Cognitive Kardashev Scale Quantifies Civilizational AI Computation
A new study published on arXiv has introduced something called the Cognitive Kardashev Scale, which measures the computational power of civilizations, kind of like the original Kardashev scale that looks at energy consumption. This new framework breaks down AI-level computation into three categories: planetary (Type I, around 10^16 W), stellar (Type II, about 10^26 W), and galactic (Type III). It takes into account total power P, how much of that power f is used for cognitive tasks, efficiency η (in operations per joule), and a benchmark based on brain processing rates. Using tech from 2024-2026, η is pegged at 10^12 FLOP/J. Currently, humanity scores around K ≈ 0.73, approaching Type I, which could mean one personal AI per person. Type II, however, is beyond imagination.
Key facts
- Paper published on arXiv with ID 2605.22840
- Builds on Kardashev's 1964 typology ranking civilizations by power
- Introduces Cognitive Kardashev Scale for AI computation
- Uses four ingredients: power P, cognitive fraction f, efficiency η, brain reference C_brain
- Anchored on 2024-2026 hardware: El Capitan, NVIDIA Blackwell, Vera Rubin
- Efficiency η = 10^12 FLOP/J for 2026 hardware
- Current humanity at K ≈ 0.73, three-quarters to Type I
- At Type I with 1% cognitive fraction: one personal AI per human
- At Type II: compute is incomprehensible
Entities
Institutions
- arXiv