Meta’s TPU Deployment Marks a Strategic Blow to Nvidia’s AI Chip Monopoly
Google’s custom silicon is now running in production at Meta, signaling a shift in the AI hardware market as major tech firms seek to break free from Nvidia's software dominance.
By Ricardo Whitaker ·

Google’s tensor processing units (TPUs) have officially entered production at Meta, marking the first time the custom silicon has been sold to an external customer for large-scale AI workloads. This development represents a significant shift in the artificial intelligence hardware landscape, as Meta, the creator of the widely used PyTorch framework, now runs its inference engines natively on Google’s chips.
The move is underpinned by the release of new PyTorch-native versions of the SGLang inference engine and the open-source vLLM engine. Previously, running these tools on TPUs required reliance on Google’s JAX framework. The new implementations allow the hardware to function as a native option for PyTorch without extensive code rewriting, a critical change for developers who have built their workflows around PyTorch rather than JAX.
Performance metrics suggest a competitive challenge to Nvidia’s current flagship offerings. According to Inferact, the startup behind vLLM, a tuned setup serving Moonshot AI’s 2.8-trillion-parameter Kimi K3 model achieved over 700 tokens per second for a single user on the seventh-generation Ironwood TPU. This figure significantly outperforms the 452 tokens per second recorded on 16 Nvidia GB200 chips using the standard vLLM configuration. Furthermore, kernels for Kimi K3 and Qwen 3.8 27B delivered 1.4 to two times the throughput of the GB200 baseline at small batch sizes when speculative decoding was not utilized.
The collaboration behind this deployment involved a three-way partnership between Meta, Google, and the SGLang team. Zongwei Zhou, Meta’s PyTorch-TPU lead at RadixArk, the startup founded by SGLang’s creators and the engine’s builder, confirmed that the system is live. He described the result as a "remarkably mature, PyTorch-native serving stack," highlighting the depth of the technical integration achieved by the three parties.
Meta’s engineering leadership emphasized the strategic necessity of hardware diversity. Jana van Greunen, director of PyTorch engineering at Meta, stated that the company "cannot be constrained by a single hardware architecture." She noted that the new TorchTPU tooling, co-developed with Google’s PyTorch team, enables the deployment of PyTorch workloads directly to Google’s infrastructure, reducing dependency on any single vendor.
This is not Meta’s first foray into Google’s hardware ecosystem. Google identifies the older PyTorch/XLA bridge as a predecessor to the current TorchTPU. Reports from SemiAnalysis indicated that Meta research teams used TPUs through that older bridge between 2020 and 2023 before canceling those contracts. The current push is widely seen as a more robust solution to the previous friction, specifically addressing Meta’s reluctance to migrate its core infrastructure to JAX.
The commercial terms of the arrangement remain partially opaque, with the physical location of Meta’s production TPUs not yet disclosed. However, industry talks reportedly included Google Cloud TPU rentals for 2026 and potential purchases in 2027. Meta is also reported to have signed a multi-year rental agreement, suggesting a long-term commitment to this alternative hardware path.
The timing of the announcement is notable, arriving just one day after Amazon revealed it was open-sourcing TorchNeuron, a native PyTorch backend for its Trainium2 and Trainium3 chips, at the same PyTorch Conference. These concurrent moves by major hyperscalers appear designed to loosen Nvidia’s software lock-in, which has long relied on the dominance of its CUDA framework to maintain market share.
Looking ahead, the SGLang-TorchTPU ecosystem is planned to expand its model support, starting with DeepSeek V4.1 Flash and Reflection Beam. RadixArk aims to achieve Day 0 TPU support for new open models by the first quarter of 2027, ensuring that new releases are immediately compatible. Google, Meta, and RadixArk are scheduled to share further details at the SGLang Summit 2026, taking place November 12-13 in San Francisco.
For the Midwest business community, this shift signals a diversification of the AI supply chain. As costs and performance become the primary drivers of enterprise AI adoption, the entry of Google’s TPUs into major production environments outside of Google’s own cloud offers a tangible alternative to the Nvidia-centric market that has dominated the sector for the past decade.
More Business · Watch Your Wallet · Report a business · Got a Story?






