On 26 August 2026 a piece of a Himalayan mountain fell into a river valley in Rasuwa, Nepal. Seven minutes and forty seconds later a wall of ice, rock and mud taller than a ten-storey building destroyed the border crossing at Rasuwagadhi, 22 km away, and turned down the Bhote Koshi into Nepal. More than 1,340 people are dead and nearly 4,900 missing. This is a plain-English account of where all that water came from, how I tested the idea, and what I got wrong along the way. It is preliminary, it is not peer-reviewed, and the parts that depend on the routing model are still unresolved. The full technical version, with every number and every assumption on display, is here.
Langtang Lirung is a 7,200 m peak about 60 km north of Kathmandu. At 8:37 on the morning of 26 August 2026 — clear, no rain, late monsoon — a piece of its north face about 600 m wide let go from around 5,200 m. It registered as a magnitude 5 earthquake. A drone at a Chinese radar post happened to be filming: a white stream of ice leading a brown cloud of rock dust. Everyone who has looked at the pictures says the same thing — the bedrock failed and took a hanging glacier down with it. Nobody has said how much of it was ice, and published estimates of how much fell differ by a factor of 400.
The mass dropped about 1,200 m onto loose ground below the face, scouring as it went, then another thousand to the stream bed, and became a torrent of ice, rock and mud that ran 22 km down the Lhende in 7 minutes 40 seconds — about 172 km/h, and it did not slow down. (This page said 7 minutes flat until 6 September; the correction is below.) It destroyed the Rasuwagadhi border crossing and the hydropower plant there, met the Bhote Koshi, and turned down it as a flood.
From there it ran the length of the Bhote Koshi and the Trishuli. Betrawati had 43 minutes' warning, Galchhi over two hours, Devghat about seven. The Flood Forecasting Division issued its first warning at 9:15 and sent 679,295 SMS messages; the river did not cross its warning level at Malekhu until 11:20, Muglin until 14:14, Devghat until 15:20. Four gauges were destroyed by the flood, two inside the first hour. Thirteen hydropower projects were damaged.
As of 8 September the reported toll stood at 1,342 dead and 4,886 missing in Nepal, with 43 dead and 519 missing on the Chinese side and more than 5,300 injured. The missing count is falling while the deaths rise — that is bodies being recovered, not people found alive. Every figure here is provisional.
A falling mountain is rock and ice. Rock doesn't flood. So where did 8,000 pools of water come from? The standard first guess after disasters like this is melted glacier ice — the fall grinds the ice up, friction heats it, and it melts into flood water.
I doubted that, for a reason you can check with high-school physics.
When something falls, its energy of height turns into heat. A fall of 1,200 metres — roughly what this mass fell before it hit the river — gives each kilogram of material only enough heat to melt about 35 grams of ice, and that's if every scrap of the heat goes into melting, which it never does.
How much water that adds up to depends on how much fell, and that is the number nobody agrees on. So I did it twice. At the collapse size my own evidence supports — 14 to 34 million cubic metres, which is a later section of this page — melting supplies about 1 million cubic metres of water, or 2.5 million if you turn every dial in melting's favour at once: the top of that size range, four-fifths ice, the longest plausible drop, and the most generous heat-sharing anyone has published. Against 20 million of "extra" water, that is short by a factor of eight.
Melting only reaches the official figure if you take the very largest published collapse — 200 million cubic metres, six times the top of my own range — and also assume four-fifths ice, a 4,000-metre drop and the most generous heat-sharing. Do all four at once and you get 20.7 million: it arrives exactly, and only, at the end of a chain where every link is set to its most favourable value. Relax any one of them and it falls away fast.
So most of the flood's water was not made by the mountain. It was already in the valley.
Picture the monsoon-swollen river as a long line of water lying in the valley, already flowing. Now send a fast, heavy wave down it. The wave acts like a snowplow: its front overruns the slower river water ahead of it, sweeps it up, and pushes it along. Kilometre after kilometre, the wave grows by collecting river that was already there.
When I do the arithmetic along the whole 199 km, roughly three-quarters of the water that passed the downstream gauge was ordinary river water, gathered up and delivered all at once. The mountain provided the push; the monsoon provided the water. Other Himalayan and Andean disasters, when examined closely, have told the same story.
This distinction is not academic. If flood volume comes mostly from the valley's own river, then the size of a future flood depends on the season — the same collapse in February and in August produce utterly different disasters.
Remember the border camera: 22 km in seven minutes forty. I simulated the collapse many different ways — small and large, dry rock, icy, wet — and asked which versions could reach that camera on time. The result was stark. Every dry-rock version arrived late — 17 to 30 minutes, never under eight — no matter how big I made it. Every icy or wet version arrived within a few minutes of it — inside the tolerance the test allows; the chart below shows how close each came.
The physics is intuitive once said aloud: dry broken rock fights itself. Every fragment grinds against its neighbours, burning off the energy of the fall — like a truckload of gravel dumped down a slope, violent but self-braking. Ice and water lubricate. A wet or icy mass keeps the fall's energy as speed. Only something lubricated — by ice, by water, or both — covers that gorge in under eight minutes. (Until 6 September this sentence said "only something running like slush", which claimed more than the test shows; the next section says what the lubricant most likely was.)
Be careful which way that arrow points. The test excludes dry rock; it does not establish slush. "Not dry" is a wide category, and my own size-and-composition search later on this page cannot narrow it: how wet the mass was, and how much of it was ice rather than rock, come out spanning almost the entire range I allowed. So the honest claim is the negative one, and the heading above says so. What I can add to it is circumstantial and consistent — ice is visible in the border footage, and several groups describe a bedrock failure that carried a hanging glacier with it.
How much water was there to sweep up before the border? The Lhende is a small river — about 150 cubic metres a second where it meets the big one, less higher up. All the water lying in its 22 km of channel that morning comes to about half a million cubic metres. Add pore water squeezed from the gravel it tore up and the little that friction had melted, and you get perhaps 2–4 million. The mountain that came down was tens of millions. So what hit the border post was at least four-fifths mountain: rock and ice, with some mud. The water story starts at the junction, not above it.
That leaves the question of how a mass of rock and ice got 22 km down a gorge in 7 minutes 40. The answer is the ice — not as water, as ice — and the physics is the physics of an ice skate.
It is not mainly pressure melting; that lowers the melting point by only 0.0074 °C per atmosphere, and you can skate at −30 °C. Three things do the work. Ice carries a disordered, liquid-like film on its surface with no load at all. Sliding heats the contact points, and at avalanche speeds the power per square metre is enormous, so a thin film melts continuously. And ice fragments as it goes, so the mass keeps making new slippery surfaces.
The literature is more careful than the slogan. Drum experiments found granular ice cuts a gravel mass's friction by up to 20% and water by about 50%, with mixtures above ~40% ice nearly liquefying — but in a field set of 64 rock–ice avalanches the same group could not confirm ice fraction as a control on runout. What showed up was water, volume, and travel over a glacier. The closest analogue is Kolka–Karmadon in the Caucasus in 2002: 100–130 million cubic metres, mostly glacier ice, which ran 18–20 km at 50–80 m/s and then stalled where the valley stopped it while the liquid part ran on. This event is that, at a fraction of the size.
The whole finding rests on one number, the border clock, so when geopera — careful satellite work, published openly — put the border arrival at seventeen minutes, it was serious. My dry-rock simulations arrive at seventeen minutes exactly. If seventeen is right, the simplest reading of the event is a dry rock avalanche and my finding is backwards.
Chasing it took a day. They never mention the camera: the footage had been censored on Chinese platforms and was not available to them. Fact-checkers who did see it converted its Beijing timestamp loosely as "around 9 o'clock in Nepal", which read literally is 23 minutes — a second vote for the slow answer. The resolution was in Chinese-language reporting: the footage is stamped 10:59 Beijing, which is 8:44 in Nepal. The conversion had been done correctly at source and "around 9" was a loose paraphrase of it.
Then I stopped taking anyone's word for it and measured the front off the video myself. That is the next section. The clock also moved in the process: it was 7 minutes flat on this page until 6 September and is now 7 minutes 40 seconds. My own number, corrected against myself.
Rather than trust anyone's reading of the border footage, I measured the front off it directly. The camera looks down the valley past a building of known size; the debris front crosses two sight lines 50 m apart in 1.066 seconds, giving 47 m/s, with an envelope of 41–63 m/s once every source of error is carried through.
Two things had to be got right first. Internet video repeats frames, so a standard tool reported only 24 of every 60 as real; the timing had to be done on genuinely distinct frames. And the front blurs into the cliff shadow, so I deliberately took the later, slower frame at each crossing, which biases the answer down rather than up.
That number agrees with geopera's independent 45–52 m/s from superelevation at bends — two different methods, two teams, the same reach. It is the piece of primary measurement I trust most on this page. The technical report has the geometry and the full error budget.
Where the flow hit the cliff at the junction head-on, it ran up the rock face and left mud lines. Run-up converts speed into height: a mass arriving at v climbs roughly v²/2g if it stops dead. Three mud lines there give a front speed near 54 m/s by a route that never touches the camera clock, and that is why they matter — an independent check on the number the whole argument rests on.
They are not a clean confirmation. Read as pure depth they say one thing, read as pure run-up another, and I published both readings after getting it wrong in each direction first. Against the corrected clock's 47.8 m/s the agreement is within about 15%, not the "few per cent" this page once claimed.
This is the number everyone disagrees about, and the disagreement is enormous: published estimates run from half a million cubic metres to two hundred million.
Arguing about it one estimate at a time gets nowhere, so I asked a different question. Not "which number is right", but which combinations of size, wetness and slipperiness are consistent with everything I can measure at once — the border clock, the speed at the border, how much valley floor the flood scoured out, how much debris it left, and how deep the water got everywhere along the river. Then run the model hundreds of times across the whole disputed range and keep only the versions that pass every test.
An early answer of 14–34 Mm³ is on this page in places and I no longer believe it. Three things say it is too small. Part of the flow ran about 3 km up the main valley on the Chinese side and came back, and the pond it left holds ~23 Mm³ on its own. A mud line at the Upper Trishuli-1 headworks, 21 km downstream, sits 45–70 m above the old bed where the passing runs put 5–21 m of water. And the mud-line map of the whole corridor, made on 7 September, needs a release nearer 110–175 Mm³ in this model.
Two measurements published since would make that ordinary rather than alarming. The first academic reconstruction maps the piece of mountain that broke away at 1.009 km²; spread 110–175 Mm³ over that and the slab was 109–173 m thick, where Chamoli in 2021 was about 135 m. And another analyst's estimate of ~100 Mm³ ±40% overlaps my range from below.
But I have not replaced the number, and I may not be able to. Between 7 and 8 September I built seven versions of the routing model, each fixing something the last one got wrong — the valley widths, the drag law, the debris limit, how long the collapse took to feed into the gorge, and finally a floodplain the flood could spread across. None passed. One version produced a single run that met all eleven tests and then failed all three measurements it had been kept away from, which is exactly what that reserve is for.
What stopped me was not the failures. It was checking my own scoring. I had been comparing the model to the mud lines as six reach averages; when I compared it point by point instead, at all 1,100 places where a mud line was measured, the one run that ever passed everything matched at 11% of the points in the first 22 km. It hit every average while being wrong at most places inside them. That is not a model needing an eighth version. It is a question not asked precisely enough to have an answer, and on 8 September I stopped and started again on a different footing.
So the size is moving up, by a lot, and the answer looks like a mass that was mostly ice and water. Read the number as under revision, and understand it may end up withdrawn rather than corrected. The details of what was tried and why each attempt failed are in the code and in the changelog rather than here.
This is the part of the work I would defend hardest, which sounds odd until you see why.
The model could always drop sediment but never pick any up — a flood that scours a valley floor was, in my equations, incapable of scouring anything. So I built that in, using erosion formulas published for entirely different rivers on other continents, with nothing tuned to this event. It worked better than it had any right to: the model tore 3.8 million cubic metres out of the corridor, against 3.2 million measured from satellite stereo imagery, and it put the erosion in the same reaches.
But the same change broke something else. Now that the model could drop sediment properly, it dropped far too much, and the flood stalled before it reached the towns downstream. So I tested it honestly: hundreds of runs across every plausible combination of inputs, asking whether any version could arrive on time and leave the right amount of debris. None could. Not one in a hundred and fifty.
That is a real result, and a harder one than a good fit. It doesn't say my numbers were slightly off; it says the shape of the model was wrong. And the fix was to stop pretending "sediment" is one thing:
And here is the part that made me trust it. geopera reached the same conclusion the same week by a completely different route: fitting their own model to their satellite measurements with machine-learning optimisation, which pulled their sediment settling rate down by a factor of three and told them the material "behaves finer than assumed, remaining in suspension longer". Two methods, one answer — but be clear about how much weight that carries. geopera is not an institution: it is Geopera Pty Ltd, and the analysis is by its founder, Darcy Weedman, published on a company blog and not peer-reviewed, exactly like this one. Two unreviewed analyses agreeing is real evidence and it is not the same as replication. Their two posts also give different border arrival times, 17 minutes in one and 23 in another, which is worth knowing when I lean on them elsewhere. Their satellite work is better than mine and I lean on it a lot; I would rather name the weight than borrow it.
One more idea from the technical work, because it explains something survivors reported: the flood came in surges, not one clean wave.
A river valley behaves surprisingly like an electrical circuit. Narrow gorges are like wires — the flood shoots through. Wide basins and side valleys are like capacitors — they fill up while the surge passes, then drain back afterwards, smoothing the peak and stretching the tail. Temporary debris dams are switches that hold, then break.
There is a striking example right at the border. The flood came down a side valley, the Lhende, and slammed into the main one. The main valley doesn't just deflect it — the water piles up and runs backwards, uphill, into the main valley's upstream arm. Tracing the height of that pile-up on the map, it should reach about 3.5 km up the wrong way. China's own reconstruction says part of the flow ran "nearly 3 km upstream", and Chinese state media reported that about 3 km of the highway approaching the border port was destroyed. Satellite images, measured along the river on 6 September, show the ground stripped bare up that arm fading out between about 2 and 3.5 km from the junction — while the two directions the flood genuinely flowed through stay stripped for the whole 5–8 km I could see. Sharper commercial imagery shows fresh sediment on the valley floor a further kilometre or so beyond that, which may be the surge or may be what settled in the reach that ponded behind the debris plug at the junction; the technical report keeps both readings. Four different clues, one answer: a large volume of this flood briefly went about three kilometres the wrong way, and came back late.
Any model can be tuned to "explain" the event it was built on. The test is whether it works somewhere else. So I ran mine, unchanged, against Chamoli 2021 in India — a similar mountain collapse in opposite conditions: winter, a nearly dry river, and a mass that was four-fifths rock. It failed at first, in a useful way, and chasing the failure taught me the missing physics: how wet a moving mass is controls how much it grinds against itself, and heat takes minutes to soak into ice before it melts. With those pieces added, one set of equations explains both events. That test stands — but as a partial pass, not a clean one. The model now delivers the front to Tapovan in 36.3 minutes against an observed 34–37, which is the headline number; inside the same run it crosses the Raini basin in 13 minutes against the roughly 27 an independent analysis published for that reach. When I fixed a bug in the arrival calculation, the arrival got better and the basin got worse. Both halves are in the technical report, because reporting only the half that worked is how a stress test stops being one.
A second test, against the 2012 Seti River flood in Nepal, I reported as a clean pass. I have withdrawn it. The river channel that test ran on was built by a routine that fills gaps in the map data with straight lines — and it had flown those straight lines across a mountain rim. The repair meant to clean up bad elevations then flattened 31 of the 54 kilometres to zero slope. The model had been slowed by terrain that did not exist. On a channel with a real gradient, the same model arrives four times too early, and a second conclusion I drew from that test reverses outright. Both are off the table.
The hardest fact in this event is that the people closest to the mountain could not have been reached by any warning that has to pass through a human being. Seven minutes forty is not enough to notice, decide, call, and have someone run. Nothing that existed anywhere that morning would have reached the border crossing in time.
But that is not true of the rest of the river. Betrawati had 43 minutes. Galchhi had over two hours. Devghat had seven hours. Those are ordinary warning timescales, comfortably enough for phone trees, sirens and evacuation, and the towns and hydropower sites along that stretch held far more people than the first 22 kilometres did. Where the dead actually were is not something I can say from a model, and the district-by-district breakdown was not published at the time of writing; the point here is about the time available, not about apportioning a toll.
The seismic network detected the collapse within minutes. It was logged as an earthquake. The warning followed at 9:15, 38 minutes after, to 679,295 people — and along the lower river it ran well ahead of the water: the river did not cross its warning level at Malekhu until 11:20, Muglin until 14:14, Devghat until 15:20. So what the 38 minutes cost was the top of the river, where, as above, no chain that involves a human being could have worked at all. It is worth being careful about what that does and does not mean. Automatically classifying a landslide seismic signature in real time and routing it to a downstream flood warning is something the research literature says is feasible — GFZ's Niels Hovius has put the achievable delay at minutes — but it is not, as far as I can find, running operationally in any national warning system, and the signal looks like a small earthquake because that is very nearly what it is. The forecasters were also losing their instruments as they worked: the gauge a kilometre below the border stopped at 8:40, Syabrubesi at 8:50, and the one at Malekhu was swept away after reading 10.48 metres. So this is a gap in what the technology can currently do anywhere, rather than anything that went wrong in one control room — and the record shows a warning that reached most of the river well ahead of the water.
An independent audit by geopera of whether this collapse could have been predicted came back null on all four satellite channels: under two metres of creep in a hundred days, steady, with no acceleration in the final 48 hours; ordinary glacier thinning; and a warm summer nearly identical to the two before it that produced no collapse. You cannot watch every peak. But you do not have to predict the collapse to reach the people two hours downstream — you only have to notice it and route it. The hours between the collapse and the flood's arrival at the lower river were real, and no part of the system was built to use them. That is a statement about how these systems are designed everywhere, not about anyone's conduct on the day.
Five results have been withdrawn and several more corrected in the two weeks this has been running — a test on another river that turned out to rest on a bad channel profile, a dammed-river explanation that needed 37 hours of blockage where there were minutes, two successive misreadings of the 70-metre mud lines, and a rule of thumb that assumed the wave travelled as a rigid block when my own earlier work had disproved exactly that. Two cold reads of these pages, on 5 and 6 September, caught several more. Every one is still up, dated, with the reason, on the changelog page. I am not listing them here; the point is the rate, not the detail.
The size of the collapse is genuinely unknown and I have not settled it. See the section above. Seven versions of the routing model have failed, and the reason I stopped was a problem with my own scoring rather than with any one version.
The model puts the debris in the wrong place. Satellite stereo shows the main deposit 40–43 km downstream where the valley opens; mine drops it too early. geopera's model got this right and mine did not.
The border speed is the number I fit worst, and it is one of the inputs the size estimate was scored against. Five estimates exist — 54, 45–52, my own video reading of 47, my routing fit of 31, and a peer-reviewed 19 — and I scored against 45–52. That is a real exposure, not a rounding error.
The lower river fails out of sample. The height it rose at Galchhi came out around 3.6 m against roughly 9 observed. The peak at Devghat met its factor-of-two test, but every passing run landed below the observation, which is "not contradicted" rather than "reproduced".
One correction worth naming, because it was mine and recent. On 8 September my literature search picked up an "additional flood volume" of 19.96 million cubic metres and I briefly treated it as a new constraint. The source turned out to be a document already in my own data folder: a two-page press release in Nepali from the Flood Forecasting Division, issued the day after the flood, saying approximately 2 crore cubic metres — 20 million, one significant figure, expressly preliminary — for the water above base flow past one gauge, Devghat, between 14:10 and 18:00. Not a measurement of the whole flood, and not comparable with the size of the collapse. The two-decimal version is not in it. I have transcribed and translated the release in full so anyone can read it rather than take my word for it.
And the standing caveat. This is fast, independent, AI-assisted analysis published days after a disaster by an engineer working outside his own field. It is not peer-reviewed and no Nepali scientist has read it. Where my numbers disagree with an agency's, both are shown. geopera's satellite work is better than mine and I lean on it.
Dave Hume, an engineer in New Zealand, working in the evenings over ten days with Claude Opus 5 (Anthropic) running in Claude Code, a terminal tool where the model can write and run programs, read files, search the web and publish pages. The AI wrote the physics code, ran the simulations, processed the satellite imagery and drafted the prose; the human set the direction, brought the original hypothesis and did most of the catching. It is on the page rather than in a footnote so that you can discount it if you think that changes what the work is worth.
What it was bad at, since that is the part worth knowing: confident errors in its own fresh work. Sediment counted twice in the ledger; a chart whose bars were drawn to the wrong axis; a bar on this page's own melt chart mislabelled for three days. All were in things written minutes earlier and believed, and none would have been caught by asking the model whether it was sure. Every substantive correction on this site was found by a human pushing back or by re-reading the pages cold, never by the model auditing itself.
Rules I worked under, kept in the repository since day one: state conclusions as ranges over disputed inputs rather than single numbers; never tune a parameter to fit the very observation you are scoring against, and if you must, declare it and hold other measurements back as a check; report failures with their diagnostic shape rather than hiding them; credit the researchers who got there first — the contribution here is arithmetic, not correction. Where those rules were broken, the breach is on the page. A longer account of the division of labour, including the specific errors and who caught them, is in PLAN.md rather than here.
Checkable. Every number on this page comes from code in a public repository that runs on open data — seconds for the budgets, minutes for the routing, about half an hour for the size envelope. The executed workings are here, and the full development history, including every retraction and the reasoning behind it, is in the commit log. Nothing rests on trusting me.
Not a request — a note on the limits, so you can weigh what you have just read. Everything here is built from what is already public.
The full gauge records from Galchhi and Devghat. A photograph has no time axis, so no amount of satellite imagery can measure how much water passed; only a gauge trace can. I do not have those traces, and almost every number here would be sharper if I did.
Anything measured inside the thirteen hydropower projects — headworks sensors, gate logs, control-system archives. Those plausibly hold the only high-frequency measurements of the flood's upper reach, where the public gauges were destroyed. I have none of it.
Repeat satellite elevation surveys. If the deposit was ice-rich it must go on losing volume as buried ice melts out, and rock does not do that. Two surveys a month apart would separate the two with no model involved. None has been published, so the question stays open here.