Intelligence per joule
The ratio of useful cognitive output — tokens, decisions, solved tasks — to the energy consumed producing it. The core measure of the Joule Wars.
Intelligence per joule is the ratio of useful cognitive output to the energy consumed producing it. The numerator is not raw tokens — it is accepted work: answers that survive verification, decisions that get acted on, tasks completed without a human redoing them. The denominator is joules, because energy is the one input in the AI stack that does not follow Moore's law.
The metric is dimensionally more fundamental than the mobile era's performance-per-watt: a watt is a joule per second, so FLOPS per watt reduces to FLOPs per joule — the seconds cancel. What remains is the physical question every AI operator eventually faces: how much useful intelligence do you get out of each unit of energy you buy?
Intelligence per joule matters twice. In the data center it is a cost curve — the margin between what a completed task earns and what its inference cost. In embodied AI it is a design constraint: a humanoid robot carries 1–3 kWh on its back, so every joule spent thinking is a joule taken from the motors. That is why the metric anchors the Joule Wars thesis: as capability commoditizes, whoever maximizes useful intelligence per joule wins the paid work.
Frequently asked questions
What is intelligence per joule?
The ratio of useful cognitive output — tokens, decisions, completed tasks — to the energy consumed producing it. It is the core metric of the Joule Wars: as AI capability commoditizes, the deciding variable becomes how much accepted work a system produces per unit of energy.
How is intelligence per joule different from performance per watt?
A watt is a joule per second, so FLOPS per watt is dimensionally FLOPs per joule — the seconds cancel. Intelligence per joule goes further by counting useful output (verified, accepted work) rather than raw operations, which makes it an economic metric, not just a hardware one.
Why does intelligence per joule matter for robotics?
A humanoid robot carries its entire energy budget — roughly 1–3 kWh — on its back, and inference competes with the motors for the same joules. Battery density improves a few percent a year while compute efficiency compounds, so cognition-per-joule becomes a design constraint on par with mass and torque.
Source analyses
Related terms
Term defined and maintained by Michał Piszczek — CTO of Archdesk, author of the Joule Wars and Proof-Adjusted Autonomy frameworks. Quote freely with attribution to piszczek.pl/glossary (CC BY 4.0).