Meta runs production AI on Google's chips, not just Nvidia's

Meta, the company behind PyTorch, is now running production AI workloads on Google TPUs through a PyTorch-native version of the SGLang inference engine, as Tom's Hardware first reported. Oracle has also agreed to offer Google TPUs through its own cloud, making OCI the first hyperscaler outside Google to sell the chips, per Google's own announcement. Analyst estimates for how big this gets diverge wildly: Piper Sandler sees $51 billion in external TPU revenue by 2027, Barclays sees $67 billion, and Morgan Stanley separately pegs Google Cloud at 48% of Alphabet's total operating profit by 2028. Google's own TPU v9 is reportedly facing a packaging-yield delay that could push shipments to early 2028, which means the demand signal is arriving faster than Google can build the chips to meet it.