Meta Deploys Google AI Chips in Production, Breaking Nvidia's Grip

Google’s custom AI chips are now running in production at Meta. The two tech giants confirmed this development through the release of new, PyTorch-native versions of the SGLang inference engine. This move marks a significant shift in how major tech companies handle their artificial intelligence infrastructure. It signals that Google’s Tensor Processing Units, or TPUs, are no longer just for Google’s internal use. They are now a viable option for external customers who rely on the PyTorch framework.

The software update allows these chips to work natively with PyTorch without requiring extensive code rewrites. Previously, running workloads on TPUs required using Google’s JAX AI framework. The new SGLang-TorchTPU and vLLM-TorchTPU versions change that dynamic. They make the hardware a direct, native option for PyTorch developers. This reduces the friction for companies that have built their AI stacks on PyTorch but wanted to diversify their hardware options.

Zongwei Zhou, Meta’s lead for PyTorch-TPU, stated that the new system is already live. He described the effort as a close collaboration between Meta, Google, and the SGLang team. Zhou noted that the setup provides a remarkably mature serving stack. His team operates out of RadixArk, a startup founded by the creators of SGLang. RadixArk worked with both Google and Meta to build this specific engine integration.

Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic importance of this move. She stated that Meta cannot be constrained by a single hardware architecture. According to Greunen, the new TorchTPU software enables the company to deploy its PyTorch workloads directly to Google’s hardware. Google confirmed it co-developed this bridge with Meta’s PyTorch team. The goal was to create a seamless connection for Meta, which had previously resisted migrating to Google’s JAX ecosystem.

This is not the first time Meta has experimented with Google’s chips. An older bridge called PyTorch/XLA existed before this new release. Reports indicate that Meta research teams used TPUs through that older bridge from 2020 to 2023. However, Meta eventually canceled those contracts. The current push appears more permanent. Last year, talks reportedly included renting TPU capacity from Google Cloud in 2026 and purchasing hardware in 2027. Meta later signed a multi-year rental agreement, suggesting a long-term commitment to this infrastructure.

The announcement landed with particular timing. It came 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 favored Nvidia. By making it easier to run PyTorch on non-Nvidia hardware, Google and Amazon are challenging the dominant position of Nvidia’s GPUs in the AI market. This could reshape the competitive landscape for AI compute resources.

Performance metrics for the new setup show strong results. A tuned version of vLLM serving the Kimi K3 model achieved over 700 tokens per second on a single user. This test used seventh-generation Ironwood TPUs. The figure comes from Inferact, a startup founded by the creators of vLLM. Inferact helped Google build the new vLLM-TorchTPU integration. The test utilized 16 Ironwood chips with speculative decoding to speed up inference.

In comparison, a similar setup using 16 Nvidia GB200 chips reached 452 tokens per second. The new TPU kernels for Kimi K3 and Qwen 3.8 27B also showed higher throughput at small batch sizes. They delivered roughly 1.4 to 2 times the performance of the GB200 baseline. These numbers suggest that Google’s hardware is becoming a serious contender for high-performance AI inference tasks, particularly for large language models.

Development work on the new software continues. SGLang-TorchTPU plans to add support for more models, starting with DeepSeek V4.1 Flash and Reflection Beam. RadixArk is also working toward Day 0 support for new open models. This means the software will support new model releases on the day they launch. The company aims to achieve this by the first quarter of 2027. Google, Meta, and RadixArk plan to share more details at the SGLang Summit 2026, which will take place in San Francisco from November 12 to 13.

The stakes for the AI industry are high. If Meta and other major players continue to adopt Google’s TPUs, it could disrupt the status quo. Nvidia has long dominated the market due to its strong software ecosystem. By making PyTorch native to TPUs, Google is removing a key barrier to entry. This move could lead to a more competitive market for AI hardware. It also gives companies like Meta more leverage when negotiating contracts with chipmakers. The integration of these systems represents a tangible step away from reliance on a single vendor.