Meta Puts Google's AI Chips to Work, Challenging Nvidia's Grip

Google has officially confirmed that its custom tensor processing units are now powering AI workloads inside Meta's data centers. This milestone was reached through the release of a new, PyTorch-native version of the SGLang inference engine. Previously, using Google’s hardware required developers to rely on the company's proprietary JAX framework. The new software stack removes that barrier, allowing Meta to deploy its models on TPUs without extensive code rewrites. This move represents the first time Google has sold its specialized chips to an external customer for production-level AI tasks.

The transition is the result of a close collaboration between Meta, Google, and RadixArk, a startup founded by the creators of SGLang. Zongwei Zhou, who leads Meta’s PyTorch-TPU efforts at RadixArk, noted that the new setup provides a "remarkably mature, PyTorch-native serving stack." Jana van Greunen, a director at Meta, emphasized that the company cannot rely on a single hardware architecture. She stated that the new TorchTPU tool enables Meta to deploy its existing PyTorch workloads directly onto Google’s hardware, offering greater flexibility in its infrastructure strategy.

This is not the first time the two companies have worked together. Google describes an older bridge called PyTorch/XLA as a predecessor to the new system. Reports from last year indicated that Meta research teams utilized TPUs through this earlier bridge between 2020 and 2023 before canceling those specific contracts. The recent push appears to be a more permanent solution, with talks reportedly involving Google Cloud rentals for 2026 and potential purchases for 2027. Meta is believed to have signed a multi-year agreement, although the exact physical location of these production chips remains unconfirmed.

The announcement landed just one day after Amazon revealed it was open-sourcing its own native PyTorch backend for its Trainium chips. Both companies appear to be making concerted efforts to loosen the software lock-in that has historically favored Nvidia. By ensuring their hardware works seamlessly with the industry-standard PyTorch framework, both Google and Amazon are making their alternatives more attractive to large-scale AI developers who are tired of being tied to a single vendor's ecosystem.

Performance benchmarks suggest the new setup is competitive. Inferact, a startup involved in building the vLLM-TorchTPU integration, reported that a tuned system serving the 2.8-trillion-parameter Kimi K3 model achieved over 700 tokens per second for a single user on seventh-generation Ironwood TPUs. This test used 16 chips with speculative decoding. In comparison, a similar setup on 16 Nvidia GB200s reached 452 tokens per second. Without speculative decoding, the new kernels reportedly deliver 1.4 to 2 times the throughput of the Nvidia baseline at small batch sizes.

Looking ahead, the teams plan to expand model support, starting with DeepSeek V4.1 Flash and Reflection Beam. RadixArk aims to provide same-day support for new open models by the first quarter of 2027. Google, Meta, and RadixArk will share further details at the SGLang Summit, scheduled for November 12-13 in San Francisco. As these partnerships deepen, the landscape for AI infrastructure is becoming increasingly multi-vendor, offering users more choice than at any point in the past decade.