Training is school. Inference is the job. Training happened before you ever showed up, when the model learned language from a huge slice of the internet. Inference is what happens every time you actually use it: the model takes your text in and predicts text back. Every message you send triggers inference, and inference is what companies pay for, every time, forever.
The distinction matters because it is the difference between a cost that already happened and a cost that never stops. The training bill was someone else's. The inference bill is yours, and it scales with usage: more people, more messages, more documents, more spend. A tool that looks cheap in a demo can look very different once a whole team leans on it constantly.
For an operator, inference is the line item to model before rollout, not after. The questions are plain ones. How many people will use this? How often? How long are the inputs and outputs? Because inference is billed in tokens, those answers translate directly into money, and a rough estimate on one page beats a surprise at the end of the quarter.
A concrete example. A company pilots an assistant that summarizes incoming support tickets. In the pilot, a handful of tickets, the cost is noise. Rolled out across every ticket the company receives, the same feature becomes a recurring bill that tracks ticket volume exactly. Nothing went wrong. That is simply what inference is: a running cost tied to real usage, and it belongs in the budget the day the tool ships.
“Inference costs scale with usage. The training already happened.”
// this page is the full text. the packaged PDF, all fifteen terms, comes with the free library.