Understanding AI cost

Where the money goes, and the few decisions that actually move it.

The shape of the cost

Two separate bills. Infrastructure is a monthly floor that exists whether or not anyone generates anything, dominated by the database. AI is per unit of work, dominated by course generation.

For a course, the expensive part is usually not what people expect: rich media generation typically costs more than the language model work behind lesson plans and questions.

What drives it

Course generationThe single largest AI line. Paid once per course — which is why cloning an already-generated course for a new term is materially cheaper than re-running the pipeline.

RegenerationRegenerating a whole course’s materials to fix one artefact pays the full price again. Regenerate the item.

Media generationInfographics, podcast audio and slide decks are the heaviest per-artefact costs.

GradingPer-attempt and modest, but it scales with cohort size and assessment frequency.

Database floorThe largest infrastructure line. The warm floor is what stops the first morning request from paying a cold start.

Controls that are already in place

Tiered models. Vision work does not run on the model used for curriculum reasoning.
A tight rate limit on LLM endpoints, far tighter than general API limits, because their cost is measured in tokens.
Content-addressed checkpoints in podcast generation, so a retry does not re-synthesise audio it already produced.
Durable jobs, so a deploy does not force a repeat.
Cached report queries per institution.

What actually saves money

In order of impact: get the syllabus right before extraction so you are not regenerating from a bad structure; clone courses instead of re-running the pipeline on an unchanged syllabus; regenerate individual artefacts rather than whole courses; and generate media for the lessons that need it rather than for everything by default.