serving stack," noting that it was the product of a three-way collaboration between Meta, Google, and the SGLang team. The SGLang team is based at RadixArk, a startup founded by the creators of the engine, which worked with Google and Meta to build this specific integration.

Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic importance of hardware diversity. She stated that Meta "cannot be constrained by a single hardware architecture" and that the new TorchTPU backend enables the company to deploy its existing PyTorch workloads directly onto Google TPUs. Google confirmed that it co-developed this TorchTPU backend in partnership with Meta’s PyTorch team, signaling a deeper collaborative relationship beyond simple vendor-client dynamics.

This is not the first time Meta has experimented with Google’s hardware, but it is the first time it has reached full production status with a native toolset. Google identifies PyTorch/XLA, an older bridge between PyTorch and TPUs, as the predecessor to the current system. Reports from last year indicated that Meta’s research teams used TPUs via this older bridge between 2020 and 2023 before canceling those contracts. The current push appears specifically designed to address Meta’s preference for staying within the PyTorch ecosystem rather than migrating to JAX.

While the exact physical location of Meta’s production TPUs remains unconfirmed, industry talks last year suggested that Meta would rent Google Cloud TPUs in 2026 before moving to purchases in 2027. Meta has reportedly signed a multi-year rental agreement, indicating a long-term commitment to this hardware strategy. This timeline aligns with the current announcement, suggesting the infrastructure is now fully operational.

The timing of this release is notable, as it came one day after Amazon announced the open-sourcing of TorchNeuron, its own native PyTorch backend for its Trainium2 and Trainium3 chips. Both announcements occurred at the PyTorch Conference and appear to be coordinated efforts to reduce the software lock-in associated with Nvidia’s GPUs. By providing native PyTorch support, both Google and Amazon are making it easier for developers to choose alternative hardware without significant code changes.

In terms of raw performance, the new setup shows competitive results. A tuned vLLM configuration serving Moonshot AI’s Kimi K3 model, which has 2.8 trillion parameters, reportedly exceeds 700 tokens per second for a single user on Google’s seventh-generation Ironwood TPU. This test, conducted by Inferact, a startup founded by vLLM’s creators, used 16 Ironwood chips with speculative decoding. In comparison, the same model on 16 Nvidia GB200 chips reached 452 tokens per second. At smaller batch sizes, the new kernels deliver roughly 1.4 to 2 times the throughput of the GB200 baseline.

The ecosystem is expected to expand rapidly. SGLang-TorchTPU will begin adding support for new models, starting with DeepSeek V4.1 Flash and Reflection Beam. RadixArk aims to provide Day 0 TPU support for new open models by the first quarter of 2027, meaning the hardware will be ready to run these models on their release date. Google, Meta, and RadixArk plan to share further details at the SGLang Summit 2026, scheduled for November 12-13 in San Francisco.