Meta Runs Google AI Chips in Production, Challenging Nvidia's Grip on the Market

The long-standing alliance between Meta and Nvidia is facing its most direct challenge yet. Google has confirmed that its custom Tensor Processing Units are now running AI workloads in production at Meta. This move marks the first time Google has sold these chips to an external customer for such a critical role, signaling a potential crack in Nvidia's dominance of the AI hardware market.

The breakthrough is driven by software. Google announced a new, PyTorch-native version of the SGLang inference engine. Previously, these engines relied on Google's own JAX framework to function on TPUs. The new versions allow the hardware to operate natively with PyTorch, the standard framework for much of the AI world, without requiring extensive code rewrites. A matching open-source version of vLLM was also announced.

Meta’s involvement is direct and hands-on. Zongwei Zhou, Meta’s PyTorch-TPU lead, stated that SGLang-TorchTPU is running in production at Meta today. He noted this result came from a close three-way collaboration between Meta, Google, and the SGLang team. Zhou said the setup provides Meta with a remarkably mature, PyTorch-native serving stack.

Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic importance of this shift. She explained that Meta cannot be constrained by a single hardware architecture. According to Greunen, TorchTPU enables Meta to deploy its PyTorch workloads to Google TPUs. Google confirmed it co-developed TorchTPU with Meta’s PyTorch team to make this transition seamless.

This is not Meta’s first encounter with Google’s silicon. An older bridge called PyTorch/XLA existed, but Google now considers it a predecessor to the new system. Reports indicate that Meta research teams used TPUs via this older bridge from 2020 to 2023 before canceling those contracts. The new push appears designed to address Meta’s previous reluctance to migrate to JAX.

While the exact location of Meta’s production TPUs is not public, recent reports suggest significant activity. Talks last year included Google Cloud TPU rentals for 2026, with purchases expected in 2027. Meta reportedly signed a multi-year rental agreement, suggesting a long-term commitment to the technology.

The timing is notable. The announcement came one day after Amazon said it was open-sourcing TorchNeuron, its own native PyTorch backend for its Trainium chips. Both moves appear aimed at loosening Nvidia’s software lock-in. By making their hardware more compatible with the industry-standard PyTorch, Google and Amazon are lowering the barrier for other companies to switch.

Performance metrics suggest the new setup is competitive. A tuned vLLM setup serving the Kimi K3 model reportedly tops 700 tokens per second for a single user on the seventh-generation Ironwood TPU. This speed comes from Inferact, a startup that helped build the engine. Their test used 16 Ironwood chips with speculative decoding. In comparison, a similar setup on 16 Nvidia GB200s reached 452 tokens per second.

For smaller workloads, the advantage is also evident. Kernels for specific models deliver roughly 1.4 to 2 times the throughput of the GB200 baseline without speculative decoding. SGLang-TorchTPU will continue adding models, starting with DeepSeek V4.1 Flash and Reflection Beam. The teams are working toward Day 0 support for new open models by the first quarter of 2027.

More details are expected at the SGLang Summit 2026, scheduled for November 12 and 13 in San Francisco. Google, Meta, and RadixArk plan to share further insights then. For now, the industry is watching closely as the first major external customer begins running Google’s chips in a high-stakes production environment.