Because 1 person is responsible for 1 person’s cost of inference.
Training is a whole industry manufacturing analogue. That cost must be amortised across every user, not attributable to one person, to be a fair apples to apples comparison.
We can compare 1 person to 1 person, or 1 industry to 1 industry. We can’t fairly compare 1 person to 1 industry.
See my reply elsewhwere to NGram.