Silicon Rivalry: Google's Custom AI Chips Enter Meta's Production Infrastructure

More than 700 tokens per second. That is the reported speed achieved by a single user when running a massive 2.8-trillion-parameter AI model on Google’s seventh-generation Ironwood tensor processing units, according to recent benchmarks. This performance figure has drawn significant attention because it demonstrates that Google’s custom silicon can rival, and in some metrics exceed, the industry-standard Nvidia chips that have long dominated the artificial intelligence market.

The milestone was announced via Google’s developer blog, which confirmed that Meta is now using these TPUs in production environments. The breakthrough is made possible by a new, PyTorch-native version of the SGLang inference engine. Previously, running code on Google’s hardware required using the company’s own JAX framework, which created a barrier for developers accustomed to PyTorch, the most widely used AI framework. The new software allows Meta to run its workloads on Google hardware without extensive code rewriting.

Zongwei Zhou, the PyTorch-TPU lead at Meta, stated that the new engine is “running in production at Meta today.” He credited a “close three-way collaboration” between Meta, Google, and the SGLang team for this achievement. Zhou noted that his team operates out of RadixArk, a startup founded by the creators of SGLang, which developed the engine in partnership with the other two tech giants. He described the result as a “remarkably mature, PyTorch-native serving stack” for Meta.

Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic importance of this move. She explained that Meta “cannot be constrained by a single hardware architecture” and that the new TorchTPU tool “enables us to deploy our PyTorch workloads to Google TPUs.” Google confirmed that it co-developed this specific integration with Meta’s PyTorch team to ensure compatibility.

This is not the first time Meta has explored Google’s hardware. The company previously used an older bridge called PyTorch/XLA between 2020 and 2023, according to reports from SemiAnalysis, before canceling those contracts. However, the current push is more integrated. While the exact location of Meta’s production TPUs is not public, reports from last year indicated that Google Cloud TPU rentals were expected in 2026, with hardware purchases potentially following in 2027. Meta is also reported to have signed a multi-year rental agreement to secure this capacity.

The timing of the announcement is notable. It came just one day after Amazon revealed it was open-sourcing TorchNeuron, a similar native PyTorch backend for its own Trainium2 and Trainium3 chips. Both announcements were made around the same PyTorch Conference and appear to be coordinated efforts to reduce the software lock-in that has favored Nvidia in the AI hardware market.

In terms of raw performance, the new setup shows promise. A tuned vLLM configuration, which was also announced as an open-source option for TPUs, reportedly outperformed Nvidia’s GB200 chips in specific tests. The vLLM setup on 16 Ironwood chips, using a technique called speculative decoding, beat the GB200’s 452 tokens per second. Additionally, the new kernels for specific models deliver roughly 1.4 to 2 times the throughput of the Nvidia baseline at small batch sizes.

Looking ahead, the SGLang-TorchTPU project plans to support more models, starting with DeepSeek V4.1 Flash and Reflection Beam. RadixArk aims to provide “Day 0” support for new open models by the first quarter of 2027, meaning the software will be ready when the models are released. Google, Meta, and RadixArk are scheduled to share further details at the SGLang Summit in San Francisco on November 12-13.

This development marks a significant step for Google, which has historically kept its TPUs for internal use. By making them a native option for external customers like Meta through standard software frameworks, Google is directly challenging Nvidia’s position as the default choice for large-scale AI deployment.