Running a trained model to produce an answer.
Inference is the act of using a model: a prompt goes in, the model runs a forward pass over its weights, and tokens come out. Training builds the model once; inference happens on every single request forever after.
That asymmetry is the whole economics of the AI industry. Training is a large capital event; inference is a recurring cost that scales directly with usage, which is why providers price per token and why cheaper inference silicon matters more each year.
Inference cost is driven by model size, context length, output length and how many requests share a GPU at once.
Inference demand is what the market is pricing when it prices accelerators: NVIDIA trades at 210.96, down 8.42% on the day. Every production chatbot request is another inference call running on hardware like it.