Meta Puts Google TPUs to Work in PyTorch, Challenging Nvidia's Grip

The dominance of Nvidia’s graphics cards in artificial intelligence has long seemed unassailable, yet a new development suggests that barrier is beginning to crumble as competitors find viable alternatives. On Wednesday, Google announced that its custom Tensor Processing Units (TPUs) are now operating in production environments at Meta, the developer behind the ubiquitous PyTorch framework. This milestone was achieved through the release of a PyTorch-native version of the SGLang inference engine, signaling a move away from the proprietary software stacks that have historically kept Google’s hardware siloed from the broader AI community.

The technical shift is significant because it removes the need for developers to rewrite code for Google’s JAX framework. Previously, running workloads on TPUs required a translation layer, but the new SGLang-TorchTPU and an open-source version of vLLM now treat PyTorch as a native environment. According to Inferact, a startup co-founded by the creators of vLLM, a tuned setup for Moonshot AI’s Kimi K3 model achieved over 700 tokens per second on seventh-generation Ironwood TPUs. This figure notably outpaces the 452 tokens per second recorded on Nvidia’s GB200 chips under similar conditions, a performance gap that raises questions about the efficiency of Nvidia’s current architecture for large-scale inference.

The collaboration behind this move involves a complex web of partnerships. Zongwei Zhou, Meta’s PyTorch-TPU lead at RadixArk, stated that the system is "running in production at Meta today," a result of a "close three-way collaboration" between Meta, Google, and the SGLang team. Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic importance of the update, noting that the company "cannot be constrained by a single hardware architecture." By enabling direct deployment of PyTorch workloads, Google has effectively addressed a long-standing friction point that previously prevented Meta from fully adopting TPUs.

This is not Meta’s first foray into Google’s hardware, though it is its first at scale with a native interface. SemiAnalysis reported last year that Meta’s research teams used TPUs via the older PyTorch/XLA bridge between 2020 and 2023 before canceling those contracts. The new TorchTPU project, co-developed with Meta, appears to be a more robust successor to that earlier experiment. While the exact location of these production TPUs remains undisclosed, industry reports suggest Meta signed a multi-year rental agreement with Google Cloud, with initial rentals potentially starting in 2026 and purchases following in 2027.

The timing of the announcement is hardly coincidental. It landed just one day after Amazon revealed it was open-sourcing TorchNeuron, its own native PyTorch backend for Trainium chips. Both moves appear designed to loosen Nvidia’s software lock-in, which has historically made it difficult for other hardware providers to compete in the AI market. For Meta and other large enterprises, the ability to run their preferred software stack on non-Nvidia hardware reduces dependency on a single supplier and may offer better cost-performance ratios for specific workloads.

The stakes for the industry are high as this ecosystem matures. RadixArk plans to add support for DeepSeek V4.1 Flash and Reflection Beam, with a goal of providing Day 0 TPU support for new open models by the first quarter of 2027. Further details on the collaboration are expected at the SGLang Summit in San Francisco in November. If these early performance benchmarks hold true, the coming year could see a more fragmented and competitive AI hardware market, forcing Nvidia to justify its premium pricing with superior software integration and performance.