Optimizing Google AI workloads on cloud GPUs requires matching your model type—training, inference, or fine tuning—