The flood that came on a clear day

There was no rain. A slab of glacier let go above the Trishuli, and a river gauge reading normal at 8:40 was silent by 8:50. What a river network can see coming, what it cannot, and what warning has to look like when the water outruns the news.

At 8:37 on the morning of August 26, seismometers around the world registered what looked like a magnitude 5.2 earthquake in the Nepal Himalaya. It was not an earthquake. It was the sound of the north side of Langtang Lirung letting go: a slab of ice and rock roughly 0.2 square kilometres across, perched at about 5,200 metres, dropping 1,200 metres into the valley below. Much of the ice melted on impact. By one USGS reading the debris briefly dammed a side river before that gave way too. Seven minutes later a security camera at the Gyirong border crossing recorded the river arriving.

The sky was clear. Nobody along the Bhotekoshi and Trishuli was watching for a flood that morning, because every flood the valley knew came with rain, and the instruments the valley trusted were built to watch rain-fed water rise.

More than 1,450 people are confirmed dead in Nepal as of the start of October, and more than 5,700 are still missing, some of them carried 240 kilometres downstream into India. Thirteen hydropower projects were damaged, and more than 900 hydropower workers were reported missing, many of them from tunnels beside the river. It is among the deadliest floods in the country’s history, and unusually well documented, because for once the record of what the gauges saw is public, minute by minute.

What the gauges saw

Three minutes after the collapse, at 8:40, the water-level station at Syabrubesi, a short way below the border, read 1.62 metres. Its warning level is 6 metres and its danger level is 9. The river was low and ordinary. By 8:50 the station had stopped transmitting. It did not report a rise. It reported nothing, because it was gone.

At 9:05 the national flood forecasting division learned that a major surge had entered the Bhotekoshi from the Tibetan side. At 9:15 it pushed an SMS alert to more than 600,000 phones along the river in four districts. At 9:20 the next station down the river, at Betrawati, went quiet too. Its last reading was 3.55 metres, well under any threshold. Forty-three minutes after the mountain spoke, two gauges were destroyed and neither had ever shown a flood.

Below Betrawati the valley opens, and the wave began to behave like something a hydrologist would recognise. At 11:26 the gauge at Malekhu crossed its danger mark; seventeen minutes later the bridge there was gone. The flood passed Muglin around one in the afternoon and reached Devghat, where the Trishuli joins the Kali Gandaki to become the Narayani, at about 3:20. The peak there came at four o’clock, 6.57 metres, and by 6:30 the river was falling. Roughly 20 million cubic metres of water that had not been in the river that morning passed Devghat in a single afternoon.

Read the whole timeline and the flood splits in two. Above Betrawati it moved faster than a car on a highway, and the only warning anyone got was the sound. Below Betrawati it was a river flood on a fast clock: hours, not minutes, with gauges upstream crossing thresholds in the right order. The same water, two different problems.

Why no gauge could have warned the border

A river gauge measures water in the channel. On the morning of August 26 the water that destroyed Gyirong Port was not in the channel. It was ice, sitting on a mountain, and it had sat there through every monsoon the region had ever measured. The Chinese and Nepali warning system covering this stretch was built to watch glacial lakes and monsoon rises, and by most accounts it had done that job well. A rapid assessment by the HiRISK group of mountain scientists found that no early-warning system for glacier collapse existed in the area at all, and that given the speed of the initial avalanche, no conventional in-channel system could have cleared the border checkpoint in time. Further downstream, the same assessment concluded, a well-run system with ten minutes or more of lead would have saved a significant number of lives.

The one thing that did move before the collapse was the glacier. Satellite measurements show the ice and rock above the valley creeping at roughly 10 millimetres a month from January to mid-August, then accelerating in the weeks before it failed. That is a small signal on a mountain nobody was paid to watch. Thousands of glaciers and rock slopes in High Mountain Asia creep without ever collapsing, and the instruments that could tell the difference are in orbit, not in the river.

The valley had also announced itself before. The 2015 earthquake shook a glacier loose in the Langtang valley next door and buried villages, killing more than 350 people. In July 2025 a glacial lake in Tibet drained into this same river and took out the bridge at this same crossing. A year later the crossing was rebuilt, the customs queue was full of trucks, and hundreds of workers were underground in hydropower tunnels beside the water.

Four kinds of flood, four clocks

Every flood runs on a clock, and the clock is set by where the water was the day before.

Snowmelt floods run on weeks. The water is banked in the snowpack all winter and you can weigh it. Lake Muskoka’s record crest in April 2026, and the Assiniboine most springs, are this kind: the end of a story anyone reading the snow surveys could have followed since February.

Rain-on-river floods run on days. The water lands across a basin and collects, and the gauges upstream report it in order, hours before it arrives downstream. This is the flood the Severn brings to Shropshire many winters, and the flood most of the world’s warning systems were built for.

Flash floods run on hours, sometimes minutes. On July 10 this summer, 6 to 12 inches of rain (150 to 300 millimetres) fell on the hills of southeast Missouri in about twelve hours, and the Black River at Annapolis crested around 28.7 feet (8.7 metres), past the 27.4-foot record from 1993. Flood stage there is 8 feet. The rain was forecast, a flash flood emergency was out overnight, and it still took eight National Guard helicopters to lift 202 campers and counsellors off a flooded summer camp. One woman died. Warning was issued. Warning is only half the job.

And then there are floods where the water was never rain: a glacier collapse, a landslide dam letting go, a glacial lake draining, an engineered dam failing. These run on minutes at the source and hours further down, and the source is usually somewhere no river gauge has ever been. Nepal on August 26 was one of these. So was Chamoli, in the Indian Himalaya in February 2021, when a rock-and-ice avalanche came down the Rishiganga on a dry winter morning and killed about 200 people, many of them in hydropower tunnels. The clock at the source is close to zero. Everything depends on how fast the news travels relative to the water.

What warning has to look like

The Nepal timeline is a checklist, and nothing on it is exotic.

Instruments have to sit upstream of people, and someone has to treat their silence as a signal. Syabrubesi and Betrawati did their job perfectly and were destroyed for it. A station that stops reporting during an event is not missing data. It is the loudest reading a river can send, and most alerting, threshold alerts included, is built to ignore it.

The message has to be faster than the water. The collapse announced itself on seismometers worldwide at 8:37. The forecast division learned of the flood at 9:05 and the SMS went out at 9:15. On the published timeline, the alert reached phones about five minutes before the Betrawati gauge died, which is to say about as the water arrived. An alert keyed to the seismic signal itself, the way tsunami warnings are keyed to earthquakes, would have had a half-hour head start. Telling a collapsing glacier from an ordinary tremor in real time is not a solved problem, but the signal existed, and nobody was listening for it.

The message has to say what to do. The official SMS told people to stay on high alert, in Nepali only, without saying where or what to do. The warnings that moved people were phone calls. Rajendra Dawadi, the head teacher at a secondary school in Trishuli, received four separate warnings that morning, one of them from Betrawati directly upstream, and walked 900 children out of the building. The flood arrived ten minutes later. The good warning was specific, it came from a place, and it came from a person.

Someone has to have the authority to say go. Nepal’s hydrology department can forecast and it can warn, but it has no legal power to order an evacuation, and no law turns a forecast into one automatically. A study after a 2021 flood that killed 120 people recommended binding procedures linking forecasts to action. They were never adopted. Instruments, models and messages all funnel into a decision, and if the decision belongs to nobody, the rest is decoration.

And some places cannot be warned. The border checkpoint at Gyirong had seven minutes. The tunnels had less. No system fixes a clock that reads zero. The fix is not to put a customs queue, a work camp and a border post where a debris flow is going to arrive at highway speed, and that is a siting decision, made years earlier, in a valley that had already shown its hand twice.

The gauge upstream of you

Nibi Metrics does not track the Trishuli. It tracks more than 17,000 gauged locations, and almost all of them run on the slower clocks, which is exactly why the Nepal record is worth reading here. On a river with hours of lead, the difference between a warning and a rescue is whether anyone knew which gauge to watch.

So find yours. Locations on Nibi Metrics show how far the water sits from typical for the day and how fast it is moving. Learn what the number means at your dock or your low-water crossing. Set an alert on the gauge upstream of you, not just the one in front of you. Official warnings come from your flood agency, and a gauge does not replace them; it tells you what the river is doing in between. And when a station in a storm stops reporting, do not read it as no data. Read it as Syabrubesi at 8:50.

The Nepal timeline here is the Flood Forecasting Division of Nepal’s Centre of Hydrology and Water Resources Research, as reported by the Kathmandu Post; the glacier findings are from the USGS, the HiRISK consortium and satellite work reported in Nature; the casualty figures are those of Nepal’s authorities as reported by the United Nations on October 1; the Missouri figures are from the National Weather Service.