Meta Deploys Google TPUs in Production, Challenging Nvidia's AI Chip Monopoly

The long-standing assumption that Meta would remain tethered to Nvidia's hardware ecosystem has been fractured. The tension at the heart of the current AI infrastructure market is no longer just about chip speed, but about software control. Google has announced that its custom Tensor Processing Units (TPUs) are now running AI workloads in production at Meta, a move that signals a significant pivot away from the single-vendor dependency that has defined the industry for years. This development marks the first time TPUs have been sold to external customers for this specific production role, directly threatening the market dominance that Nvidia has maintained through its CUDA software lock-in.

The technical breakthrough enabling this shift is the release of a PyTorch-native version of the SGLang inference engine. Previously, running these engines on TPUs required developers to use Google’s JAX framework, a language many large AI labs, including Meta, resisted adopting. The new implementation allows the hardware to operate natively with PyTorch, largely eliminating the need for extensive code rewrites. Zongwei Zhou, Meta’s PyTorch-TPU lead, confirmed that the SGLang-TorchTPU system is currently live in production at Meta, the result of a three-way collaboration between Meta, Google, and the SGLang team. This partnership provides Meta with what Zhou described as a remarkably mature, PyTorch-native serving stack, removing a major friction point in their deployment pipeline.

Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic necessity of this diversification. She noted that Meta cannot be constrained by a single hardware architecture, a statement that underscores the risk of relying exclusively on one supplier in a market as volatile as high-performance computing. By enabling the deployment of PyTorch workloads directly onto Google TPUs, the new TorchTPU stack offers Meta a viable alternative path. Google co-developed this tool specifically with Meta’s PyTorch team, targeting the specific needs of a major customer that was hesitant to migrate its entire engineering base to JAX.

This is not the first time Meta has explored Google’s hardware, but it is the first time it has reached full production maturity. Google identifies PyTorch/XLA, an older bridge between PyTorch and TPUs, as the predecessor to the current TorchTPU. Industry reports indicate that Meta research teams utilized TPUs via this older bridge between 2020 and 2023 before canceling those contracts. However, the current push appears more robust, with reports suggesting that last year's negotiations included Google Cloud TPU rentals for 2026 and potential purchases in 2027, followed by the signing of a multi-year rental agreement. The precise location of these production TPUs remains undisclosed, but the contractual framework suggests a long-term commitment.

The timing of this announcement is particularly pointed, arriving just one day after Amazon revealed it was open-sourcing TorchNeuron, its own native PyTorch backend for Trainium chips. Both Google and Amazon are executing similar strategies: by providing first-class PyTorch support for their custom silicon, they are attempting to loosen the grip that Nvidia’s software ecosystem has on the AI community. For developers, this means the choice of hardware is becoming decoupled from the choice of programming framework, fostering a more competitive and flexible market.

Performance metrics from the new stack suggest that Google’s hardware is competitive with Nvidia’s latest offerings. A tuned vLLM setup, built in collaboration with Inferact, achieved over 700 tokens per second for a single user on the seventh-generation Ironwood TPU while serving Moonshot AI’s Kimi K3 model. In comparison, Nvidia’s GB200 chips reached 452 tokens per second in similar published tests. Furthermore, the new kernels for models like Kimi K3 and Qwen 3.8 27B deliver throughput that is roughly 1.4 to 2 times higher than the GB200 baseline at small batch sizes, even without speculative decoding. These figures indicate that the software optimization has closed the gap that previously favored Nvidia in raw inference speed.

Looking ahead, the ecosystem is preparing for broader model support. SGLang-TorchTPU is scheduled to add support for DeepSeek V4.1 Flash and Reflection Beam next. RadixArk, the startup behind the SGLang engine, aims to achieve Day 0 TPU support for new open models by the first quarter of 2027. This means that when new models are released, they will be immediately compatible with Google’s hardware. Google, Meta, and RadixArk plan to share further details at the SGLang Summit 2026 in San Francisco in November, suggesting that this infrastructure shift is just the beginning of a broader realignment in the AI hardware market.

For readers in the Great Lakes and Ohio Valley, where major data center hubs and tech infrastructure investments are increasingly common, this shift signifies a more diversified supply chain. The era of the 'Nvidia-only' AI stack is ending, replaced by a multi-vendor environment where software compatibility, rather than proprietary lock-in, drives adoption. This could lead to more competitive pricing and greater flexibility for organizations building AI systems in the region, as they are no longer forced to choose between the two dominant players.