Inference is the phase where a trained AI model is actually used to produce predictions or answers, as opposed to training, where it learns from data. Every time you ask ChatGPT a question or a bank scores a loan application, that is inference. It is like the difference between a student studying for years and finally practicing as a doctor: training happens once and costs a lot, while inference happens millions of times and must be fast and cheap. Inference cost and speed drive most real-world AI budgets.