Different questions need different costs
A user’s subscription price is not the complete production cost of an AI service. Hardware, labour, model development, operations and electricity can all matter. Training a model and serving many requests are distinct activities with different timing and cost structures.
Consider the next unit
Average development cost and the cost of an additional request answer different questions. Greater volume can spread some fixed costs, while capacity constraints or new infrastructure can change the marginal cost. A simple “AI is cheap” or “AI is expensive” claim needs an explicit unit and boundary.
Include effects outside the firm
Data-centre electricity demand connects AI to power systems. The social comparison may also include emissions, local constraints and other effects not captured by the user’s bill. Energy impacts differ by technology, location and electricity supply; not every AI task has the same footprint.
Measure benefits too
A full appraisal should compare useful output and alternatives, not just resource use in isolation. Ask what activity is displaced, how much demand changes and which impacts are measured. Environmental economics and econometrics provide complementary tools for examining those claims without assuming every application has the same result.