At 07:34 local time on 10 August 2026, a magnitude 7.4 earthquake struck western Colombia. Its epicentre was about 5 km from San José del Palmar, on the border between Chocó and Valle del Cauca, but the number that governs everything else is the depth: 110 km. First reports put the toll above 230 lives lost and more than 900 injured, concentrated in Cali and Pereira, with damage running from Chocó through the coffee belt. This is an engineering reading of the event, written with our thoughts for everyone affected.
This was not a shallow crustal earthquake like the one that struck Venezuela in June. It happened inside the Nazca plate as it descends beneath the Andes, and that single fact changes the shape of the loss: where the damage lands, how many cities it reaches at once, and whether a risk model saw it coming at all.
A rupture 110 km down has no near field
In a shallow earthquake, the damage concentrates: the closest sites are a few kilometres from the rupture, and intensity falls away quickly. At 110 km depth the geometry inverts. The nearest point of the source is already 110 km from every site on the surface, so moving 100 km sideways only takes the distance from 110 to about 145 km. The energy is spread over an enormous footprint instead of concentrated on one town, and the shaking lasts longer.
That is exactly the pattern of 10 August: severe damage in Cali, Pereira, Manizales, Armenia and Chocó at the same time, in cities up to 200 km apart. For an insurance portfolio that is not five events, it is one event with correlated loss across five urban concentrations, and it is precisely the accumulation case that a model calibrated on the intuition of shallow earthquakes understates.
The deep corridor beneath the coffee belt
The dashed lines are the depth of the descending plate. Under Manizales it passes about 107 km down, under Pereira about 112 km, under Armenia about 124 km. The coffee belt does not sit near this source. It sits on top of it.
The corridor had warned twice. The catalogue read it too small.
This corridor is not quiet and never was. Since 1960, 110 earthquakes of magnitude 4.5 or greater have occurred between 70 and 200 km depth beneath the Cauca valley and the coffee belt. Twice it went much further. In February 1938 a magnitude 7.0 broke about 150 km beneath Filandia, in the middle of the coffee belt. In November 1979 another one broke 5 km from the 10 August epicentre, at almost the same depth. The region had already given its warning, twice.
| Date | Mag. | Location | Depth | Note |
|---|---|---|---|---|
| 10 Aug 2026 | 7.4 | 5 km S of San José del Palmar, Chocó | 110 km | This event. Intensity VIII in the coffee-belt valleys, 60 to 110 km away. |
| 23 Mar 2019 | 6.1 | Versalles, Valle del Cauca | 122 km | Felt across the west of the country, limited damage. |
| 02 Sep 1997 | 6.8 | Roncesvalles, Tolima | 199 km | The deepest of the large ones, and still damaging across the coffee belt. |
| 19 Aug 1995 | 6.6 | Neira, Caldas | 120 km | Damage in Pereira and Manizales, directly above the source. |
| 23 Nov 1990 | 6.1 | Salento, Quindío | 145 km | Beneath the coffee belt. |
| 23 Nov 1979 | 7.2 | El Cairo, Valle del Cauca | 110 km | Same patch of the descending plate as 10 August 2026. Moment magnitude 7.2; the body-wave magnitude in the same record is 6.4. |
| 03 Apr 1973 | 6.2 | Filandia, Quindío | 158 km | Beneath the coffee belt. |
| 30 Jul 1962 | 6.5 | Tadó, Chocó | 64 km | Surface magnitudes between 6.5 and 6.8 in the reviewed record. |
| 05 Feb 1938 | 7.0 | Filandia, Quindío | ~150 km | Directly beneath the coffee belt. Surface magnitude 7.0; depth fixed by the analyst, so it carries real uncertainty. |
Magnitudes and depths from the reviewed international bulletin and the global moment tensor catalogue. For 1938 and 1962 only surface and body-wave magnitudes exist, and the depths are fixed rather than solved: they are precedents, not precision.
Look closely at the 1979 row, because it is the whole problem in one line. Read the body-wave magnitude that most catalogues carry as the headline number and that earthquake is a 6.4. Read the moment magnitude, the one that measures the energy actually released, and it is a 7.2. The same earthquake, five kilometres from the 10 August epicentre and at the same depth, is either a moderate event or a major one depending on which column you read. Body-wave magnitudes saturate: above roughly 6.5 they stop growing with the earthquake. A model built on an unhomogenised catalogue inherits that ceiling and prices the corridor as if its largest event had been a 6.4.
That is the first trap a catalogue sets for any risk model, and it compounds with the second: with intervals of decades to centuries between the largest events, a 60 to 100 year record does not contain the tail of the distribution anyway. What was missing in this corridor was never the activity, and after 1938 and 1979 it was not even the size. What was missing is that the cities built on top of it had never been tested by one.
The zonation we already carried, and why the coffee belt is in it
Our Colombia model is built in four layers: the mapped crustal faults, the deep sources of the descending plate, the great offshore rupture on the Pacific margin, and a smoothed grid that carries the moderate seismicity none of the named sources own. It carries 377 mapped fault traces, and six near-field faults were curated by hand down to the geological slip rates published by the national survey, because the raw database inflated them.
The decision that matters for 10 August was made months earlier, and for a completely different reason. Colombia's large historic inland earthquakes, a magnitude 7.3 in 1917 at 221 km depth among them, are not crustal events: they are deep, inside the plate. Every depth filter we inherited was cutting them out, and the result was that the interior of the country had no source at all for its largest historical earthquakes. So we extended the deep corridor inland, eastward, until it covered the whole descending slab under the Andes, and rebuilt its rate from the seismicity that actually occurs there.
That corridor runs from 60 to 250 km depth, and it carries a maximum magnitude of 7.8. Beneath the coffee belt it overlaps two other families: the Cauca-Romeral fault system at the surface, which produced the Armenia earthquake of January 1999, magnitude 6.1 at 17 km depth, close to 1,200 lives lost and a reconstruction bill of the order of 1.5 billion dollars; and the smoothed background grid. Manizales, Pereira and Armenia are therefore modelled with three independent source families stacked under them, not one.
The Cauca-Romeral system runs straight through the coffee belt. Shallow, close, and the source of the Armenia disaster. This is the layer everyone models.
The descending plate, from 60 to 250 km, extended inland so the large historic interior earthquakes finally had a source. 10 August, magnitude 7.4 at 110 km, sits inside it.
The moderate seismicity no named source owns, smoothed from the catalogue so it is neither double counted nor lost between layers.
The claim here is a narrow one, so let us be precise about it. We are not claiming to have predicted this earthquake; nobody predicts earthquakes. We are claiming something a (re)insurer can actually audit: the 10 August event falls inside a source our Colombia model already carried, below the maximum magnitude we already assigned to it, in a corridor we had drawn under those exact cities. The model was frozen before the event. How we check it is deliberately two-level, and we say which level we are citing: consistency first, against the open USGS model for South America, where our Colombian control cities sit inside the ±25% band we require; and accuracy second, which no benchmark can give you and only a real earthquake can. That second level is what 10 August now makes possible, and it is below.
Model under discussion: Sismicus Colombia v2.0, frozen August 2026. Every figure on this page refers to that version.
How a routine filter deletes 90% of a deep source
Before rates are computed, catalogues are "declustered" to separate independent earthquakes from aftershocks. The classic window algorithms were designed for shallow crustal seismicity: they remove everything within a radius and a time window of a larger event. Applied to a deep cluster, where thousands of genuinely independent earthquakes concentrate in a small volume, they do something quietly catastrophic.
We measured it while calibrating Colombia. In the Bucaramanga nest, the densest of the country's intermediate-depth clusters, the standard window at magnitude 4.5 spans 34.7 km and 77 days, which is larger than the nest itself. It flagged roughly 1,632 of some 1,800 earthquakes as aftershocks, and the rate collapsed from about 29 events a year to about 2. The hazard under Bucaramanga and Bogotá fell with it. The fix is not a fudge factor: the volume is declared a continuous process, its earthquakes stay independent, and the rate that feeds the model is the real one.
It was measurable, identifiable and correctable, but only because someone went looking. That is the honest question this event puts to the whole industry, ours included: how many models in production have looked? The same treatment is now being applied to the corridor that broke on 10 August: its volume declared continuous, its rate recomputed from the restored record, and the ceiling of the deep source moving from a single number to a branch of three values, so that the uncertainty is carried instead of assumed. Both changes go into the next version of the Colombia model. Both rest on physics that was on the table before 10 August, which is the only reason we are allowed to call this an improvement rather than a correction.
Where the ground decides
The intensity maps agree on something that should unsettle anyone pricing this region: the mountainous epicentral zone was not where the ground shook hardest. The strongest shaking landed on the populated sedimentary valleys around it, where soft ground amplified the waves. Pereira, 61 km from the epicentre, reached intensity VIII. So did Cartago and Quibdó. The municipality above the epicentre reached VI.
The epicentre was in the mountains. The shaking landed in the valleys.
A rupture 110 km down has no near field: the shaking spreads instead of concentrating. The model that matters is the one carrying that deep source, not only the faults you can see at the surface.
The mountains above the epicentre were shaken less than the sedimentary valleys 60 to 110 km away, where the exposure is. Site response and building stock decide where intensity becomes damage.
Cali, Pereira, Manizales, Armenia and Quibdó were hit by the same event, spread over 200 km. For a portfolio that is correlated loss, and it is the accumulation case a shallow-calibrated model understates.
This is a design conviction in Sismicus, our hazard layer, not a post-event observation: hazard is delivered at the surface, with the dynamic amplification of the ground already integrated, site by site, across the full range of building periods from 0.01 to 10 seconds. A hazard map on rock is an intermediate product, not an answer. If a portfolio model applies generic site factors by soil class, or applies none at all, 10 August is the case study for why that is not enough.
From engineering to risk intelligence
"Most probabilistic seismic models are calibrated on generic or historical data. Ours is built on a unique, proprietary dataset generated through the design of critical infrastructure — from Panama Canal bridges to metro systems and the high-speed rail between Dubai and Abu Dhabi. This gives us a fundamental advantage: a hazard engine that is not only statistically sound, but informed by real-world engineering at the highest level."
— Carlos Caramés Molero, Founder & Partner, DynamisWhat to ask your model this week
- How does it treat deep in-slab sources? Which algorithm computed their rates, and was its behaviour ever checked on an intermediate-depth cluster rather than assumed?
- What does it hold beyond the catalogue? Does the event set contain deep scenarios larger than the largest one observed, with rates that are physically justified rather than extrapolated from a 60-year record?
- Is the hazard on rock or at the surface? And at what resolution is basin amplification modelled under the urban concentrations that actually carry the exposure?
- Can it hindcast? Can it reproduce this event, predicted response against the records of the national network, and show you the result?
The fourth is the one that separates an auditable model from a black box. Solvency II does not ask for faith in the vendor; it asks for understanding of the model and diversity of views. An event like this one is exactly the occasion to exercise both.
Our commitment
We are going to run that exercise in public, and it is already scheduled. Within three weeks we will put the 10 August earthquake through the model as a scenario, delivered at the surface, with the current generation of deep-source attenuation models, and compare it against the shaking maps of the Colombian national survey, which are built on its own network of around 119 stations. We audit against the national survey and not against the model we use as our consistency benchmark, deliberately: an audit that reuses the benchmark it is meant to test is circular. We will publish the residual maps and the median bias, where we were right and where we were not, because a model that will not be audited by reality has no business insuring anyone else's.
We are structural and earthquake engineers: the same team that designs and reviews the buildings you insure. If 10 August taught us anything, it is that the risk is not in what the catalogue shows. It is in the decades of silence it does not.
Key questions
Where did the 10 August 2026 Colombia earthquake strike?
About 5 km from San José del Palmar, on the Chocó and Valle del Cauca border, but 110 km below the surface: inside the Nazca plate as it descends beneath the Andes, not on a fault at the surface. The strongest shaking, up to intensity VIII, was recorded 60 to 110 km away in the valleys of the coffee belt, at Pereira, Cartago, Manizales and Armenia.
Why does a deep earthquake damage several cities at once?
A rupture 110 km down has no near field. The closest point of the source is already 110 km from any site, so moving 100 km across the surface only takes the distance to the source from 110 to about 145 km. The shaking spreads over an enormous area instead of concentrating, and the duration lengthens. On 10 August that meant simultaneous damage in Cali, Pereira, Manizales, Armenia and Chocó, cities up to 200 km apart. For a portfolio it is a correlated loss across several urban concentrations, the accumulation case that models calibrated on shallow earthquakes understate.
Had this deep corridor produced large earthquakes before?
Twice, and both under the coffee belt. In February 1938 a magnitude 7.0 broke about 150 km beneath Filandia, in Quindío. In November 1979 another broke 5 km from the 10 August epicentre and at almost the same depth, with a moment magnitude of 7.2, although the body-wave magnitude in the same record reads 6.4. Between them, 110 earthquakes of magnitude 4.5 or greater have occurred at 70 to 200 km depth beneath the Cauca valley since 1960. The corridor was active, well known and had already produced this size. What it had never done is test the cities that now sit on top of it.
What should an insurer ask its model after this event?
Four things: how the model treats deep in-slab sources and how their rates were computed; whether its event set contains deep scenarios above the largest one observed, with physically justified rates; whether hazard is delivered on rock or at the surface with site amplification included; and whether the model can reproduce this event, predicted response against recorded response, and show the result. The fourth question is the one that separates an auditable model from a black box.
Sources
- USGS, M7.4, 10 August 2026, 5 km S of San José del Palmar, Colombia (event us6000tjl2): origin, depth, ShakeMap intensity and alert level.
- USGS earthquake catalogue: intermediate-depth seismicity of the Cauca corridor, 1960 to 2026.
- International Seismological Centre, reviewed bulletin: the 1938 (event 902357), 1962 and 1979 (event 656068) earthquakes of the corridor, with the full set of reported magnitudes.
- Global Centroid Moment Tensor catalogue: moment magnitude 7.2 for the 1979 El Cairo earthquake.
- Hayes and others (2018), global slab model: depth to the descending Nazca plate beneath the coffee belt.
- GEM Global Active Faults Database (CC BY-SA 4.0): Cauca-Romeral fault traces.
- Colombian national design code (NSR-10): the reference against which our Colombia model is validated city by city.
- Historical events: Armenia 1999; Roncesvalles 1997; Neira 1995; El Cairo 1979.
Magnitude, depth and intensity from USGS; casualty figures are early reports and will change. Fault traces on the maps are drawn from the open global database and are indicative at this scale. Not a substitute for a formal risk report.