The moment a model answers
In AI, inference is a trained model producing output for new input. Sending a question to ChatGPT and getting a reply, or a photo app recognizing a face — that live execution is inference.
Training and inference
AI work splits into two phases. Training is the preparation, adjusting parameters against large amounts of data at enormous computational cost over a long period. Inference is the performance: using the finished model to produce an answer, at a far smaller cost per request.
In cooking terms, training is years of apprenticeship and inference is making one dish to order. As users, we only ever interact with inference.
Inference costs add up
One request is light, but hundreds of millions of people asking questions daily makes the total enormous. Subscription pricing and free-tier limits exist mainly to cover inference. It is also why chip makers feature so heavily in AI business news — both phases need vast amounts of silicon.
How this relates to reasoning models
You will also see the phrase reasoning model, which describes models that spend time thinking in stages before answering. That is a different sense from inference as a processing phase. Telling which meaning is intended makes AI coverage easier to follow.
It is increasingly said that inference, not training, is now the main battleground. Reasoning models spend heavily at answer time, so the quality and efficiency of inference has become a competitive factor. Seeing AI cost as the sum of training and inference compute explains a lot about pricing and company news.
Competition to build chips specialized for fast, cheap inference is intense, and it has become a defining theme of AI infrastructure.