Google’s custom AI chips, called tensor processing units (TPUs), are now running AI workloads in production at Meta, according to a Google developer blog announcement reported by Tom’s Hardware. The setup uses a PyTorch-native version of SGLang, software used to run AI models and generate responses. Google also announced a PyTorch-native version of vLLM, another widely used inference engine.
The change gives Meta a way to use Google’s chips with PyTorch, the AI framework Meta created, instead of relying on Google’s JAX framework as before. Google says the new versions let developers use TPUs mostly without rewriting their software. The article’s headline also says Google has sold TPUs to external customers for the first time, but the supplied report does not identify those customers or give sales details.
A route beyond JAX
The software connection matters because a chip is only useful to developers if their tools can make use of it. Previously, SGLang and vLLM ran on Google TPUs through JAX. The PyTorch-native versions make the chips an option for teams already using PyTorch, potentially reducing the work needed to move an AI workload onto TPU hardware.
Google says the vLLM version is open source. The report does not say whether the SGLang version has the same licensing or access terms, or explain how developers can obtain either version. It also gives no figures for how many TPUs Meta is using, which models they serve, or how the setup performs.
What changes for users?
For Meta, the confirmed change is another way to run AI workloads in its data centers. That could give its engineers more hardware options, but the report offers no evidence of changes to the speed, quality or availability of Meta’s products for users. It also does not say whether Meta’s deployment is a test or a broader rollout beyond the statement that it is in production.
Why it matters
Meta’s AI engineers can now use Google TPUs with PyTorch-native versions of SGLang and vLLM, rather than relying on JAX for those engines. That adds a hardware option for Meta, but the source does not show whether people using Meta’s products will notice any difference in speed or features.
