The Sun Does Not Send Us Energy

Here is a fact that ought to be more disturbing than it is. Over any reasonable stretch of time, Earth radiates back into space almost exactly as much energy as it receives from the sun. The imbalance—the part that is currently heating the planet, and which is an emergency—runs at about nine tenths of a watt per square metre against an incoming flux of two hundred and forty. Better than a 99.6 per cent match.

So energy does not accumulate here. It arrives and it leaves. Which raises a question that sounds naive and is not: if nothing is piling up, what exactly is the sun providing that makes forests and hurricanes and cities possible? You cannot build anything out of a quantity that leaves as fast as it comes in.

The answer is that the sun is not supplying energy. It is supplying order—or, put the way a physicist would, it delivers energy in a low-entropy form and Earth returns it in a high-entropy one, and the difference between those two is the entire budget out of which every ordered thing on this planet, including you, is paid for.

The Photon Accounting

You can see it most clearly by counting photons rather than joules.

Sunlight arrives from a surface at about 5,772 kelvin. Earth radiates from an effective temperature of about 255 kelvin. For thermal radiation, the average energy per photon is proportional to the temperature of the source. So if the energy flowing in and the energy flowing out are equal, but each outgoing photon carries roughly a twenty-second of the energy of an incoming one, then about twenty-two photons must leave for every one that arrives.

⤢ enlarge

Twenty-two ways to distribute the energy instead of one. That is the whole trick, and it is measurable: the planet exports entropy at roughly six hundred trillion watts per kelvin. Most of that is bare radiative degradation—sunlight becoming heat—but a slice of it, somewhere between three and six per cent, is produced by the irreversible processes of the climate system itself: winds, currents, the water cycle, the whole restless machinery.

This is the correct way to understand a living planet. It is not a system that accumulates energy. It is a gradient, standing between a very hot source and a very cold sink, and everything interesting that happens here happens in the flow between them.

What the Gradient Buys, in Watts

Earth absorbs about 122,000 terawatts of sunlight after reflection. That is the budget. Here is what it is spent on.

Global photosynthesis fixes something like 105 billion tonnes of carbon a year, which works out at roughly 130 terawatts of chemical free energy. Against 122,000 terawatts of absorbed sunlight, that is about one tenth of one per cent. The entire biosphere—every forest, every plankton bloom, four billion years of it—captures a thousandth of what arrives.

The atmosphere is far greedier: the planet’s total generation of free energy runs to about 6,300 terawatts, almost all of it geophysical—the winds, the lifting of water, the ocean circulation. Life is a rounding error on the weather.

And humanity? Global primary energy consumption is around 590 exajoules a year, which as a continuous power is about 18.7 terawatts. Set against absorbed sunlight that is 0.015 per cent—a number so small it is regularly used to argue that human energy use is trivially sustainable.

But that is the wrong denominator, and this is the most important comparison in the essay. Set 18.7 terawatts against the 130 terawatts of global photosynthesis and you get about fourteen per cent. One species is now running an energy flow equal to roughly a seventh of what all photosynthetic life on Earth manages. That is the ratio that should be quoted, and it almost never is.

Entropy and Information Are Not the Same Thing

Now to the second half, and to a confusion that has to be cleared before anything else can be said honestly.

“S = k log W”the inscription on Boltzmann’s grave, Vienna

Boltzmann’s entropy counts the number of microscopic arrangements consistent with what you can see macroscopically. Shannon’s information entropy, written down eighty years later, has an almost identical form: a sum over probabilities times their logarithms. The resemblance is exact and it has generated an enormous amount of loose talk.

Here is the careful statement. The two coincide only when the probability distribution in question is over the physical microstates of the system you are talking about. When that condition holds, one bit corresponds to about 9.6 × 10⁻²⁴ joules per kelvin. When it does not, the two quantities are simply unrelated numbers that happen to share an equation.

The scale of the gap is worth feeling. Take a one-terabyte drive. Its thermodynamic entropy—over the positions and momenta of every atom in the thing—is on the order of a thousand joules per kelvin. The Shannon entropy of the files on it, converted to the same units, is about ten billionths of a joule per kelvin. Fourteen orders of magnitude apart. The bits you care about are a vanishingly thin layer of structure on top of an ocean of microscopic degrees of freedom that do not care what you saved.

So no: information is not entropy, deleting your photographs does not measurably warm the room, and the fact that both are written with a logarithm is a mathematical fact about counting, not a bridge between physics and meaning.

But There Is a Real Bridge, and It Is Narrow

The genuine connection was found by Rolf Landauer in 1961 and it concerns one specific operation. Erasing a bit—taking a system that could be in either of two states and forcing it into one, discarding which it was—must dissipate at least k_B T ln 2 of heat into the surroundings.

At room temperature that is 2.87 × 10⁻²¹ joules, or about eighteen thousandths of an electron-volt. It is a preposterously small quantity and its importance is entirely conceptual: it is the point at which a fact about logic becomes a fact about heat.

Two things about it are constantly got wrong. First, it is erasure that costs, not computation. A logically reversible operation—one you could run backwards to recover its inputs—has no thermodynamic floor at all. The bill is charged for throwing information away. Second, the bound is attained only in the limit of infinite slowness; any erasure performed at a finite speed costs strictly more.

It has been measured. The first confirmation, in 2012, watched a single colloidal particle pushed between two wells of an optical trap and found the dissipated heat saturating at the bound. By 2021 an experiment using an underdamped micromechanical oscillator was hitting the limit to within one per cent in protocols lasting a tenth of a second.

This is also what finally killed Maxwell’s demon, and the resolution is not the one usually taught. Szilard and Brillouin thought the demon was defeated by the cost of measurement. Charles Bennett showed that measurement can in principle be done reversibly, for free—and that the irreducible cost is clearing the demon’s memory so it can start the next cycle. The demon does not fail because looking is expensive. It fails because forgetting is.

Real demons have since been built. In 2010 a Brownian particle was ratcheted up a spiral potential by feedback alone, converting information into free energy. In 2015 an autonomous single-electron device on a chip acted as an information-powered refrigerator—cooling the system it was watching while the demon circuitry itself heated up, the bill arriving exactly where the theory says it should.

You will also see occasional headlines about erasure below the Landauer bound. They are real experiments and none of them breaks anything: each pays somewhere else—a bath held out of equilibrium, a squeezed memory, an asymmetric potential. The most recent, from early this year, went more than twenty per cent below the bound, and its own authors identify the mechanism as an embedded Maxwell demon exploiting information hidden in the apparatus’s hysteresis. The generalised bound holds. It always has.

How Far Is Our Computing From the Floor?

Nowhere near it, and the size of the gap is the single most clarifying number in this whole subject.

⤢ enlarge

A current high-end accelerator delivers around 0.7 picojoules per low-precision arithmetic operation. Since one such operation involves on the order of a thousand elementary bit manipulations, the cost per bit operation lands somewhere around a hundred thousand times the Landauer bound; per whole operation the factor is nearer a hundred million. A recent review compresses it to: real devices use about a million times more energy per operation than thermodynamics requires.

Now do the arithmetic that follows. Even taking the most aggressive estimate of how many bits the world produces annually and imagining every one of them erased, the total Landauer cost would be a few joules per year. Not gigawatts. Joules. Enough to lift an apple off a table.

The thermodynamic limit is not a constraint on anything we are doing or will do this century. Whatever is making data centres expensive, it is not the second law. It is gate capacitance, wire resistance, leakage current, the redundancy needed for reliable switching, and the fact that heat has to physically leave a building.

What has changed, and is genuinely worrying, is the rate of improvement. Koomey’s law held that computations per joule doubled every 1.57 years from the 1940s to about 2000—a hundredfold gain per decade. After 2000, with the end of Dennard scaling, the doubling time stretched to about two and a half years. An independent re-examination in 2024, looking at high-performance systems from 2008 to 2023, found efficiency doubling every 2.29 years while raw performance doubled every 1.85. Efficiency is now improving more slowly than capability, which is a new and structurally uncomfortable situation: every generation is more capable and each unit of capability costs relatively more.

The Planet’s Information Budget, With the Honest Caveats

How much information does the technosphere hold? The honest answer begins with an admission: we measured this properly exactly once.

Martin Hilbert and Priscila López spent years constructing an inventory across sixty technology categories from 1986 to 2007, in optimally compressed bytes. Their results are the only rigorous historical figures anyone has. Total storage: 2.6 exabytes in 1986, 295 exabytes in 2007. And the transition buried inside that: the digital share of all stored information went from 0.8 per cent in 1986 to 25 per cent in 2000 to 94 per cent in 2007. The changeover from analogue to digital as humanity’s memory happened inside a single decade, and mostly inside seven years of it.

After 2007, the study stops, and everything you have ever read about zettabytes is an industry forecast of a different quantity. The widely quoted figures—around 181 zettabytes for 2025—measure data *created, captured, copied and consumed* in a year. Most of that is transient: video streams nobody stores, sensor telemetry discarded on arrival, replicas of replicas. The stock is smaller than the flow by roughly two orders of magnitude, and no methodologically transparent number for the stock exists after 2007.

⤢ enlarge

For scale, put the biosphere on the same axis. A 2015 estimate put the total DNA in living cells at about 5 × 10³⁷ base pairs—roughly 10³⁸ bits at two bits a pair, held in something like 5 × 10³⁰ cells. That is around fourteen orders of magnitude more than everything humanity generates in a year, and unlike our data it is all in active use, all being copied, and all being error-corrected.

There is a companion figure I find genuinely arresting. Modelling chemical reactions as elementary logical operations, Seth Lloyd and a collaborator estimated that all life on Earth performs somewhere between 10³³ and 10³⁵ such operations per second—and that a single human being performs around 10²⁰ to 10²², which they argue is comparable to the total information-processing capacity of every computer, phone and server on the planet combined. At most a few per cent of that occurs in neural firing. The overwhelming majority is metabolism: your cells, doing chemistry.

What the Machines Actually Cost

Since this is where the public argument sits, the numbers deserve stating plainly. Data centres consumed about 415 terawatt-hours of electricity in 2024 and around 485 in 2025—a little over 1.5 per cent of global electricity. The International Energy Agency’s base case has them reaching roughly 950 terawatt-hours by 2030, close to 3 per cent, with the AI share of that rising from somewhere between 5 and 15 per cent today to perhaps 35 to 50 per cent.

Three per cent is a real number and worth arguing about. It is not the ten or twenty per cent that circulates, and using the inflated figure makes the genuine local problems—grid queues, water, transmission—easier for their proponents to dismiss.

The IEA also supplies the counterpoint, and it is the more interesting half. Energy per AI task has been falling by at least an order of magnitude annually in recent years. Total demand is rising anyway, because usage is growing faster than efficiency. That is a Jevons dynamic, not a physical limit—which means it is a question about economics and policy rather than about thermodynamics, and it will not be solved by a better chip.

One more piece of accounting, because it puts the technosphere in its place. Human-made mass overtook all living biomass around 2020, give or take six years, and it has been doubling roughly every twenty years. Roughly 1.1 trillion tonnes of life; rather more than that of concrete, aggregate, asphalt, brick, metal and plastic. Whatever we are building, it now outweighs everything that grew.

A Catastrophe That Is Not One

You may have encountered the claim that information has mass and that we are heading for an “information catastrophe.” It deserves reporting precisely and then deflating, because it is the sort of idea that survives on not being examined.

The proposal is that a stored bit possesses a rest mass of about 3.2 × 10⁻³⁸ kilograms, and that at twenty per cent annual growth in bit production, digital content would outnumber Earth’s atoms in about three hundred and fifty years and exceed half the planet’s mass in five hundred.

The mass figure is exactly the Landauer energy divided by the speed of light squared. Which means the proposal equates the heat dissipated when a bit is erased with the rest mass of a bit while it is stored. Those are different quantities; there is no reason a storage state should weigh what its destruction radiates. A rebuttal published in a physics journal argues the identification is unsound in principle, the proposed experiment—weigh a full drive against an empty one—has produced no confirmed result, and the catastrophe follows entirely from compounding twenty per cent for two and a half centuries. Compound any exponential for two hundred and fifty years and it eats the planet. That is a fact about exponentials, not about information.

Where Quantum Mechanics Comes In

It is not decoration. Quantum mechanics changes what information *is* in ways that matter for everything above.

Start with a fact that has no classical analogue. Two particles can be in a maximally entangled pure state, in which the pair as a whole has exactly zero entropy—there is nothing you do not know about it—while each particle taken alone has the maximum possible entropy. The information is not in the parts. It is in the relation, and no examination of either half will find it.

Which leads to the area law, one of the loveliest results in many-body physics. For the ground state of a system whose interactions are local, the entanglement entropy of a region scales not with its volume but with the area of its boundary. The information binding a chunk of matter to its surroundings lives on the surface. This is why tensor-network methods can simulate quantum systems at all, and it is the many-body echo of the fact that a black hole’s entropy is proportional to the area of its horizon rather than the volume inside.

Quantum information is also, in principle, indestructible. The no-cloning theorem says an unknown state cannot be copied; the no-deleting theorem, its mirror, says that given two copies you cannot cleanly destroy one. Both follow from linearity, and both express the deeper point that Schrödinger evolution is unitary and therefore reversible. Information is conserved. That is precisely why Hawking’s 1974 calculation—a pure quantum state collapsing into a black hole and emerging as thermal radiation carrying no memory of it—was a paradox rather than a curiosity.

The story since 2019 is genuinely exciting and routinely overstated. Using quantum extremal surfaces and gravitational path integrals that include replica wormholes, several groups derived the Page curve—the turnover in radiation entropy that unitarity demands—from gravity itself, in controlled low-dimensional models. That is a real achievement.

But it is not a solved problem, and papers published this year say so. Whether the “island” prescription describes genuine information recovery in four-dimensional gravity with dynamical, massless gravitons—or is an artefact of coupling the black hole to an external reservoir and then restricting which observables you allow—remains actively disputed as of early 2026. Anyone who tells you the information paradox was settled in 2019 is describing a model, not the universe.

The last piece is quantum Darwinism, and it answers a question the rest of this essay assumes. Why is there a classical world at all—one where facts are objective and many people can check the same thing without disturbing it? Wojciech Zurek’s answer is that the environment acts as a broadcast channel. A system’s pointer states—the ones robust against decoherence—get imprinted redundantly into many independent fragments of the surroundings, so that thousands of observers can each read a separate copy and agree. Superpositions do not proliferate, so nobody sees them. Objectivity is redundancy, and it has been tested experimentally on nitrogen-vacancy centres, photonic simulators and quantum hardware since 2018.

The Room That Is Left

Which brings me to the number I find hardest to stop thinking about.

The entropy of the observable universe has been carefully totalled. It comes to about 3 × 10¹⁰⁴ in units of Boltzmann’s constant, and it is overwhelmingly dominated by supermassive black holes—so much so that every star, every photon of the microwave background, every neutrino and every atom of gas together contribute less than a millionth of a per cent. On the entropy ledger, the visible universe is black holes and rounding errors.

Now compare that with the maximum the universe is permitted, given by the area of the cosmic event horizon: about 2.6 × 10¹²².

⤢ enlarge

Eighteen orders of magnitude of unused capacity. The universe has run for nearly fourteen billion years and used up perhaps a million-billion-billionth of its entropy budget. It is not, as the Victorians feared, drifting toward an imminent heat death. It is barely started.

And that headroom is not a curiosity. It is the precondition for everything. Ordered structures—stars, galaxies, cells, brains, sentences—can exist only where there is somewhere for their disorder to go. A universe near equilibrium permits nothing. This one has room to spare, and the whole history of complexity is a history of local structures exploiting that gap.

The same reasoning sets the ceilings on computation. The holographic bound allows about one bit per four Planck areas—roughly 10⁶⁹ bits per square metre of enclosing surface. Seth Lloyd worked out what a kilogram of matter in a litre of volume could do if every degree of freedom were used: about 5 × 10⁵⁰ operations per second on 10³¹ bits. A current supercomputer is around thirty orders of magnitude short on speed. We are not near any limit that physics imposes.

What This Adds Up To

Every ordered thing is paid for by exporting disorder somewhere else. A crystal forming, a cell dividing, a sentence being written, a data centre running—each is a local decrease in entropy funded by a larger increase outside. Earth has been running that trade at six hundred trillion watts per kelvin for four billion years, and the trade is not close to exhausting either the sun or the sky.

What is new is not that we use energy. We use fifteen thousandths of one per cent of what arrives, and the thermodynamic floor beneath our computing is a factor of a million below where we operate. We are nowhere near any limit that nature has set.

What is new is the rate at which one species has begun organising matter—an energy throughput around a seventh of all photosynthesis, an artefact mass now exceeding all life and doubling every twenty years, and an information stock whose growth rate outran every previous technology by an order of magnitude in the two decades we bothered to measure it.

And there is a final asymmetry worth ending on. The constraints that will actually bind in the next few decades are not the ones in this essay. They are grids, water, transmission queues, planning permission and politics. Physics has left us an almost unimaginable amount of room. The ceilings we are hitting are the ones we built ourselves, which is a more tractable problem than the second law—and a considerably more embarrassing one.