AI Infrastructure Shift: Meta Integrates Google TPUs via New PyTorch Engine

For organizations managing large-scale artificial intelligence operations, the choice of hardware and software has become a critical strategic decision. Effective immediately, Meta has begun running production AI workloads on Google’s custom silicon, marking a significant diversification of its infrastructure stack. This development allows the social media giant to utilize Google’s Tensor Processing Units (TPUs) natively within its existing PyTorch framework, reducing the need for extensive code rewrites and offering a more flexible deployment model for its engineers.

The operational change relies on the release of SGLang-TorchTPU, a new inference engine designed specifically for this hardware-software combination. Previously, running PyTorch models on TPUs required bridging through Google’s proprietary JAX framework, a process that many developers found cumbersome. The new engine eliminates this friction, providing a mature serving stack that integrates directly with the tools Meta’s teams already use daily.

Meta’s PyTorch-TPU lead, Zongwei Zhou, confirmed that the system is already active in production environments. He described the achievement as the result of a three-way collaboration between Meta, Google, and the SGLang team. Zhou noted that the engine provides his team with a remarkably mature, PyTorch-native serving stack, signaling that the hardware is no longer just a theoretical option but a viable component of Meta’s daily operations.

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 and that the new TorchTPU implementation enables the company to deploy its PyTorch workloads to Google TPUs seamlessly. Google confirmed that it co-developed this specific version of the software with Meta’s PyTorch team to ensure compatibility and performance.

Contextually, this is not Meta’s first foray into TPU usage, but it represents a more mature integration. Earlier efforts relied on PyTorch/XLA, which Google now designates as a predecessor to the new TorchTPU. Industry reports indicated that Meta research teams utilized TPUs through this older bridge between 2020 and 2023 before canceling those contracts. The recent push by Google was heavily targeted at Meta, a company that had previously resisted migrating its core operations to the JAX framework.

The logistics of this hardware acquisition remain partially opaque, though industry chatter suggests a structured transition. Reports from last year indicated discussions regarding Google Cloud TPU rentals for 2026, with potential purchases scheduled for 2027. Meta was subsequently reported to have signed a multi-year rental agreement, suggesting a long-term commitment to diversifying its compute resources beyond its current primary providers.

This announcement arrived one day after Amazon revealed that it was open-sourcing TorchNeuron, a native PyTorch backend for its Trainium2 and Trainium3 chips. Both moves appear strategically aligned to loosen Nvidia’s dominant software lock-in. By ensuring that major cloud providers’ custom chips work natively with PyTorch, the industry is working to reduce the dependency on a single vendor’s ecosystem, potentially leading to more competitive pricing and broader hardware options for AI developers.

Performance benchmarks also support the viability of this new stack. A tuned vLLM setup serving Moonshot AI’s Kimi K3 model reportedly achieved over 700 tokens per second for a single user on the seventh-generation Ironwood TPU. This test, conducted by Inferact, used 16 Ironwood chips with speculative decoding and outperformed a comparable setup on 16 Nvidia GB200s, which reached 452 tokens per second. For smaller batch sizes, the new kernels deliver roughly 1.4 to 2 times the throughput of the Nvidia baseline.

Looking ahead, the ecosystem plans to expand rapidly. SGLang-TorchTPU will begin adding support for new models, starting with DeepSeek V4.1 Flash and Reflection Beam. The teams are also working toward Day 0 TPU support for new open models by the first quarter of 2027. Google, Meta, and RadixArk, the startup behind SGLang, plan to share further details at the SGLang Summit 2026, scheduled for November 12–13 in San Francisco.