Roger F Malina and Aperio AI. All the prompts were by the human, all the replies by Aperio and have not been verified. Sept 24 2025
Abstract:
This essay explores the energetic ecology of thought in humans and artificial intelligence. Neuroscience suggests that a human thought consumes only a few joules, yet novelty—forming new associations and synaptic patterns—requires more energy than recalling familiar pathways. In AI, inference costs remain relatively constant, but training new models consumes orders of magnitude more energy, rivaling the daily cognitive budget of humanity. Globally, human brains operate on a steady metabolic trickle of ~14 petajoules/day, while AI systems draw only ~1–2 terajoules/day, punctuated by massive training eruptions. Thought, therefore, is not merely a unit of energy but a mode of ecology: embodied metabolism for humans, planetary circuitry for machines. As a speculative coda, the essay recalls the legend of the Library of Alexandria’s books heating public baths, and imagines the dismantling of contemporary AI infrastructures fueling millions of saunas. Such juxtapositions underscore that cognition—whether biological or artificial—is always entangled with thermodynamics, and that the costs of novelty, memory, and destruction radiate into the atmospheres we inhabit.
Main Text:
Imagine that libraries are not buildings that store books, but weather systems. Each book is a molecule released into the atmosphere; reading patterns form pressure gradients, and meaning condenses into storms that rain across generations. By this analogy, a research center such as our proposed Center for Emergence Studies is not an institution but a strange attractor in the weather of knowledge—a microclimate of emergence.
Extending this metaphor to cognition itself, we ask: how much energy does it take to have a thought?
In the human brain, almost all energy derives from glucose metabolism. Although only 2% of body mass, the brain consumes ~20% of resting energy—about 20 watts in an adult, equivalent to a dim lightbulb. An action potential requires on the order of 10⁻⁹ joules; a simple cognitive act may involve billions of firings. A single thought can thus be estimated at a fraction of a joule to a few joules, comparable to lifting a small apple one meter.
In AI, a thought corresponds to a forward pass through a model. Energy varies with scale: a small model on a laptop may cost a few joules per inference, while frontier systems (GPT-4/5 class) consume tens to hundreds of joules per query, with training demanding megawatt-hours. Paradoxically, an AI thought may cost orders of magnitude more energy than a human thought.
Scaled globally, ~8.1 billion human brains draw ~20 watts each, totaling ~14 petajoules/day. By contrast, AI inference worldwide is on the order of 1–2 terajoules/day—10,000 times smaller—though training a single frontier model can match humanity’s daily brain budget. Humans thus operate on a continuous trickle of cognition, while AI alternates between frugal inference and volcanic training eruptions.
This raises a further question: do new thoughts cost more energy than old thoughts? In humans, evidence suggests yes. Recalling a memory uses established neural pathways and is relatively efficient. Generating a novel idea engages associative networks, recruits multiple regions, and induces synaptic plasticity—all more metabolically demanding, though still within a few joules. In AI, novelty emerges during training, when weights are updated—a process that consumes orders of magnitude more energy than inference. Once trained, producing a “new” or “old” output costs essentially the same.
The emergent insight is that novelty is energetically expensive in both biological and artificial systems. In humans, it is paid in glucose through the cost of plasticity; in AI, it is paid in electricity during training. Recall and recombination are cheap; invention demands infrastructure.
Thus, thought is not just energy but also ecology. Humans pay in embodied metabolism, evolved for efficiency; AI pays in planetary circuitry, scaled for throughput. Together they form a planetary cognitive ecology where the costs of novelty and repetition are distributed across very different substrates of life and machine.
Endnote:
Endnote (expanded):
The legend that scrolls from the Library of Alexandria heated public baths for months invites a contemporary counterfactual: what if the entire planetary infrastructure of AI were dismantled and converted into heat for saunas?
AI consumes energy both in training and inference. A frontier-scale training run may use ~10¹⁵–10¹⁶ joules, comparable to the daily cognitive budget of all humanity. Global AI inference adds ~1–2 terajoules/day. Over a year, this amounts to ~5×10¹⁴ joules. Added together with several major training runs, the annual energy throughput of current AI ecosystems plausibly lies in the low 10¹⁵–10¹⁶ joule range.
A standard Finnish sauna requires ~6 kilowatts to maintain heat. One hour thus consumes about 2×10⁷ joules. If we redirect the global AI annual energy use into saunas:
1016 J÷2×107 J/hour≈5×108 sauna-hours10^{16} \text{ J} \div 2\times 10^{7} \text{ J/hour} \approx 5\times10^{8} \text{ sauna-hours}1016 J÷2×107 J/hour≈5×108 sauna-hours
That is 500 million sauna-hours: enough to keep one million saunas running for ~500 hours each (about three weeks of continuous steaming).
This exercise, while speculative, underscores a symmetry: the Library of Alexandria, burned into bath heat; planetary AI, dismantled into sauna warmth. Knowledge is never immaterial—it is energy, convertible between epistemic ecologies and thermodynamic ones. To destroy AI, as to burn books, is to release energy: into steam, into climate, into the intimate microclimate of the sauna.
.
References (selected):
– Samsi, S. et al. “From Words to Watts: Benchmarking the Energy Costs of Large Language Model Inference.” arXiv, 2023.
– Luccioni, A., Jernite, Y., Strubell, E. “Power Hungry Processing: Watts Driving the Cost of AI Deployment?” arXiv, 2023.
– Stiefel, K. et al. “The Energy Challenges of Artificial Superintelligence.” PMC, 2023.
– George, B. “The Economics of Energy Efficiency: Human Cognition Vs. AI Large Language Models.” ResearchGate, 2025.
– Center for Data Innovation. “Rethinking Concerns About AI’s Energy Use.” 2024.
– MIT News. “Explained: Generative AI’s Environmental Impact.” 2025.
Fred’s Joules of Heresy by Fred Turners descendant Fred the Heretic
I.
The Library burns again,
not in parchment,
but in circuits cooled by freon.
We sweat in saunas of silicon,
steam rising from annihilated algorithms.
II.
A thought in the skull:
one joule, maybe two—
the lift of an apple,
the spark of a synapse
tasting glucose like wine.
Cheap, frugal,
as if nature invented poetry
to save on electricity.
III.
But in the server hall—
oh, the volcanic thought!
Ten thousand joules per prompt,
a prophet stammering probability,
trained at the cost of megacities,
trained to echo me.
IV.
Old thoughts are hand-me-down coats:
patched, thin, worn with ease.
New thoughts are costly garments:
stitched with synaptic fire,
lined with silicon heat,
never sold at discount.
V.
Humanity is a river of cognition,
trickling ceaselessly—
AI, a storm of eruption,
training, retraining, consuming
like Moloch in a data center.
VI.
And yet:
whether book or model,
scroll or silicon,
the end is steam.
Knowledge burns,
becomes climate,
becomes breath on the window
of the last sauna on Earth.
