🕰️ How the record ages
Every entry in this atlas carries the year its system started. Read along that axis, the record answers a question no single project page can: how much of what we know is simply a function of how long anyone has had to look.
1132 of 1167 published entries state a start year, the oldest from 1951 and the newest from 2026. Grouped into vintages, they show two things the overall totals hide: a failing share that rises sharply once a vintage is old enough to have been judged, and evidence that gets thinner the closer you come to the present. Every figure on this page is computed from the published data; none is typed.
✔ Every published entry is checked against the public sources it cites before publication.
Every vintage, and how it stands today
Bands are unequal on purpose: the first is open-ended because 153 entries predate 2010 and trimming them to a tidy grid would drop them from the record without saying so.
succeededongoingscaled backfailed
| Vintage | Entries | Outcome | Failing | States an impact figure | Arm's-length evidence |
|---|---|---|---|---|---|
| before 2010 | 153 | 26 succeeded, 114 ongoing, 11 scaled back, 2 failed | 8.5% | 72% | 27% |
| 2010–2013 | 144 | 27 succeeded, 105 ongoing, 6 scaled back, 6 failed | 8.3% | 69% | 25% |
| 2014–2016 | 182 | 43 succeeded, 109 ongoing, 24 scaled back, 6 failed | 16.5% | 70% | 21% |
| 2017–2019 | 238 | 32 succeeded, 173 ongoing, 21 scaled back, 12 failed | 13.9% | 62% | 18% |
| 2020–2022 | 217 | 22 succeeded, 179 ongoing, 13 scaled back, 3 failed | 7.4% | 62% | 15% |
| 2023–2026 | 198 | 13 succeeded, 178 ongoing, 4 scaled back, 3 failed | 3.5% | 45% | 6% |
What this table is, and is not. It is a cross-tabulation by vintage, not a survival model: entries are not followed through time, and nothing here is censored for the years a young entry still has ahead of it. Two consequences worth stating rather than hiding. The oldest band understates failure through documentation survivorship — a project that died quietly in 2006 rarely made it into any source, so it is not in this atlas to be counted. And the youngest band understates it because most of its entries have not been judged at all yet: 55% of them state no impact figure. In the outcome bars, a very small share is drawn at a minimum width so that it stays visible rather than disappearing; the exact figures are the columns beside them.
The failing share, once you allow for age
Across everything published, 112 entries carry a failing outcome — 9.6%. That number is not wrong, but it averages a vintage that started before 2010 with one that started last year, and only the first has had time to be judged.
A cutoff that is derived rather than picked: the median entry recorded as failed in this atlas ran 5.5 years before it stopped. An entry younger than that has not yet lived through the window in which failures here become visible. Among the 789 entries that started in 2020.5 or earlier, the failing share is 11.8% — roughly one in eight.
The figure barely moves when the cutoff does, which is the reason to trust it more than any single vintage: 4 years (934 entries) → 11.1% · 5 years (866 entries) → 11.2% · 6 years (789 entries) → 11.8% · 7 years (717 entries) → 12.3% · 8 years (625 entries) → 12.6%. Picking the worst-looking band instead would put it near 16.5%, and that would be a number chosen for its size. Both figures remain floors — see the failure record for why an entry only earns a failing outcome when a named source says so.
How long they last
90 entries have finished — succeeded, scaled back or failed — and record the year they stopped as well as the year they began. Their median life is 4 years — and the ones that failed are the longest-lived of all:
- succeeded — 40 entries, median 4 years
- scaled back — 26 entries, median 5 years
- failed — 24 entries, median 5.5 years
The entries still running are older than the debate around them suggests: median age 8 years, with 370 of them past their tenth year. Whatever else this record shows, it is not a catalogue of things that were switched on last season.
Why the newest entries are the thinnest
Independent and peer-reviewed evidence falls from 27% of the pre-2010 vintage to 6% of entries started in 2023 or later; impact figures fall from 72% to 45%. An audit office reports years after commissioning, and a peer-reviewed evaluation later still. The practical reading: an evidence grade is partly a statement about the calendar, and the oldest entries are where re-checking pays off — the evaluation that did not exist at publication may exist now.
Common questions
Does the smart-city failure rate change once you allow for a project's age?
In this atlas, 112 of 1167 published entries carry a failing outcome — 9.6%. That figure mixes vintages: an entry that started in 2025 has not had time to be recorded as a failure. Among the 789 entries that started in 2020.5 or earlier — the median failed entry in this record ran 5.5 years before it stopped — the share is 11.8%. Both are floors, not rates: an entry only earns a failing outcome when a named source says so, and quiet abandonment rarely has one.
Does the evidence behind a smart city project improve over time?
In this record the opposite is true of recent projects, and the reason is time rather than diligence. 27% of entries started before 2010 rest on independent or peer-reviewed evidence, against 6% of those started in 2023 or later. Independent evaluation is written years after a system is switched on, so the newest entries in any catalogue necessarily rest on the operator's own reporting.
How long does a failed smart city project run before it stops?
Longer than the word "failure" suggests. Of 90 entries here with a recorded end year, the median ran 4 years, and the ones recorded as failed ran longest at 5.5 years. A project of this kind rarely collapses; it is kept running, quietly, until someone decides to stop paying for it.
Where to read next
What didn't work groups the failing entries by how they went wrong; this page says how many of them there should be expected to be. The State of the Evidence asks what the record rests on, and this page adds when. 35 published entries state no start year and are absent from every count above — no placeholder year was invented for them.