Seven Minutes to the Border

At 08:37 in the morning on 26 August 2026, seismometers across the Himalaya registered a shock near the Nepal — China border. Within minutes, a wall of water was moving down the Bhote Koshi — Trishuli river system. It destroyed Timure and swept through Rasuwagadhi, Syabrubesi and Betrabati. Damage was reported as far downstream as Galchhi and Malekhu, more than a hundred kilometres away. Hundreds of people were killed or went missing on both sides of the border.

Several of the river gauging stations that would normally record such an event were themselves destroyed by it. That absence is why this reconstruction leans on satellites, seismographs and physics rather than instrument records.

The early explanation, carried by most outlets in the days afterward, was a glacial lake outburst flood: a meltwater lake perched on the Purepu Glacier had drained catastrophically. It was a reasonable first guess. That exact lake had done something similar, on a smaller scale, on 8 July 2025.

But a single number in the record does not fit that story, and it turns out to be the number that matters most.

A surveillance camera at the Gyirong Port border crossing recorded floodwaters arriving roughly seven minutes after the seismic signal. The distance from the probable source area to that crossing is about 21.6 kilometres. That gives an average front speed of just over 51 metres per second.

Rivers do not do this. A flood wave moving down a channel is governed by friction against the bed and banks, and that physics imposes a ceiling. To find out where the ceiling sits for this valley, we built a hydraulic model of it — and the model, calibrated against the flood's own arrival time at a town 112 km downstream, moves its flood wave at about 12.6 metres per second.

The observed front was roughly four times faster than the physics of river flooding allows.

The investigationWhat we actually did

Answering this required assembling the event from scratch: locating the source, sizing the mass that failed, working out what it built in the valley, and modelling what came downstream. Each step used a different, independent line of evidence, which matters — because when independent methods converge on the same answer, that convergence is itself the result.

Six experiments, in sequence

The reconstruction ran as a chain, each step feeding the next. Where a step failed to confirm something, that failure is reported here as a result rather than quietly dropped — several of them constrain the answer as tightly as the successes do.

Experiment 01 & 02Where did it start?

The received account named a specific glacial lake. Testing that meant looking at the lake itself, and at the corridor of destruction leading away from it.

The lake is real. NDWI scanning confirmed the Purepu supraglacial lake system at roughly 28.404 • N, 85.59 • E — persistent, mapped, and genuinely the source of the July 2025 event. What no available satellite pass could establish was that it drained on 26 August 2026. Cloud cover over the source region in the only usable post-event optical pass ran to 75 — 79%, and the only Sentinel-1 radar pass in range predates the event by two days.

So we approached from the other direction: instead of asking whether a particular lake emptied, we asked how far up the valley the destruction actually reaches. Flood damage strips vegetation, and that leaves a clear signature in NDVI.

The honest conclusion from remote sensing alone: no specific origin point was visually confirmed. The convergent evidence brackets the origin to roughly the 15 — 25 km reach — consistent with, but not proof of, the Langtang Lirung north-slope location that glaciologists had named within days of the event.

That bracket was enough to proceed. The next question was not where but how much.

Experiment 03Weighing a mountain with a seismograph

A large landslide announces itself seismically. As an enormous mass accelerates downslope and then decelerates against the valley floor, it applies a long-period force to the Earth that seismometers record — and which, crucially, looks different from the sharp rupture of a tectonic earthquake.

Ekstr • m and Stark established the quantitative link between that signal and the mass responsible in a 2013 paper in Science. Their relations connect the peak force to a landslide magnitude, and the force to the mass. Applied here, they turn a seismic reading into an estimate of how much mountain came down.

Before trusting it, we checked it against the paper's own worked example, the 2010 Hunza-Attabad landslide in Pakistan: the formula returns 3.24 • 10¹¹ kg where field and inversion estimates give 1.1 — 1.4 • 10¹¹ kg. That is inside the factor-of-two uncertainty the authors state themselves — so the method is usable, provided nobody pretends the output is precise.

The magnitude will not sit still

Here the record becomes genuinely difficult. The event was first catalogued as a magnitude 4.4 earthquake. It was later reclassified by USGS as a magnitude 5.2 landslide, once analysts determined the shaking came from mass movement rather than fault slip.

A revision of 0.8 magnitude units sounds minor. Run through the scaling relation, it changes the implied mass roughly sixty-fold.

The likely explanation is not a record-breaking landslide but a mismatch of measurement scales. The Ekstr • m — Stark relation is calibrated against their own catalogue of events analysed by a specific long-period single-force inversion. A routine catalogue reclassification from "earthquake" to "landslide" comes from a different, more generic waveform assessment. There is no guarantee the two numbers mean the same thing, and treating them as equivalent would be exactly the kind of error this reconstruction tried to avoid elsewhere.

This work therefore proceeds on the original M4.4 — 4.5 reading, giving a slide volume near 1.3 — 2.2 million cubic metres, and flags the M5.2 figure as a real and unresolved open question rather than silently adopting or discarding it.

Experiment 04What that mass built in the valley

Knowing roughly how much material fell is not the same as knowing what it did on landing. A landslide dam's height depends entirely on the shape of the valley it lands in — the same volume that barely dents a wide floodplain can plug a narrow gorge completely.

So rather than assume a shape, we solved for one. At each candidate location along the bracketed origin reach, the DEM supplies a real cross-section; from that we build an actual area-versus-height relationship for that specific spot, and solve for the dam height at which the local cross-sectional area, filled to that height over its own width, accounts for the estimated slide volume. No geometry is imported from anywhere else in the valley.

An independent check falls into place

That dam site was derived purely from seismology and terrain geometry. It knows nothing about photographs of the mountain.

Separately, satellite-imagery review published in a hazard assessment put the avalanche detachment at roughly 5,200 m elevation, dropping about 1,200 vertical metres to the valley floor. Commentary circulating on social media described the fall as around 1.5 km.

The bed elevation at our independently solved dam site is 3,848.6 m. The drop from 5,200 m to that point is 1,351 metres — sitting almost exactly between the two reported figures. Two completely separate chains of evidence, one seismic and topographic, one photographic, landed in the same place.

The fall height also passes a physics check. For a drop of 1,200 — 1,500 m, frictionless free fall would cap velocity around 150 — 170 m/s. The independently reported descent speed of roughly 54 m/s is about a third of that ceiling — squarely inside the 20 — 60% efficiency that real rock and ice avalanches achieve once friction and entrainment are accounted for.

Experiment 05Modelling the flood that followed

With a dam height and an impounded volume, the next question is what happens when it fails. This needs two things: a description of the breach, and a way to route the resulting flood downstream.

The breach

Rather than prescribe a plausible-looking hydrograph shape, we used published regression equations from Froehlich (1995), obtained through a US Bureau of Reclamation review by Wahl (2001). That source was chosen deliberately: it does not merely cite competing breach formulas, it tests ten of them against a database of 108 real historical dam failures and reports which perform best. Froehlich's peak-outflow relation came out with the smallest demonstrated uncertainty of any method examined.

The 45 m breach width sits in the same range as the surveyed 27 • 9 m breach from the 2016 Zhangzangbo outburst in this same basin — larger here, consistent with a taller dam. That analog event, modelled in detail by other researchers, is the single most useful calibration reference available for this valley.

Routing it downstream

Experiment 06The test the model nearly failed

A flood model that gets the timing right can still be wrong about everything else. The independent check is width: how far across the valley did the water actually spread? Sentinel-2 imagery answers that directly, and we measured it at 36 cross-sections through the Betrabati — Malekhu reach.

Run as clear water, the model predicted a median inundation width of 52 m. The satellite says 385 m. The model was seven to eight times too narrow.

That gap is not a rounding error, and no adjustment to the peak discharge fixes it. What it points to is that this was never clear water. Every account describes an ice-and-rock avalanche; the flow that resulted carried an enormous sediment load, and a debris-laden flow behaves differently from water in two specific ways — it is bulkier, and it is far rougher.

The remaining factor of three had a more interesting cause than "the physics is wrong". Looking at individual cross-sections rather than the median, the mismatch ranged from 0.17 • to 18 • . At some points the model was already too wide. The worst failures clustered precisely where the observed damage was widest — which is exactly where the measurement method itself breaks down.

Our width extraction walks two straight rays outward from the channel centerline until the terrain rises above the modelled water surface. Where floodwater spread laterally around a low rise into a connected floodplain, those two rays stop at the rise and never find it.

So we re-measured the identical model output a second way: an eight-connected flood fill, taking any terrain topologically connected below the water surface. That method has the opposite failing — it has no concept of momentum or travel time and will happily fill any low ground it can reach.

The answerAn avalanche pushing water, not a dam filling up

Return to the seven minutes.

The instinct that a landslide dam must fill before it bursts is sound physics, and we tested it directly. Filling the 1.11 million m • impoundment at plausible monsoon inflow rates takes somewhere between two and thirty-one hours — around eight hours at a mid-range estimate. That is the right description of a blockage that forms and then slowly ponds up behind itself. It cannot describe water arriving 21.6 km downstream seven minutes after the slope failed.

Three independent lines of evidence resolve it.

Speed. At 51 m/s, the front was moving four times faster than friction-governed channel flow permits. An impulse wave — water displaced by a mass entering the channel — is not bound by that limit, because it is not being routed, it is being shoved.

Appearance. The border footage shows what looks like water, not a visible wall of rock and ice. That fits a leading edge made largely of the river's own pre-existing water, pushed ahead of the slower, heavier debris front, which drags against the bed in a way the water surge does not.

No filling time was needed. This closes it. At the M4.4-derived mass, an ice-rock avalanche that is even 20 — 30% glacier ice by mass carries 0.65 — 1 million m • of water-equivalent; at 50 — 100% ice it reaches 1.6 — 3.2 million m • . Those figures are comparable to, or larger than, the entire impoundment volume solved from the dam geometry. The water did not have to accumulate from the river. Most of it was already sitting frozen on the mountain, and it arrived as ice.

The two processes are separated in time, and the confusion between them is what made the original account seem plausible.

That last entry matters. Nobody is claiming landslide-dammed lakes do not form and slowly breach in this basin — one did, forty-eight hours later, in the textbook manner. Having a clean example of the slow process in the same valley days afterward is what makes the case that 26 August was something different, rather than a quarrel about definitions.

LimitsWhat this reconstruction does not know

Ordered by how much each would change the picture if resolved.