What each substrate pays for a unit of work

The same star, two ways of spending it. Biology runs within about 18x of the thermodynamic floor; a data-centre GPU runs about 250 million times above it.

v2 revised 2026-09-15 The two columns end to end — sources, storage, conversion, currency, consumers. v1 ›

Both columns below start with the same input: sunlight, 1.7 × 10¹⁷ watts of it, arriving at the same planet. Everything that follows is a difference in what gets captured, how it is stored, what converts it, and what spends it.

Read them side by side. Where the two line up, they line up for a reason.

Lifefour billion years, on sunlight
Computingeighty years, on electricity
Where the energy comes from
1.7 × 10¹⁷ WSunlightThe same star, arriving at the same planet. Everything below is a choice about what to do with it.
1.7 × 10¹⁷ WSunlightIdentical input. The two columns diverge only in what they capture and how.
≈ 130 TWPhotosynthesisUnder 1% of incident sunlight is captured globally — but across the whole lit surface, continuously, for free.
20–22%PhotovoltaicsTwenty times more efficient per square metre, over a vanishingly smaller area, and it has to be built.
independent of lightChemosynthesisVent bacteria run on sulphide. Life found a second power source and kept it.
≈ 3.2 TWThermal & nuclear plantsMost electricity still comes from burning something — largely photosynthesis, stored and then dug up.
How it is stored
37 kJ/gBody fatThe densest chemical store biology uses. A person carries weeks of it without noticing.
≈ 0.9 kJ/gLithium-ionAbout forty times less dense than fat. Every constraint on untethered computing starts here.
≈ 17 kJ/gGlycogenThe fast store. Minutes to hours, not weeks — biology keeps two tiers deliberately.
secondsCapacitors & UPSOur fast tier holds seconds, and exists only to survive the gap before a generator starts.
The converter, and the currency
1,000–2,000 per cellMitochondriaThe power supply, replicated per process. A liver cell runs thousands in parallel.
1 per machinePower supply & VRMCentralised, and a single point of failure. Biology does not build these one to a body.
30.5 kJ/molATPOne universal currency, spent by every process in every organism. No conversion, no exchange rate.
≈ 3.2 TWElectricityOur universal currency — and like ATP, almost everything converts into it before being spent.
What spends it
≈ 20 WThe brainA fifth of a person’s entire energy budget, for the thinking. Runs on porridge.
≈ 700 WOne acceleratorThirty-five brains, for one card, doing something narrower.
≈ 100 WBasal life supportWhat it costs merely to stay alive and organised — pumps, gradients, repair, temperature.
≈ 20% of facilityCooling and overheadAt a PUE of 1.2, a fifth of the power never touches computation. Biology spends its overhead the same way.
≈ 25% efficientMovementMuscle converts a quarter of chemical energy into work. The rest is heat, which is also the temperature control.
embodiedManufacturingA server’s build cost is paid once, before it runs. Biology pays for construction continuously, out of the same budget.
continuousGrowth and repairNever finished. An organism is a thing that is always being rebuilt out of current income.
no counterpart — hardware does not repair itself, so replacement is a capital event rather than a metabolic one
What these figures count
counts
Energy densities are per gram of the storage medium itself. Fat at 37 kJ/g is the textbook 9 kcal/g; lithium-ion at ~250 Wh/kg is a current pack-level figure. The 40× ratio is the robust part — cell chemistry moves it by a few percent, not by an order of magnitude.
excludes
Fat's density excludes the oxygen needed to burn it, which the organism does not carry. A battery carries both halves of its reaction. That is a real advantage to the battery and it narrows the gap, though nowhere near closing it.
uncertainty
Mitochondrial counts vary from one to several thousand by cell type. Muscle efficiency is ~20–30% depending on task and measurement. PUE varies from 1.1 in the best facilities to well over 1.5.
the trap
A brain and an accelerator are not doing the same job, so "35 brains per card" is a power comparison and nothing more. It says what each draws, not what each accomplishes.

The floor underneath both columns

There is a hard limit under the cost of computing. Landauer's principle says erasing one bit at temperature T must dissipate at least kT ln 2 — about 2.9 × 10⁻²¹ joules at room temperature. Nothing gets under it. It is the only number on this page that is not an estimate.

Measured against that floor, the two substrates are not close.

Energy per elementary operation, as a multiple of the Landauer limit
18×  vs  250,000,000×
One ATP hydrolysis releases about 5.1 × 10⁻²⁰ J — roughly 18 times the thermodynamic floor. One dense FP16 multiply-add on an H100-class GPU costs about 7.1 × 10⁻¹³ J — roughly 250 million times the floor. Seven orders of magnitude separate them.
What this figure counts
counts
ATP hydrolysis at standard free energy, −30.5 kJ/mol, the value carried in the ai-bio-energetics catalogue as ATP-001. The GPU figure is chip-level: rated board power divided by dense FP16 throughput.
excludes
Cooling, power conversion and the rest of the facility. At a PUE of ~1.2, plus host and memory overhead, the real figure per useful FLOP is two to three times higher. The biological side excludes the cost of making the ATP.
uncertainty
In cellular conditions ATP hydrolysis is nearer −50 to −60 kJ/mol, putting biology at roughly 32× the floor rather than 18×. That changes nothing about the conclusion.
the caveat
A FLOP and an ATP hydrolysis are not the same unit of work. One drives a myosin power stroke or a pump cycle; the other is a multiply-add. This compares the elementary energy quantum of each substrate, not equivalent computations — and no honest version of this comparison can do better.

In aggregate, the ordering reverses

Per operation, biology wins by seven orders of magnitude. In total power the ranking is less flattering to the thing everyone is currently worried about: human bodies draw roughly 17 times more power than every data centre on Earth combined — 820 GW of metabolism against 47 GW of computing — and the biosphere runs on about 2,700 times what the data centres do.

Why this matters

Two facts sit awkwardly together. Per operation, computing is seven orders of magnitude from the floor that biology nearly touches. In aggregate, computing is still small.

The second is why the first is urgent rather than academic. There is enormous headroom to grow into, and the cost per operation decides how far. An agent economy that settles its work in tokens is settling in energy whether it says so or not, and at 2.5 × 10⁸ times the floor there is a great deal of room between what is being paid and what must be.

That is why this site treats settlement as one of four acts rather than an afterthought. The engine that runs the loop has to account for what a run costs, because the cost is nowhere near fixed.

Sources