The Brain Runs on 20 Watts. Why Doesn't AI?

A chart comparing a brain to a server rack tells you nothing until you can say what task both are doing, what counts as finishing it and what's included in each watt.

A screenshot keeps circulating. It places a human brain, a local AI model running on a home workstation and a refrigerator-sized server rack on the same scale, as if they belonged there together. A follow-up post corrected one of the numbers. Both posts share the same assumption: that stacking a brain against a chip reveals something clean about intelligence. It doesn't, and the reason is specific enough to state as a rule. A brain-versus-AI energy comparison only means something when the task is defined, the system boundary matches and the unit is energy per accepted result, not energy per hour of existing.

The seductive comparison

Picture the chart. A brain sits at one end, running on about 20 watts. A workstation with a local model sits in the middle and pulls noticeably more power. A server rack sits at the far end and draws far more again. Three numbers, one implied conclusion: more watts bought more thinking, and nature got the better deal.

The problem is what each number measures. The brain's wattage covers a resting adult already shaped by decades of eating, sleeping and learning. The rack's wattage might cover training a model from scratch, serving millions of users or sitting half idle. Lining these up on one axis doesn't compare intelligence. It compares incompatible accounting choices and calls the result insight.

Why parameters aren't brain parts

The urge to count doesn't stop at watts. People count parameters the way they once counted neurons, as if scale alone explained thought. The human brain holds roughly 86 billion neurons and on the order of hundreds of trillions of synaptic connections. That's a number describing wiring, not a performance spec.

A parameter is a stored value a model uses to turn input into output. Some systems carry enormous parameter counts yet touch only a fraction per question. DeepSeek-V3 holds 671 billion parameters in total but activates just 37 billion per token, routing each input through a subset of its expert networks. Total parameter count tells you about storage. Active parameter count tells you about the work done on this particular question. Neither tells you whether the answer is any good, and neither maps onto a neuron count in any unit that survives translation.

The 20-watt half-truth

The brain's 20-watt figure gets repeated so often it sounds like the whole story. It covers current resting metabolic power, not the sugar, sleep and years of development behind that resting state, and not the body that carries the brain around and feeds it sensory data. The boundary between "the brain" and "the system that makes the brain useful" is a choice the analyst makes. Nature didn't draw that line.

That's the exact move AI critics accuse the industry of making. Quote a chip's running wattage and skip the energy spent training the model sitting on it. Quote a brain's resting wattage and skip the decades behind it. Both numbers are real. Neither is the full cost.

The strongest objection, and why it survives only partway

A fair critic will push back here. Even granting every accounting problem, the brain still runs broad, real-time sensory and motor control on a power budget smaller than a laptop charger, and AI training costs get paid once, then spread across millions of later uses. Both points are true, and neither rescues the meme chart. Amortized training cost still has to be measured per completed task to mean anything, and the brain's efficiency at broad biological control says nothing about how it stacks up against a model solving one narrow problem. Conceding the brain's real efficiency doesn't restore the chart. It confirms the chart was never comparing the right thing.

What the energy numbers actually include

Training a model, running it afterward and cooling the building around it are three separate costs, and reporting rarely keeps them apart. Deloitte projects global data-center electricity consumption will reach roughly 536 terawatt-hours in 2025, climbing to about 1,065 terawatt-hours by 2030 as generative AI training and inference outgrow every other use. AMD has set a target of a twentyfold increase in rack-level energy efficiency for AI training and inference by 2030, measured from a 2024 baseline. Efficiency is genuinely improving, yet total electricity use is still projected to climb because demand is growing faster than efficiency gains can offset it. A more efficient chip and a hungrier industry aren't a contradiction. They're the same story from two angles.

What this means for buying local AI hardware

Tokens per second measures how fast a chip moves, nothing more. A model that answers instantly and gets it wrong twice before getting it right hasn't saved you time or electricity, whatever the spec sheet says. The question worth asking isn't how many brains fit on your desk. It's how much accepted work the machine finishes, for your actual tasks, per dollar and per watt.

The Framework: Same Boundary, Same Task, Same Unit

Before trusting any brain-versus-chip or chip-versus-chip comparison, run it through three checks.

Same boundary: confirm both systems count the same inputs, whether that's one chip's draw or the full stack including cooling, storage and training.

Same task: define what success looks like before either system runs. "Faster" and "bigger" aren't tasks.

Same unit: measure energy per accepted result, not watts per hour of operation. Divide total energy by the number of outputs a person would actually keep, corrections included.

A chart that skips any of the three isn't measuring intelligence. It's measuring whatever was easiest to put on a slide.

Sources

1. The 20 W sleep-walkers - PMC
2. Power of a Human Brain - The Physics Factbook
3. Lower Bound Estimates for Electrophysiological Power Dissipation in Human Gray Matter | bioRxiv
4. Suzana Herculanohouzel | Vanderbilt Expert | News | Vanderbilt University
5. Billions of neurons, trillions of synapses – UCLA Brain Research Institute (BRI)
6. DeepSeek-V3 Technical Report | alphaXiv
7. Data center sustainability | Deloitte insights
8. Data Center Sustainability