The cost of AI is broader than a subscription price. A useful assessment can include compute, energy, water, hardware, materials, human labor, review time, attention, trust and the burden created downstream.
What model, workload, frequency and infrastructure are required? Could a smaller or local system do the job?
Water and facility impacts vary by location, cooling design, electricity mix, hardware and accounting boundary. Avoid universal per-query numbers without context.
Accelerators, servers, networking, replacement cycles and embodied impacts belong in lifecycle discussions where material.
Review time, correction, training, monitoring, cognitive load and trust are resources too.
Instead of “How much energy does AI use?” ask: which system, doing what task, how often, on what infrastructure, in which location, compared with what alternative, over what lifecycle?
Q10's EE/EH approach is designed to make those boundaries explicit and to prevent resource savings from silently overriding truth, safety or human authority.