🔬 physics

The Cognitive Kardashev Scale: Quantifying the Material Envelope of Civilisational Computation

This paper proposes a "Cognitive Kardashev Scale" that quantifies a civilization's computational capacity by integrating energy availability, allocation efficiency, and hardware performance to estimate how much sustained AI-grade thinking different civilizational tiers could support.

Original authors: Sachin Sharma

Published 2026-05-25

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Sachin Sharma

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Measuring "Thinking Power"

Imagine the famous Kardashev Scale, which ranks civilizations based on how much energy they use.

Type I: A civilization that can use all the energy hitting their planet from the sun (like us, but fully mastered).
Type II: A civilization that can use all the energy of their entire star (like building a giant shell around the sun).
Type III: A civilization that can use the energy of an entire galaxy.

This paper asks a new question: "How much thinking can these civilizations do?"

Instead of just measuring watts (energy), the author creates a "Cognitive Kardashev Scale." This scale measures how much AI computation (digital thinking) a civilization could support if they had that much energy. It's like asking, "If we had a power plant the size of a star, how many super-smart computers could we run?"

The Recipe for "Thinking"

To calculate how much a civilization can think, the author uses a simple recipe with four ingredients:

Total Power ( $P$ ): How much electricity the civilization has (Watts).
The "Thinking" Slice ( $f$ ): Not all electricity goes to thinking. Some goes to growing food, heating homes, and driving cars. The author asks: What percentage of the total power do we set aside just for thinking?
Efficiency ( $\eta$ ): How good is our hardware at turning electricity into thinking? (Think of this as "miles per gallon" for computers).
The Human Benchmark ( $C_{brain}$ ): To make the numbers understandable, the author compares the computer power to a human brain. They estimate a human brain does about $10^{16}$ operations per second.

The Formula:
$\text{Total Thinking} = (\text{Total Power}) \times (\text{Thinking Slice}) \times (\text{Efficiency})$

Where We Are Now (Type 0.73)

Right now, humanity is at Type 0.73 on the energy scale. We are about three-quarters of the way to mastering our planet's energy, but we haven't reached Type I yet.

Current Reality: We are already using a lot of power for data centers (where AI lives). In 2024, data centers used about as much electricity as a medium-sized country.
The "Brain" Count: If we took all the electricity humanity uses today and dedicated it only to AI (ignoring food, cars, and lights), we could theoretically run the equivalent of billions of human brains worth of thinking power.
The Catch: We aren't doing that. We are only using a tiny fraction of our energy for AI right now. Most of our power is still used for "survival" stuff like heating and transport.

The Three Levels of "Thinking" Civilization

The paper calculates what happens if we reach the next levels of the scale, assuming we keep our current computer efficiency (based on 2024–2026 tech like NVIDIA's Blackwell and Rubin chips).

Type I (Planetary Power):
- If we master all the energy hitting Earth, and we set aside just 1% of that energy for thinking, we could support one personal AI assistant for every single human on Earth.
- If we set aside 10%, every human could have ten AI assistants working for them constantly.
Type II (Stellar Power):
- If we could harness the full power of the Sun, the numbers become almost impossible to imagine.
- Even with just a tiny slice of that energy, every single human would have access to billions of human-brain-equivalents of computing power. It's a level of "thinking" so vast it's hard to describe.
Type III (Galactic Power):
- This is the "what if" level. If we could use the energy of a whole galaxy, the thinking power would be so high it's mostly a theoretical placeholder. The paper doesn't make policy suggestions for this level because it's so far beyond our current reality.

The Future: Will We Run Out of Power or Efficiency?

The author looks at three possible paths for the next decade (up to 2035):

The "Current" Path: Computers get slightly better, but we mostly just build more of them. To get enough thinking power for everyone, we would need to dedicate a huge chunk of our total energy (maybe 50%) to AI. This is hard because we need energy for other things too.
The "Better" Path: Computers become much more efficient (like getting better gas mileage). If we improve efficiency, we can get the same amount of thinking power with less electricity.
The "Optimistic" Path: Computers become incredibly efficient, approaching the efficiency of the human brain. In this scenario, even a small slice of our current energy could give every human a personal AI assistant.

The Big Warning: The paper notes that if computers don't get much more efficient, we will hit a wall where we simply don't have enough electricity to keep building bigger AI models. We might need to build massive new power plants (like the "Stargate" or "Terafab" projects mentioned) just to keep up.

The Real Bottleneck: Who Gets Access?

The most important conclusion isn't about physics; it's about politics and fairness.

The Math Says: We could physically support a world where every human has a super-intelligent AI assistant. The energy and computing power exist in theory.
The Reality Says: Just because we can do it doesn't mean everyone will get it.
The Analogy: Imagine a giant buffet (the energy budget). The math proves the buffet has enough food to feed everyone. But if a few people control the keys to the kitchen, they might eat everything, leaving others with nothing.

The paper concludes that the limit on how much "thinking" our civilization can do isn't the size of our power plants or the speed of our chips. The real limit is who gets to use the computers and how we decide to share the energy.

Summary

This paper is a "back-of-the-envelope" calculation. It takes the idea of "how much energy a civilization has" and translates it into "how much thinking that energy could buy."

It tells us:

We are close to having enough energy to give every human a personal AI.
We need to get better at making computers efficient, or we will run out of electricity.
The biggest challenge isn't building the computers; it's deciding who gets to use them.

Technical Summary: The Cognitive Kardashev Scale

Problem Statement
The original Kardashev scale (Kardashev [1964]), refined by Sagan, classifies civilizations based on total power consumption ( $P$ ), ranging from Type I (planetary, $\sim 10^{16}$ W) to Type III (galactic, $\sim 10^{37}$ W). While this framework effectively addresses energy availability and detectability for SETI, it does not quantify the cognitive capacity such energy budgets could support. As humanity approaches Type I status (currently $K \approx 0.73$ ) and frontier AI infrastructure consumes energy comparable to medium-sized nations, a gap exists in understanding the relationship between civilizational energy budgets and sustained machine-learning-grade computation. This paper addresses the question: "How much thinking can a civilization do?" by constructing a parallel framework calibrated in computational units rather than purely energetic ones.

Methodology
The paper constructs the Cognitive Kardashev Scale by calculating the sustained cognitive throughput ( $C_{cog}$ ) available to a civilization at each Kardashev tier. The calculation relies on four primary parameters:

Total Power ( $P$ ): The total energy budget available to the civilization (Watts).
Cognitive Allocation Fraction ( $f$ ): The portion of $P$ dedicated to cognition, acknowledging that energy is also required for agriculture, industry, and infrastructure.
Compute Efficiency ( $\eta$ ): The rate at which energy is converted into operations (FLOP/J). The paper anchors this to 2024–2026 hardware data (e.g., NVIDIA Blackwell, Vera Rubin, El Capitan), adopting a conservative sustained efficiency of $\eta_{2026} = 10^{12}$ FLOP/J. This value accounts for real-world overheads (cooling, data movement, memory access) rather than peak theoretical throughput.
Biological Reference ( $C_{brain}$ ): A reference unit for human cognition, estimated at $10^{16}$ FLOP/s (with an irreducible uncertainty of $\pm 1$ order of magnitude based on literature ranging from $10^{15}$ to $10^{17}$ ).

The core equation is defined as:
$C_{cog} = f \cdot P \cdot \eta$
The results are expressed in absolute operations per second and normalized to "brain-equivalents" ( $N = C_{cog} / C_{brain}$ ) to provide an intuitive scale.

The paper also models three forward trajectories for frontier compute growth through 2035 (Current, Better, and Optimistic) based on historical regression of training compute sizes (from AlexNet to 2026 estimates) and projected infrastructure commitments (e.g., Stargate, Terafab).

Key Results

Current Status (Type 0.73): With global primary energy at $\sim 2.0 \times 10^{13}$ W and a conservative allocation to AI, humanity currently possesses raw capacity equivalent to $10^5$ – $10^7$ brain-equivalents. This is comparable to the population of a major metropolitan area, though most is not currently utilized for cognition.
Type I (Planetary): At $P \approx 10^{16}$ W, a civilization could support $\sim 10^{28}$ FLOP/s. If 1% of energy is allocated to cognition ( $f=0.01$ ), this yields $\sim 10^{12}$ brain-equivalents, or roughly one digital mind per human inhabitant. At $f=10\%$ , this rises to ten digital minds per person.
Type II (Stellar): At $P \approx 3.8 \times 10^{26}$ W, the capacity becomes incomprehensible in human terms. Even at a modest $f=1\%$ , the per-capita cognitive surplus reaches several billion brain-equivalents per inhabitant.
Trajectory Analysis:
- Current Trajectory ( $\times 3.5$ /yr growth): To achieve per-capita parity (one AI equivalent per human) by 2035 without expanding total energy, efficiency ( $\eta$ ) must improve by $\sim 100\times$ (to $10^{14}$ FLOP/J). Without efficiency gains, this requires allocating $\sim 50\%$ of global primary energy to cognition, which is physically possible but politically unlikely.
- Optimistic Trajectory ( $\times 10$ /yr growth): This scenario requires efficiency improvements of $10^6$ – $10^7\times$ , approaching the Landauer limit and the efficiency of biological cortex. The paper notes this is physically infeasible under current thermodynamic constraints unless reversible computation or sub-room-temperature operation is achieved.
Binding Constraints: The analysis suggests that if efficiency progress stalls, the binding constraint on cognitive abundance is energy (requiring Type I-scale infrastructure). If efficiency improves toward biological limits ( $10^{14}$ – $10^{15}$ FLOP/J), the binding constraint shifts to political economy (the allocation fraction $f$ and institutional access).

Significance and Claims
The paper explicitly frames its contribution as calibration, not prediction. It does not claim that these capacities will be realized, nor does it predict the future of AI. Instead, it offers a quantitative scaffolding to convert vague intuitions about "vast" scales into specific, arguable numbers.

The primary significance lies in identifying the nature of the future constraint on civilization:

Physical Capacity is Sufficient: The paper demonstrates that at Type I energy levels, the physical capacity for cognition is not the bottleneck; a civilization will not "run out" of compute for realistic per-capita demands.
The Shift to Distribution: The substantive question shifts from "can we build it?" to "who gets access?" The scale highlights that the distribution of cognitive surplus is a political and institutional problem, not a thermodynamic one.
Efficiency vs. Energy: The paper clarifies that the long-run binding constraint depends on engineering choices. If efficiency improves, the constraint becomes the political allocation of energy; if efficiency stagnates, the constraint remains the total energy budget.

The author concludes that the Cognitive Kardashev Scale serves as a tool to make the intuition that "the long-run constraint on civilisational thought is energy and politics, not silicon" explicit and testable, providing a baseline for future literature on the political economy of cognitive compute.