Meta Puts Google TPU Chips to Work in Production AI Systems

Google’s custom AI chips are now running production workloads at Meta, marking the first time the company has sold its tensor processing units to an external customer. This development breaks a long standing pattern of Google using its TPUs exclusively for internal services. The move is significant because it introduces a direct alternative to Nvidia’s dominant position in the AI hardware market.

The change is enabled by a new version of the SGLang inference engine that is native to PyTorch. Previously, running code on Google TPUs required using the JAX framework. The new software stack allows developers to use PyTorch directly on the hardware. This removes a major barrier for companies that have built their AI models on the PyTorch ecosystem.

Zongwei Zhou, the PyTorch TPU lead at Meta, confirmed that the system is live. He noted that the engine provides a remarkably mature serving stack. The software was built through a close collaboration between Meta, Google, and the SGLang team. This group operates out of RadixArk, a startup founded by the creators of SGLang.

Meta has long sought to avoid being tied to a single hardware provider. Jana van Greunen, director of PyTorch engineering at Meta, stated that the company cannot be constrained by one architecture. The new TorchTPU backend allows Meta to deploy its existing PyTorch workloads onto Google hardware. Google co developed this software with Meta’s engineering team to facilitate this transition.

This is not the first time Meta has experimented with Google chips. Research teams at Meta used TPUs between 2020 and 2023 through an older bridge called PyTorch/XLA. Those contracts were eventually canceled. The new push is more robust because it integrates deeply with the PyTorch framework that Meta relies on for its core AI development.

Reports from last year suggested that Meta was in talks to rent Google Cloud TPUs in 2026. The company reportedly signed a multi year rental agreement for purchases in 2027. While the exact location of these production chips is not public, the announcement confirms that the hardware is now serving live AI traffic.

The timing of this announcement is notable. It came one day after Amazon open sourced TorchNeuron. That is a native PyTorch backend for its own Trainium chips. Both Google and Amazon are working to loosen the software lock in that has favored Nvidia. By making their hardware easier to use with popular frameworks, they aim to attract more enterprise customers.

Performance metrics for the new setup are promising. A tuned version of the vLLM engine achieved over 700 tokens per second for a single user. This test used a large open weight model called Kimi K3 on seventh generation Ironwood TPUs. The setup used 16 chips with speculative decoding to boost speed.

Inferact, the startup behind vLLM, helped build this version. It compared the results against Nvidia GB200 chips, which reached 452 tokens per second in a similar test. The new kernels also show a throughput advantage of roughly 1.4 to 2 times the Nvidia baseline at small batch sizes. This indicates that the Google hardware can compete on speed as well as software compatibility.

Development on the software stack will continue through the end of the year. The team plans to add support for new models including DeepSeek V4.1 Flash and Reflection Beam. RadixArk aims to provide day zero support for new open models by the first quarter of 2027. More details on this partnership will be shared at the SGLang Summit in San Francisco in November.