The 18 cables that are all exactly 1,100, 1,200 or 1,300 km
Showing the misleading chart
A route-planning briefing we drew takes every submarine cable system on TeleGeography’s map that publishes a length — 634 of them, 1,949,028 km between them — and plots the lengths at one bar per 10 km below 2,000 km and 500 km above it, on a linear count axis from zero, with nothing smoothed, dropped or rescaled. The distribution comes out spiky: six systems at exactly 1,100 km, six at exactly 1,200, six at exactly 1,300, four at exactly 20,000. The desk reads the spikes as engineering and recommends scoping the next route to the nearest class. They are not engineering. A published length is a number somebody wrote down, 270 of the 634 end in a nought where 63 would if the last digit carried information, and the rounding grows with the cable: 4% of systems under 500 km sit on an exact multiple of 100 km against 68% of those over 10,000 km — so the roundest figures are the ones carrying most of the kilometres.
01The claim
Cable systems are built to standard route lengths. Every submarine cable system on TeleGeography’s map that publishes a length is on the slide — 634 of them, 1,949,028 km between them, 553 in service and 81 planned, none dropped. The count axis is linear and starts at zero, the bins are a round 10 km below 2,000 km and 500 km above it, and nothing is smoothed, logged, indexed or rescaled. Read at that resolution the distribution is not a smooth curve with a long tail: it is a field of one- and two-system bars with towers standing out of it. Six systems are published at exactly 1,100 km. Six at exactly 1,200. Six at exactly 1,300. Three at 1,000 and three at 1,400. The same step recurs at the top of the range, where four systems are exactly 20,000 km, three exactly 25,000, three exactly 15,000 and three exactly 10,000. Thirty-seven of the 634 sit on an exact multiple of 1,000 km and 147 on an exact multiple of 100. Between the classes the shelf is thin — of the 63 systems from 1,000 to 1,500 km inclusive, only 38 fall anywhere other than a round hundred. Read-out for the desk: regional trunk routes are engineered to a 100 km step and long-haul systems to a 5,000 km step. Recommendation: scope the next regional system to the nearest class, and price the long-haul option at 20,000 km rather than paying for a bespoke route survey.
02The trick
The drawing is not doing anything. One field, one unit, one population; a count axis that is linear and starts at zero; bins of one round width, the same right across each panel; every system with a published length present and nothing reweighted, smoothed or binned wide. The picture survives any audit you care to run on it, because what produced the towers is not in the picture — it is in the column. A published cable length is not a measurement of a cable. It is a number somebody wrote down, and the map records what it is given. Count the last digits and the field says so: 270 of the 634 lengths end in a nought, where 63 would if that digit carried information, and 64 end in a five. Between them the two favourite digits take more than half the set. The rest are ordinary — 42 ones, 35 twos, 34 threes, 44 fours, 32 sixes, 37 sevens, 47 eights, 29 nines — which is the tell that this is heaping rather than some property of seabeds. So the 1,100 km class is six unrelated routes: Norte Conectado (Infovia 01), which is a river system with eleven Brazilian landing points along the Amazon and no seabed at all; an inter-island link in the Bahamas; another in the Azores; the Nigeria–Cameroon crossing; and PASELA and JAKASUSI, two Indonesian domestic cables. Nothing joins them but the shape of the number. The second half of the effect is the one that generalises, and it is what makes this worse than an ordinary comb: the rounding is not uniform across the set, it scales with the thing being measured. Take the systems in bands. Under 500 km, 4% are an exact multiple of 100 km — twelve of 281, where chance alone would give about three, so even the least-rounded band is rounded; it is simply the floor the rest is measured against. From 500 to 1,000 km it is 15%. From 1,000 to 2,000, 33%. From 2,000 to 5,000, 34%. From 5,000 to 10,000, 45%. Over 10,000 km, 68% — and 36% sit on an exact multiple of 1,000. As to why, the honest answer is that the map records no provenance, so nothing in the file says whether a figure was surveyed or announced. The likely story is the obvious one — a short crossing is laid and measured, while a 45,000 km system is announced by a consortium years before a ship sails — and there is a proxy in the file worth taking seriously, because a planned system’s length can only be an announcement. Planned systems are rounder than built ones: 38% of the 81 sit on a multiple of 100 km against 21% of the 553 in service, and the gap holds in five of the six bands. Eighty-one systems is not proof. It points the same way. What is not in doubt is the consequence, which is that roundness is over-represented in the kilometres rather than spread evenly through them. The 270 systems whose length is a multiple of 10 km carry 66.6% of the 1,949,028 km. The 147 that are multiples of 100 km — 23% of the systems — carry 57.3%. The 37 multiples of 1,000 km carry 26.8%, and sixteen systems that are multiples of 5,000 km carry 16.9% on their own. Which is what makes the seven-digit total a fiction of typography rather than of arithmetic. Treat each published figure as an interval the width of its own rounding — the largest round step that divides it — and the sum moves by up to ±60,797 km, or about ±8,317 km if the roundings are independent and mostly cancel. That is 0.4% of the total, so the sum is not pushed in any direction and is not wrong; it is simply good to two or three figures where 1,949,028 prints seven. Four in five of the inputs have three significant figures or fewer, and the single largest contributor — a planned system published at 50,000 km — has one. TeleGeography itself quotes the in-service figure as “over 1.5 million kilometres”, which is two significant figures and the right number of them. Two more things the column is quietly doing. Seventy-five of the 709 systems tracked publish no length at all — 54 of them in service — so every total here is a sum over 553 of the 607 in-service systems; and 81 of the 634 that do publish one are planned rather than built, contributing 403,042 km of announced figures. (The slide is our demonstration in the manner of a route-planning briefing — the desk, the classes, the read-out and the recommendation are invented, and every figure behind them is the map’s own, retrieved on 23 September 2026. One domain note: a published route length and the cable actually laid are not the same quantity, since slack adds a few per cent on the way to the seabed.)
03The fix
Bin at the resolution the numbers actually have, not the resolution they are printed at. At 500 km bins every one of the trunk classes falls inside a single bar and the modes disappear, because there were never five populations — there was one population and a preference for round hundreds. The last-digit tally is the check that gets you there, and our blood-pressure exhibit has the argument for it and the ten-bar chart to run: count the final digit of every value and see whether each of the ten gets about a tenth. Here nought gets 270 of 634 and five gets 64. Then run the test this dataset needs and most heaped datasets do not, because it is the half that is easy to miss: split the values into bands and ask whether roundness rises with size. Where it does — 4% at the short end, 68% at the long — the granularity is correlated with the quantity, which does not bias a total in either direction, since rounding to nearest stays symmetric however coarse it gets. What it does is stop the error shrinking the way independent noise would: instead of being spread across 634 rows and cancelling, it is carried by a few dozen of them. The practical difference is not that the total is wrong but that it is good to two or three figures rather than seven, and a reader who has not run the test will happily quote all seven. So round a total to the digits its inputs can carry. “About 1.9 million km across 634 systems, of which about 1.5 million is in service” says everything 1,949,028 km said and stops claiming what it cannot support — and printing the count of systems the total is built from, against the count tracked, is the line that tells a reader 54 in-service systems are missing from it. Never put a threshold or a bin edge on a round number, either. A cut at “1,000 km and over” lands on a pile of three systems published at exactly 1,000, and whether they fall in or out is a decision about the rounding rule rather than about cable. The same goes for a rank, a quantile boundary or a histogram’s first edge. And where the last digits genuinely matter, go and get a measurement instead of a figure: the map publishes several hundred lengths to the kilometre, MAREA at 6,605 and Grace Hopper at 7,191 and Amitié at 6,792 among them, and those are what a surveyed route length looks like sitting in the same column as an announced one. That is the sentence worth carrying away, because it outlives cables: one column can hold two kinds of number, and nothing about the column says which is which. A field that is sometimes measured and sometimes declared has to be read as the second, and the last digit is usually the only place it admits it.