Your Brain Runs on 20 Watts. AI Wants a Power Plant

The human brain runs on roughly 20 watts. The world's fastest supercomputer, LineShine in Shenzhen, draws 42.2 million watts. That gap has become the internet's favorite argument about AI energy use, and most of it is wrong.

The comparison itself holds up. However, the numbers circulating on social media trace back to a single paper. The most striking one has been misattributed for three years.

China just topped the global supercomputer ranking for the first time since 2017.

LineShine. Shenzhen. 2.198 exaflops. 2 quintillion calculations per second. 20% faster than the US's El Capitan.

built entirely on domestic Chinese CPUs. no Nvidia. no US chips. no export controls... https://t.co/Tuuk6svl6l pic.twitter.com/fwFKWN0Mr4 --- IT Guy (@T3chFalcon) June 24, 2026

AI Energy Use: What 20 Watts Actually Buys {#h-ai-energy-use-what-20-watts-actually-buys}

The 20-watt figure rests on decades of metabolic measurement. The brain accounts for about 2% of body weight and roughly 20% of resting oxygen consumption.

Neuron counts are shakier than they appear. The widely quoted 86 billion rests on four male brains and is currently under dispute in the journal Brain.

Viral posts often use 12 watts rather than 20. That figure appears in a 2023 paper in Frontiers in Artificial Intelligence, stated without any citation at all.

The same paper produced the number everyone shares. Its authors estimated that digitally recreating a human brain would draw 2.7 billion watts.

That estimate came from extrapolating a 10-million-neuron simulation to mouse scale, then multiplying by a thousand.

The paper also states that the simulation ran about 30,000 times slower than biology. Social posts drop that detail. Secondary sources then credit the figure to the Blue Brain Project, which never published it.

Reliable numbers do exist elsewhere. Epoch AI estimated a typical ChatGPT query at 0.3 watt-hours in early 2025. A peer-reviewed study in Joule later landed on 0.31.

Two independent methods agreeing that closely is unusual. However, the figure changes sharply with workload, and reasoning models that produce longer answers can cost several times as much.

What Biology Does Differently, and What Silicon Copied {#h-what-biology-does-differently-and-what-silicon-copied}

Cortical activity is sparse. Average firing rates are below 1 Hz, and energy follows change rather than clock cycles.

Modern AI reached the same conclusion independently. Kimi K2 activates 32.6 billion of its 1.04 trillion parameters per token, close to 3.1%.

That ratio is falling fast. Mixtral used roughly 28% of its parameters in 2023, while DeepSeek-V3 now uses 5.5%.

Biology also computes at low precision. Nothing inside a neuron resolves to 32 bits.

Chipmakers followed the same path. DeepSeek trained a 671-billion-parameter model in eight-bit precision. NVIDIA has since pretrained a 12-billion-parameter model in four-bit.

The third difference is the largest and the least copied. Brains hold memory and computation in the same physical place.

Digital machines separate them. Stanford's Mark Horowitz showed the cost of that split. Fetching an operand from memory can consume hundreds of times more energy than the arithmetic itself.

The Brain-Shaped Chips That Never Arrived {#h-the-brain-shaped-chips-that-never-arrived}

Hardware built explicitly to imitate neurons has struggled. No neuromorphic or analog system has trained or run a frontier model in production.

Intel's Hala Point packs 1.15 billion artificial neurons across 1,152 chips. It remains a research prototype installed at Sandia National Laboratories. Mike Davies, director of Intel's Neuromorphic Computing Lab, speaking to The Register in 2024, said:

"We're not mapping any LLM to Hala Point at this time. We don't know how to do that."

The commercial picture is thinner still. BrainChip is the sector's flagship listed company. It reported $700,000 in customer receipts against $5.3 million of operating outflow last March quarter.

Others have stalled outright. Rain AI, which sought $150 million and failed to raise it, explored a sale in 2025.

Researchers inside the field describe a circular problem. Catherine Schuman, assistant professor of electrical engineering and computer science at the University of Tennessee, Knoxville, stated:

"The hardware companies are waiting for there to be a killer application, but it's really hard to understand how to build those applications without having hardware to prototype on."

More than 20 researchers signed a 2025 consensus paper in Nature. It argued that the field still lacks the ecosystem it needs.

Biology's principles won. The hardware built to embody them did not.

Everyone Is Bidding for the Same Electrons {#h-everyone-is-bidding-for-the-same-electrons}

Efficiency matters now because electricity has become the binding constraint. The International Energy Agency put global data center consumption at 485 terawatt-hours in 2025.

AI-focused facilities grew 50% during that year alone. The agency expects them to triple by 2030.

Grid access, rather than chip supply, now gates construction. Median time from an interconnection request to commercial operation exceeds five years, according to Lawrence Berkeley National Laboratory.

Microsoft chief executive Satya Nadella said in November that his company holds processors it cannot plug in. The shortage is powered buildings, not silicon. Institutional investors have raised similar questions about grid readiness.

Bitcoin miners spent a decade solving exactly that problem. They hold energized sites, signed power agreements, and interconnection rights that newcomers wait years to secure.

The result has turned mining into an energy and infrastructure business. Retrofitting a working site costs roughly $3 million to $4 million per megawatt. Greenfield construction runs $10 million to $12 million, VanEck estimates.

Announced deal values are enormous. Public miners have signed AI contracts worth more than $70 billion in aggregate.

Delivered capacity tells a quieter story. Second-quarter 2026 filings show roughly 750 megawatts actually energized across the sector.

Core Scientific accounts for about 437 of those 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 costly. Combined quarterly losses at miners MARA and CleanSpark reached $851 million in August.

Most mining capacity will never convert. Preliminary Cambridge survey data presented in July showed that about 10% of miners had allocated power to AI.

The obstacles are physical. Mining tolerates interruption, whereas AI tenants demand firm power, dense cooling, and fiber that remote sites rarely have.

Even so, the direction of travel is clear. Core Scientific now earns 83% of its revenue from colocation and just 13% from mining itself.

Investors have priced that shift in. Miners holding signed leases trade at far higher multiples of their energized power. Meanwhile, the next AI bet increasingly looks like electricity rather than chips.

Why Efficiency Will Not Fix AI Energy Use {#h-why-efficiency-will-not-fix-ai-energy-use}

Efficiency gains have absorbed demand growth in the past. Global data center compute grew by 550% between 2010 and 2018, while energy use rose by about 6%.

Then the pattern broke. United States data center consumption climbed from 58 terawatt-hours in 2014 to 176 in 2023.

Evolution optimized under a hard ceiling. A skull drawing 200 watts would have killed its owner, so efficiency became the only available answer.

AI has never faced that ceiling. It has faced a capital ceiling instead, and capital stretches far more easily than electricity does.

That is now changing. The open question is no longer whether biology is more efficient. It is what AI becomes once power, rather than money, decides what gets built.