Everything K-culture — comebacks to K-beauty, straight to your inboxGet it in your inbox

METAL MEDIA

K-Culture GlossaryWhere everyone starts

Training & Inference

The two phases of AI's life — the time it spends learning (training) and the time it spends working (inference). The economics of the two are completely different.

In plain words

An AI's life splits into two phases: the time it spends learning (training) and the time it spends putting that learning to work (inference).

Training is like years of studying for an exam. The model spends months reading through massive amounts of data, adjusting its parameters as it goes, which requires tens of thousands of GPUs and an astronomical electricity bill. Headlines like "Company X spent hundreds of billions of won training its new model" are about this phase. Training itself splits further into "pre-training," which builds the model's basic capabilities, and follow-up training that layers on manners and practical judgment.

Inference is what happens after graduation, once the model starts its job. Every time a fully trained model receives a question and produces an answer, that's inference — and every time we talk to a chatbot, it triggers an inference cost somewhere in a data center. As the user base grows, inference ends up costing more than training, which is why most news about AI companies' pricing plans and data center investments is really about inference costs.

How it shows up in the news

"A new chip that cuts inference costs by 90%" — this means it lowers the cost of serving the model (inference), not the cost of building it (training). Once you can tell the two phases apart, chip and semiconductor news gets a lot easier to follow.

"Reasoning" is used to mean two different things

Is it about actually running the model?, Is it about thinking before answering?

This is the most dangerous overlap in Korean AI reporting. English has two distinct words, reasoning and inference, but both have often been rendered with the same Korean term. A sentence like "inference costs have spiked" is usually about the cost of running the model, while "reasoning ability has improved" is about the model's capacity to think. Same word, different event.

AspectTraining & InferenceReasoning Model
Original terminference — running a trained model to produce an answerreasoning — thinking things through logically
Oppositetraining — the time spent learninganswering immediately without thinking it through
When it improvesanswers come out faster and cheaperit solves hard problems like math or coding
In articles"inference cost," "inference-only chip," "inference demand""reasoning ability," "spent more time thinking"
How they relateso this cost rises — this is behind today's chip shortagethe longer it thinks, the longer it takes to run

Rule of thumbIf it's preceded by "cost," "chip," or "demand," it's about running the model (inference); if it's preceded by "ability" or "process," it's about thinking (reasoning).

See also

Stories using this term

No story has used this term yet. New ones attach here automatically.

Browse every entry