Training
Also known as: model training
Training is the process of adjusting a model’s weights so it gets good at a task. It runs the model on a huge dataset, measures each error, and nudges the weights to reduce it — billions of times.
It’s the expensive, one-off phase:
- Training: takes weeks, burns megawatts, happens once (or rarely).
- Inference: runs the finished model per request — cheap, fast, constant.
This train-once / run-many split is the fundamental economics of the whole field.
Cost intuition: you pay heavily once to *make* the model, then pay a little each time you *use* it. That's why providers charge per token — they're recovering the training bill.