Meta Puts Google TPUs to Work in Its AI Infrastructure

The artificial intelligence hardware landscape has long been dominated by a single vendor, but that monopoly is showing cracks. This week, Google announced a significant step that places its custom tensor processing units (TPUs) directly into the hands of one of its biggest competitors, Meta.

The core of this development is a new software integration that allows Google’s TPUs to run natively on PyTorch, the AI framework that Meta uses for its own models. Previously, running PyTorch code on TPUs required using Google’s proprietary JAX framework as an intermediary. The new versions of the SGLang and vLLM inference engines remove that barrier, letting Meta deploy its existing workloads on Google hardware without a massive rewrite.

Zongwei Zhou, Meta’s lead for PyTorch-TPU, confirmed that the new SGLang-TorchTPU engine is already running in production at Meta. He described the achievement as the result of a tight collaboration between Meta, Google, and the team at RadixArk, the startup behind SGLang. Zhou noted that the engine provides Meta with a highly mature serving stack that feels like a natural extension of their current tools.

Jana van Greunen, who directs PyTorch engineering at Meta, emphasized the strategic importance of this move. She stated that Meta cannot limit itself to a single hardware architecture and that this new integration enables the company to deploy its PyTorch workloads directly to Google TPUs. Google confirmed it co-developed this technology with Meta’s PyTorch team to ensure compatibility.

This is not the first time Meta has explored Google’s silicon, but it is the most seamless integration yet. Google’s previous bridge, PyTorch/XLA, was used by Meta research teams between 2020 and 2023, but those contracts were eventually canceled. The new TorchTPU solution is a more direct successor, designed specifically to address Meta’s preference for staying within the PyTorch ecosystem rather than migrating to JAX.

While the exact location of these production TPUs remains undisclosed, industry reports suggest that negotiations last year involved Google Cloud rentals for 2026, with hardware purchases potentially following in 2027. Meta has reportedly signed a multi-year rental agreement, indicating a long-term commitment to diversifying its compute resources beyond its preferred partners.

The timing of this announcement is notable, as it came just one day after Amazon revealed its own open-source backend, TorchNeuron, for its Trainium chips. Both moves appear to be strategic efforts to loosen the software lock-in that has historically favored Nvidia. By making their hardware native to the industry-standard PyTorch framework, these companies are making it easier for developers to switch or add new hardware to their stacks.

Performance benchmarks released alongside the news show promising results. A tuned version of vLLM running on the seventh-generation Ironwood TPU reportedly achieved over 700 tokens per second for a single user when serving a massive open-weight model. This test, conducted by Inferact, used 16 Ironwood chips with speculative decoding. In comparison, a similar setup using 16 Nvidia GB200 chips reached 452 tokens per second. The new kernels also delivered roughly 1.4 to 2 times the throughput of the Nvidia baseline at small batch sizes without speculative decoding.

Looking ahead, the developers plan to expand support for more models, starting with DeepSeek V4.1 Flash and Reflection Beam. They are also targeting Day 0 support for new open models by the first quarter of 2027. More details on these developments are expected to be shared at the SGLang Summit in San Francisco in November.