Most of the people a nuclear war kills are not killed by the weapons. They starve, some years later, a long way from anywhere a weapon fell. This is how that happens, one link at a time.
Nuclear weapons burst over cities set the cities on fire. A city fire of a certain intensity does something an ordinary fire does not: it builds its own weather, and the column of smoke it raises punches through the layer of the atmosphere where rain happens, into the layer above, where nothing washes anything out.
Soot up there is black. It absorbs sunlight, warms, and rises further under its own heating, which puts it higher still and makes it last longer. Under it the ground gets less light and turns cold, and stays cold for years, because there is no mechanism to bring the soot down quickly.
Crops need warmth, light, water and a season without frost, and a shortage of any one of them is enough. The world holds a few months of grain and no more. So the harvests fail everywhere at once, the stores run out, and the famine that follows is not confined to the countries that fought.
The physics of the middle of this chain — soot in the stratosphere makes it cold and dark, and cold and dark ruins harvests — is not seriously disputed by anybody. Groups that disagree about everything else agree that if the soot gets up there, the rest follows.
What is disputed is the first link: how much soot a burning city actually puts above the weather. That single quantity decides everything downstream, it has been measured properly for no modern city, and the published estimates for it differ by more than a factor of a hundred. The section on the fires says why, and the fuller argument is set out in the soot question.
The part that is already imagined, and the smallest part of the answer.
DocumentedA weapon over a city kills by blast, by fire and by prompt radiation, in that order of reach. For a hundred-kilotonne weapon burst at the height that does the most damage, most buildings collapse out to about three kilometres, exposed skin burns out to about four and a half, and windows break out to ten. Prompt radiation reaches barely two — above about two kilotonnes it is entirely inside the radius where the blast has already arrived, which is why it hardly appears in modern casualty estimates.
DocumentedThe published totals for the war itself. A hundred weapons of fifteen kilotonnes between India and Pakistan: about 27 million dead in the first days. Four thousand four hundred weapons of a hundred kilotonnes across the northern hemisphere: about 360 million. Those are enormous figures and they are not the subject of this document, because in both cases the number who die afterwards is roughly ten to fifteen times larger.
The pointEverything that follows starts from the fires and not from the deaths. The chain runs through what burns, not through who dies — though in practice the two are nearly the same thing, because the fuel that makes the smoke is the buildings people live in. On the published assumptions the arithmetic works out at about 154 kilogrammes of stratospheric soot for every person inside the fire. That proportionality is worth holding on to: it means no plan can trade one against the other. A war cannot be made cleaner by burning the same cities differently.
The link the whole chain hangs from, and the one nobody has measured properly.
DocumentedA fire big enough and hot enough stops behaving like a fire and starts behaving like a weather system. Many separate fires, lit at once across several square kilometres, draw in air from all round the burning area. That inrush feeds the fire, which burns hotter, which draws harder. Inward winds of hurricane strength are recorded at Hamburg in 1943, at Dresden in 1945 and at Hiroshima. The single column of rising air above such a fire is what carries the smoke up, and an ordinary fire has nothing like it.
DocumentedWhether that happens turns on one number: how much burnable material stands on each square metre of ground. The handbook figure for a firestorm is about four grammes per square centimetre — forty kilogrammes of wood, furnishings, paper and plastics for every square metre, averaged across the whole burning area, streets and parks included. Below it the fires stay separate and their smoke stays low. Above it they join up.
ContestedAnd that number is not known for a modern city. Hiroshima, the one place the world has actually calibrated against, is published at 3.9, at 10 and at 16 grammes per square centimetre by three different authorities. An American survey of the 1990s gave ordinary urban land 1.4 to 2.1. The most-cited nuclear-winter papers assume 12 to 95 for Indian and Pakistani targets. A study that found no significant cooling used 0.14 — a figure its critics obtained the underlying map for and identified as suburban Atlanta, golf course included.
| Fuel loading | What it is | Reaches the stratosphere | Soot from the large war |
|---|---|---|---|
| 0.14 g/cm² | Suburban lawns and low houses | nothing | 0 Tg |
| 1 g/cm² | Ordinary American urban land | 2% | 1.5 Tg |
| 4 g/cm² | The handbook’s firestorm threshold | 15% | 46 Tg |
| 10 g/cm² | A dense city centre | 50% | 381 Tg |
| 16 g/cm² | The upper Hiroshima estimate | 70% | 854 Tg |
Across the range of values that serious people have published for the same quantity, the answer runs from nothing whatever happens to five times the worst case anyone models. This is not a weakness of one model; it is the actual state of the field, and the National Academies said so in 2025 when it reviewed the whole argument and declined to pick a winner.
It is also why the model at the years after puts the fuel loading on a slider rather than burying it in a constant. The slider is not a toy. It is the disagreement, made movable.
Two layers of atmosphere, and why it matters enormously which one the smoke ends up in.
DocumentedThe lower atmosphere has weather in it, and weather cleans. Everything below about twelve kilometres — the troposphere — overturns, forms cloud and rains. Smoke that stops there is washed out in days to weeks. This is why ordinary wars, ordinary volcanic ash and ordinary forest fires do not cool the planet for years: whatever they put up comes down again almost at once.
DocumentedAbove that lies the stratosphere, which is stable and dry, and does not rain. It is stratified — warm above, cool below — so it does not overturn, and there is no mechanism to scrub it. Anything lifted into it comes down only by settling, slowly, over years. Volcanic sulphate put there takes about a year to clear. Soot takes longer, because of what happens next.
DocumentedSoot lifts itself. It is black, so it absorbs sunlight and warms the air around it, and warm air rises. A layer of soot that arrives at fifteen kilometres does not stay at fifteen: it climbs. This is not a theory awaiting a test. Smoke from the British Columbia fires of 2017 was watched rising from twelve to twenty-three kilometres over two months, and the Australian fires of 2019 lofted smoke from about fifteen kilometres to thirty-four in some forty days. The higher it goes, the longer it lasts — the same observations found smoke below eighteen kilometres clearing in four or five months and smoke above nineteen taking closer to a year.
DocumentedSo the residence time is years, and counter-intuitively the smaller injection lasts longer. Five teragrams of soot takes about six years to fall to a third of itself; a hundred and fifty teragrams takes four and a half. The reason is that a thick layer shades its own underside: only the top of it is heated enough to keep climbing, so the rest never gets the lift.
ContestedWhether the plume reaches the stratosphere at all is where the argument lives. The case for “no” rests on a model of a dry plume, and dry plumes do not rise far: a study of 2022 found a dry column needs a temperature excess of sixty degrees or more to reach the tropopause, and firestorms deliver less than half that. The same study re-ran that dry case with realistic humidity — letting the water vapour in the rising air condense and release its heat — and the plume went from topping out at five kilometres to reaching twelve. Latent heat, in other words, is not a refinement to the calculation. It is the calculation.
What a layer of soot overhead actually does to the ground beneath it.
DocumentedThe dark comes first and matters more than the cold. Soot is not a blanket; it is a shade. It intercepts sunlight on its way down, which means less energy arriving at the ground, which means cooling — but it also means less light for plants, and less heating of the ocean surface, and therefore less evaporation, and therefore less rain. Rainfall falls further, in percentage terms, than temperature does.
| 100 weapons · 5 Tg | 4,400 weapons · 150 Tg | |||||
|---|---|---|---|---|---|---|
| When | Sunlight | Rain | Over land | Sunlight | Rain | Over land |
| First month | 86% | 92% | −1.4 K | 26% | 58% | −4.6 K |
| Six months | 92% | 95% | −1.4 K | 21% | 55% | −8.8 K |
| Year two | 89% | 94% | −1.6 K | 7% | 47% | −11.6 K |
| Year three | 91% | 95% | −1.3 K | 9% | 48% | −12.4 K |
| Year six | 95% | 97% | −0.8 K | 28% | 59% | −10.5 K |
| Year eleven | 98% | 99% | −0.4 K | 66% | 80% | −5.3 K |
| Year sixteen | 99% | 99% | −0.2 K | 87% | 92% | −2.1 K |
Worth pausing onSeven per cent of normal sunlight, in the second year of the large war. That is not dusk. It is closer to a heavily overcast day that does not end, over the whole northern hemisphere, for a year. And the cooling figure — twelve degrees below normal, averaged over land and over the whole year — understates what it does to farming, because an average conceals the frosts. What ruins a crop is not a cool summer. It is one night below freezing in the wrong week.
DocumentedThe small war is not a small effect. A hundred weapons of fifteen kilotonnes — a fraction of one per cent of the world’s arsenals, weapons of the size dropped in 1945 — cools the planet more than any year of the last thousand, and holds it there for the better part of a decade. Frosts arrive earlier and leave later: the published figure is a growing season shortened by ten to forty days a year for five years, everywhere.
DocumentedThe ultraviolet, which arrives later and points the other way. Soot in the stratosphere also warms it, and a warmer stratosphere destroys ozone faster. At the large injection the ozone column falls to about a quarter of normal by the third year and takes fifteen years to recover. But the soot is also blocking the sunlight the ozone would have been filtering, so for the first two years the ultraviolet at the ground is lower than normal, not higher. It crosses back at about year three and peaks around year eight, at levels not recorded anywhere on Earth today. In the model those years read: an index of 6 in the first year, 8 in the second, 25 in the third and 30 in the fourth. The cold comes first; the burning comes later, to whoever is still there.
Why a crop does not average its conditions, and what that means for a bad year everywhere at once.
DocumentedA crop is limited by whichever of its needs is scarcest, not by the sum of them. This is an old idea in agronomy — Liebig’s law of the minimum — and it is the reason a nuclear winter is worse for farming than the temperature figures suggest. Warmth cannot make up for missing light. Rain cannot make up for a frost. A field that has ninety per cent of its usual light, seventy per cent of its usual rain and one killing frost in August yields what the frost allows, which may be nothing at all.
DocumentedFour things are needed and all four are hit. Heat, counted as the accumulated warmth over the season, which the cooling takes away. Light, which the soot takes away directly. Water, which falls with the rainfall. And a season without frost, which shortens from both ends. In the temperate grain belts — the American Midwest, the Ukrainian and Russian black earth, the north Chinese plain — all four move the wrong way at once, and those belts are where the world’s exportable surplus is grown.
| Year after the war | 100 weapons · 5 Tg | 4,400 weapons · 150 Tg |
|---|---|---|
| One | 95% | 39% |
| Two | 93% | 26% |
| Three | 95% | 34% |
| Four | 96% | 37% |
| Six | 97% | 46% |
| Eight | 98% | 58% |
Worth pausing onA seven per cent shortfall is not a small thing. The five-teragram case takes about seven per cent off the world’s harvest for a few years, and that reads mild. It is not: the world’s grain trade is a thin surplus on top of what growers eat themselves, and a seven per cent fall in production is a much larger fall in the amount available to buy. The countries that lose most are not the countries that fought. They are the ones that import.
Four buffers stand between a failed harvest and a famine. They are smaller than they look.
DocumentedStored grain. The world holds a few months of consumption, not years, and it is held where it was grown rather than where it will be needed. Stocks cover the first year of a bad harvest and are gone in the second. Every published run finds the worst year is the second, not the first, and this is why.
DocumentedGrain fed to animals. A large share of the world’s crop calories is eaten by livestock, and feeding them to people instead recovers a great deal. This is the single largest buffer in the system and the published work models it explicitly: divert half of it and the five-teragram case is survivable; divert all of it, cut household waste to nothing and distribute the result equitably and the world gets through forty-seven teragrams. Nothing at all gets it through a hundred and fifty.
DocumentedWaste. Roughly a fifth of household food is thrown away. Halving that covers the sixteen-teragram case; eliminating it covers twenty-seven. These are not policies anyone has ever executed, and they are counted here as the outer bound of what the system could do if it behaved perfectly.
DocumentedTrade — and this is the one that surprises people. At the small end, trade is decisive: at five teragrams, a world that keeps trading loses about seventeen million people to hunger and a world that closes its borders loses four hundred and nine million. At the large end it is nearly irrelevant: at a hundred and fifty teragrams the figures with and without trade are five and a half billion either way. Trade moves a surplus from where it is to where it is needed. When there is no surplus anywhere, there is nothing to move.
The number most often quoted — five billion for a war between the large arsenals — is from Xia and thirteen co-authors in Nature Food, 2022. It carries four conditions that are almost always dropped when it is repeated. It is the variant with no international trade. Survivors are fed at 1,911 calories a day, which is a survival ration and not a diet. Half the grain that would have gone to animals goes to people instead. And it is calculated on the world population of 2010, which was 6.7 billion.
Most importantly: the paper’s own term is “people without food”, not deaths. It is the number the food system cannot keep alive at the worst of it. What happens to those people is not the output of a demographic model, because nobody has built one for this. It is left as the plain statement of a shortfall, and that is how the model at the years after states it too.
A war between the large arsenals, as the chain unfolds. Every figure is from the model at the years after.
| When | What is happening |
|---|---|
| Hour one | The weapons fall. About 360 million people are killed by blast, fire and prompt radiation. Some 380,000 square kilometres of city are burning — an area the size of Germany. |
| Day one to three | The fires organise into firestorms and their columns punch into the stratosphere. This is the whole of the argument, and it is over within seventy-two hours. Everything after it is consequence. |
| Month one | A quarter of normal sunlight reaches the ground. Land temperatures are four and a half degrees below normal and falling. The soot is still spreading and still climbing. |
| Months two to six | The layer closes over the northern hemisphere. Sunlight falls to a fifth of normal, rainfall to just over half. The growing season that year is largely lost in the temperate belts. Stored grain covers what is missing. |
| Year two | The worst of it. Sunlight at seven per cent of normal — a permanent overcast. Land twelve degrees below normal. The harvest is a quarter of a normal year. The stores are empty. This is the year the famine figure belongs to. |
| Years three to five | The cold is at its deepest, around twelve degrees, and the sunlight begins to return. Harvests recover to a third of normal and stop there. The ozone column is at a quarter of normal, and the ultraviolet at the ground crosses back above where it started. |
| Years six to ten | The soot thins out; by year six about a third of it is gone and sunlight is back to a quarter. The ultraviolet peaks around year eight at an index of thirty, which is higher than anywhere on Earth today. Harvests reach half of normal. |
| Years eleven to sixteen | Two thirds of the sunlight is back at year eleven, and nearly all of it at sixteen. Land temperatures are still two degrees below normal in the sixteenth year — further from normal than the whole of the warming of the twentieth century, in the other direction, a decade and a half after the war ended. |
The shape of itNothing in this sequence is fast. The weapons take an afternoon and the fires take three days; everything after that is measured in years, and the worst of it arrives when the war has been over for eighteen months. That is the sense in which the horror is unfolding rather than sudden, and it is the reason a model of a nuclear war that stops at the last detonation is answering a different question from the one worth asking.
Stated plainly, because a document that only argued one way would not be worth reading.
ContestedThe fuel loading, which is the first link and decides everything. See section 02. This is the real argument and it is not settled.
ContestedHow much of what burns becomes black carbon. The standard chain assumes two per cent of the fuel mass. The best-known dissenting study assumed effectively all of it, which its author conceded in print made the estimate high by a factor of ten to a hundred — while using a fuel loading tens of times too low. The two errors nearly cancelled, which made the two camps’ headline soot masses look like agreement when they were nothing of the kind.
ContestedHow many of the results rest on one model run. The five-teragram case has been run several times by several groups. Every case above it rests on a single ensemble member of a single Earth system model, one crop model, one fishery model and one year of food-balance data. That is a property of an expensive field with few groups in it, not a criticism of any of them, but it means the error bars round the larger figures are unknown rather than small.
EstimateWhere the independent work has landed since 2020. Three groups outside the original camp have now tested pieces of the chain — a national laboratory sweeping the fuel loading, a separate group testing plume rise with realistic humidity, and satellite observations of two continental wildfire events. On the physics of lofting and residence time they support the original result. On the fuel loading itself, none of them settles it, and the National Academies’ own survey of American urban fuel loads sits closer to the sceptical figure than to the standard one.
The largest omissions, which all run the same way.
Each of the things left out would make the outcome worse, and none of them would make it better. The published figures are therefore not a worst case dressed up; they are a floor, arrived at by counting only the parts that can be counted. The people who produced them say so in their own papers, and the model at the years after prints the same list on its own readout rather than leaving the absence to imply innocence.
That is the honest summary of the whole subject. The middle of the chain is solid, the first link is genuinely disputed, and everything not modelled points one way.
The principal papers behind this explainer are Crutzen and Birks (1982); Turco, Toon, Ackerman, Pollack and Sagan in Science (1983); Robock, Oman and Stenchikov in J. Geophys. Res. (2007); Toon and others in Science Advances (2019); Coupe and others (2019); Bardeen and others (2021); Xia and thirteen others in Nature Food (2022); and the National Academies’ review of 2025. They are cited properly, with page and table references, in the soot question.