Meta Deploys Google AI Chips in Production, Shaking Up Market Dynamics

Meta is now running Google’s custom artificial intelligence silicon in live production environments, marking the first time Google has sold its Tensor Processing Units (TPUs) to an outside customer. This milestone shifts the balance of power in the AI hardware market, challenging the long-standing dominance of Nvidia. The integration is powered by a new, PyTorch-native version of the SGLang inference engine, which allows Meta to utilize the chips without rewriting its existing codebase. For the first time, both SGLang and vLLM, two major inference engines, support this hardware natively through the standard PyTorch framework, removing the previous requirement to use Google’s proprietary JAX framework.

The path to this collaboration was neither immediate nor straightforward. Between 2020 and 2023, Meta’s research teams experimented with TPUs using an older software bridge known as PyTorch/XLA, but those contracts were eventually canceled. Google then focused its efforts on creating a more seamless connection, recognizing that Meta preferred to stick with PyTorch rather than migrate to a different framework. This led to the development of TorchTPU, a co-developed tool that serves as a direct successor to the older bridge. The current deployment represents a renewed commitment by Meta to diversify its hardware sources, ensuring it is not tied to a single architecture.

The technical implementation involves a close three-way partnership between Meta, Google, and RadixArk, a startup founded by the creators of SGLang. Zongwei Zhou, Meta’s lead for PyTorch-TPU integration, confirmed that the SGLang-TorchTPU system is active in production today. He described the result as a remarkably mature serving stack. Jana van Greunen, who directs PyTorch engineering at Meta, emphasized that the company cannot be constrained by one hardware vendor. This flexibility allows Meta to deploy its workloads directly onto Google’s infrastructure, a capability that was previously limited to internal Google use.

The timing of this announcement is particularly significant, as it came just one day after Amazon open-sourced its own native PyTorch backend for its Trainium chips. Both moves appear designed to loosen the software lock-in that has historically favored Nvidia. Performance benchmarks suggest the new setup is competitive. A tuned vLLM configuration serving a massive 2.8-trillion-parameter model reportedly achieved over 700 tokens per second on the seventh-generation Ironwood TPU. In comparison, Nvidia’s GB200 chips reached 452 tokens per second in similar tests, indicating a significant speed advantage for the new Google hardware in certain configurations.

Looking ahead, the ecosystem is set to expand rapidly. RadixArk plans to add support for new models like DeepSeek V4.1 Flash and Reflection Beam to the SGLang-TorchTPU platform. The team is also targeting Day 0 support for new open models by the first quarter of 2027, ensuring that developers can use the latest AI models on TPUs immediately upon release. Google, Meta, and RadixArk plan to share further details about this collaboration at the SGLang Summit, scheduled for November 12–13 in San Francisco. This event will likely provide deeper insights into how these three organizations plan to continue shaping the future of AI infrastructure.

The location of Meta’s production TPUs remains somewhat ambiguous, though reports suggest a mix of rentals and purchases. Last year, discussions included Google Cloud rentals for 2026 and hardware purchases for 2027, followed by a reported multi-year rental agreement. This hybrid approach allows Meta to scale its AI capabilities without the upfront capital expenditure of buying all the chips outright. As the industry watches this development, the shift signals a broader trend toward multi-vendor strategies in AI, where companies seek to avoid dependency on a single supplier for both hardware and software.