Your Brain Runs on 20 Watts. AI Wants a Power Plant
The human mind runs on roughly 20 watts. The world’s quickest supercomputer, LineShine in Shenzhen, attracts 42.2 million watts. That hole has turn into the web’s favourite argument about AI power use, and most of it’s mistaken.
The comparability itself holds up. However, the numbers circulating on social media hint again to a single paper. The most hanging one has been misattributed for 3 years.
AI Energy Use: What 20 Watts Actually Buys
The 20-watt determine rests on many years of metabolic measurement. The mind accounts for about 2% of physique weight and roughly 20% of resting oxygen consumption.
Neuron counts are shakier than they seem. The broadly quoted 86 billion rests on 4 male brains and is presently under dispute within the journal Brain.
Viral posts usually use 12 watts quite than 20. That determine seems in a 2023 paper in Frontiers in Artificial Intelligence, acknowledged with none quotation in any respect.
The identical paper produced the quantity everyone shares. Its authors estimated that digitally recreating a human mind would draw 2.7 billion watts.
That estimate got here from extrapolating a 10-million-neuron simulation to mouse scale, then multiplying by a thousand.
The paper additionally states that the simulation ran about 30,000 occasions slower than biology. Social posts drop that element. Secondary sources then credit score the determine to the Blue Brain Project, which by no means printed it.
Reliable numbers do exist elsewhere. Epoch AI estimated a typical ChatGPT question at 0.3 watt-hours in early 2025. A peer-reviewed examine in Joule later landed on 0.31.
Two unbiased strategies agreeing that intently is uncommon. However, the determine modifications sharply with workload, and reasoning fashions that produce longer solutions can value a number of occasions as a lot.
What Biology Does Differently, and What Silicon Copied
Cortical exercise is sparse. Average firing charges are under 1 Hz, and power follows change quite than clock cycles.
Modern AI reached the identical conclusion independently. Kimi K2 activates 32.6 billion of its 1.04 trillion parameters per token, shut to three.1%.
That ratio is falling quick. Mixtral used roughly 28% of its parameters in 2023, whereas DeepSeek-V3 now makes use of 5.5%.
Biology additionally computes at low precision. Nothing inside a neuron resolves to 32 bits.
Chipmakers adopted the identical path. DeepSeek trained a 671-billion-parameter mannequin in eight-bit precision. NVIDIA has since pretrained a 12-billion-parameter mannequin in four-bit.
The third distinction is the biggest and the least copied. Brains maintain reminiscence and computation in the identical bodily place.
Digital machines separate them. Stanford’s Mark Horowitz showed the price of that break up. Fetching an operand from reminiscence can devour a whole bunch of occasions extra power than the arithmetic itself.
The Brain-Shaped Chips That Never Arrived
Hardware constructed explicitly to mimic neurons has struggled. No neuromorphic or analog system has skilled or run a frontier mannequin in manufacturing.
Intel’s Hala Point packs 1.15 billion synthetic neurons throughout 1,152 chips. It stays a analysis prototype put in at Sandia National Laboratories. Mike Davies, director of Intel’s Neuromorphic Computing Lab, talking to The Register in 2024, said:
“We’re not mapping any LLM to Hala Point right now. We don’t know the way to try this.”
The business image is thinner nonetheless. BrainChip is the sector’s flagship listed firm. It reported $700,000 in buyer receipts in opposition to $5.3 million of working outflow final March quarter.
Others have stalled outright. Rain AI, which sought $150 million and failed to boost it, (*20*) a sale in 2025.
Researchers inside the sector describe a round drawback. Catherine Schuman, assistant professor {of electrical} engineering and laptop science on the University of Tennessee, Knoxville, stated:
“The {hardware} corporations are ready for there to be a killer utility, however it’s actually exhausting to grasp how you can construct these functions with out having {hardware} to prototype on.”
More than 20 researchers signed a 2025 consensus paper in Nature. It argued that the sector nonetheless lacks the ecosystem it wants.
Biology’s rules received. The {hardware} constructed to embody them didn’t.
Everyone Is Bidding for the Same Electrons
Efficiency issues now as a result of electrical energy has turn into the binding constraint. The International Energy Agency put international information heart consumption at 485 terawatt-hours in 2025.
AI-focused amenities grew 50% throughout that yr alone. The company expects them to triple by 2030.
Grid entry, quite than chip provide, now gates development. Median time from an interconnection request to business operation exceeds 5 years, in keeping with Lawrence Berkeley National Laboratory.
Microsoft chief government Satya Nadella said in November that his firm holds processors it can not plug in. The scarcity is powered buildings, not silicon. Institutional traders have raised related questions about grid readiness.
Bitcoin miners spent a decade fixing precisely that drawback. They maintain energized websites, signed energy agreements, and interconnection rights that newcomers wait years to safe.
The consequence has turned mining into an energy and infrastructure business. Retrofitting a working website prices roughly $3 million to $4 million per megawatt. Greenfield development runs $10 million to $12 million, VanEck estimates.
Announced deal values are monumental. Public miners have signed AI contracts value greater than $70 billion in mixture.
Delivered capability tells a quieter story. Second-quarter 2026 filings present roughly 750 megawatts really energized throughout the sector.
Core Scientific accounts for about 437 of these megawatts. Galaxy’s Helios campus delivered 133; TeraWulf 102; IREN 50; and Riot 25. Hut 8 has contracted 949 megawatts and energized none.
The pivot has been pricey. Combined quarterly losses at miners MARA and CleanSpark reached $851 million in August.
Most mining capability won’t ever convert. Preliminary Cambridge survey information introduced in July showed that about 10% of miners had allotted energy to AI.
The obstacles are bodily. Mining tolerates interruption, whereas AI tenants demand agency energy, dense cooling, and fiber that distant websites hardly ever have.
Even so, the route of journey is evident. Core Scientific now earns 83% of its income from colocation and simply 13% from mining itself.
Investors have priced that shift in. Miners holding signed leases commerce at far larger multiples of their energized energy. Meanwhile, the next AI bet more and more seems like electrical energy quite than chips.
Why Efficiency Will Not Fix AI Energy Use
Efficiency features have absorbed demand progress up to now. Global information heart compute grew by 550% between 2010 and 2018, whereas power use rose by about 6%.
Then the sample broke. United States information heart consumption climbed from 58 terawatt-hours in 2014 to 176 in 2023.
Evolution optimized underneath a exhausting ceiling. A cranium drawing 200 watts would have killed its proprietor, so effectivity grew to become the one accessible reply.
AI has by no means confronted that ceiling. It has confronted a capital ceiling as a substitute, and capital stretches way more simply than electrical energy does.
That is now altering. The open query is now not whether or not biology is extra environment friendly. It is what AI turns into as soon as energy, quite than cash, decides what will get constructed.
The submit Your Brain Runs on 20 Watts. AI Wants a Power Plant appeared first on BeInCrypto.