An honest AI invoice would have two totals.
The first covers the visible request: electricity, cooling, chips, network traffic, and labor.
The second covers everything cheap intelligence invites people to add: agents that never sleep, endless retries, personalized persuasion, synthetic video, surveillance, and services left running because each individual call feels negligible.
Efficiency can lower the first total while detonating the second. “Energy per token” then improves beautifully as total demand climbs. The serious unit is useful human outcome per full system cost—including whose grid, water, wages, and neighborhood absorb the remainder.