**Training compute:** GPU/CPU time for training and retraining models. Costs scale with model complexity, dataset size, and retraining frequency. - **Inference compute:** CPU/GPU time for serving predictions. For real-time models, this cost scales directly with traffic. For batch models, it scales w