Google, AMD Reportedly Team Up on 10th-Generation TPU

Google has long relied on its in-house Tensor Processing Units to train and run artificial intelligence models, reducing its dependence on general-purpose GPUs. Reinforcement learning and agentic AI, however, require frequent switching between model inference, decision-making and conventional computing tasks. Bringing CPU and tensor-processing resources closer together could reduce data-transfer bottlenecks, improve latency and lower power consumption as those workloads become more complex.

Google is working with Advanced Micro Devices on a 10th-generation TPU that would combine CPU and TPU cores in a hybrid package, according to market reports. The design is intended to optimize reinforcement-learning and agentic-AI workloads by shortening the distance between general-purpose and tensor computation. As of Aug. 17, 2026, the companies had not publicly disclosed a production timetable, manufacturing process, performance targets or power-efficiency figures for the reported chip.