The Software Survival Record · observed 2026-09-04

Far more software is being launched than funded, and the unfunded goes quiet 21.4× faster

We observed 24,889 company domains on a single day - every Y Combinator company from the 2025-26 batches, every Show HN launch in the same window, and every vendor this network has written about - and set them against two federal records of how many companies are being started and funded at all.

LIMIT The two ends are counted by different instruments on different populations and must never be multiplied together into a single funnel probability. An EIN application is not a company, and the trajectory arm follows only the funded. This is the SHAPE of the spectrum, not a survival probability for any one founder.

26,980observations, hash-chained 12runs on the record 99.71%of dark readings reproduced on a second look 50quarters of federal filing data

Commercial interest, stated up front. This record is produced by Lattice, which operates a network of comparison and reference properties and sells placements on them. It does not sell any outcome in this dataset, and no company’s presence, absence or observed state here is or has ever been purchasable. The raw ledger is published so that this statement is checkable rather than trusted.

What was measured

A domain check is a weak signal. That is the reason to publish the raw one.

Every observation here is structural: does the domain resolve, does it serve, where does it redirect, does it still run mail. Nothing about revenue, headcount or whether anyone is still working on it.

That limit is severe and it is stated first because everything downstream inherits it. A serving website is weak evidence that a company is alive - domains outlive the companies that registered them, and a parked page renews itself for years. The inverse is weaker still: a domain can go quiet while the company behind it is perfectly healthy and simply moved.

So this is not a list of companies that died. It is a dated record of what 24,889 domains were doing on one day, published with the method and the errors attached, and the reader draws their own inference.

The field · one mark per domain

20,233 serving 2,568 excluded as inconclusive 2,088 no longer serving

Three bands, oldest cohort first: Y Combinator, funded (1,281) · vendors we have written about (1,488) · Show HN launches, indie (22,120). The speckle is not noise - it is the density of domains no longer serving, and it thickens visibly down the field.

LIMIT Marks are grouped by cohort but not sorted by outcome inside one, so density is readable as texture. This is a proportion made visible, not a map: no mark identifies anyone. The excluded marks are domains that returned a bot policy rather than an answer - 10.3% of the whole - and they sit outside every rate on this page rather than being guessed at in either direction.

Finding 1

The funded cohort and the unfunded cohort are 21.4× apart at the same age

Same observation date, same instrument, same eighteen-month launch window. The only difference is which door the company came through.

0.48% Y Combinator, 2025-26. 6 of 1,260 observable domains no longer serving. Every batch taken whole.

10.25% Show HN, 2025-26. 2,029 of 19,804 observable domains no longer serving. Every launch in the window.

LIMIT This is not the causal effect of money. Funded companies are selected for viability before they are funded, so the gap measures funding and whatever got them funded, together and inseparably. Show HN is also its own population: technical, English-speaking, developer-tools-heavy, and many rows are projects rather than businesses. A plumber’s booking software appears in neither.

Finding 2 · the curve

Attrition is not spread evenly across the first two years

Each Show HN launch has a known launch date and one observation date, so every domain has a known age. Asking what share of each age band is still serving produces a survival curve from a single day of observation - the cross-sectional estimate, not a cohort followed forward.

Every launch cohort, by month of entry

18 monthly cohorts · 22,120 launches · ordered oldest first

LIMIT One band per launch month. Band length is exposure time; the frayed share is what was no longer serving at the single observation. It is NOT a timeline of when anything stopped - we looked once, so no date of failure is known or implied.

Share of launch domains still serving, by age

22,120 Show HN launches, observed 2026-09-04

share still serving domains observed in each band the sharpest step, marked not smoothed

LIMIT Cross-sectional, not longitudinal - ages are known, the cohort was not followed forward. If the launch population changed across the window, this confounds age with cohort, and Show HN volume did grow sharply over exactly this period. The curve declines monotonically, which is evidence against that confound but does not eliminate it. The 13→14 month step of 8.1 points is drawn rather than smoothed: it is either a real cliff or a cohort artifact and we do not yet know which. Bands under 200 domains are excluded from every figure quoted here. A second round in thirty days is what turns this estimate into a measurement.

Finding 2b · the question this record was started to answer

Leading with AI barely moves the needle. Which door you came through moves it 21.4×.

Registered before any outcome was seen: do AI launches go dark faster than the rest? Either answer was publishable - if yes, the “AI slop” thesis gets its first number; if no, the failure narrative is overstated. The answer is mostly no.

11.0%Launches that lead with AI in the title - 25.3% of the cohort - no longer serving, adjusted for launch month

10.0%Launches that do not - no longer serving, same adjustment

A gap of about one point on a base of ten, in the direction of the slop thesis, and it reverses in 8 of 18 launch months. Set against the 21.4× gap between funded and unfunded, whether a launch says “AI” is close to irrelevant to whether its domain is still serving.

LIMIT This compares launches that LEAD WITH AI IN THE TITLE against launches that do not - not AI companies against non-AI companies, because a short title under-reports. The gap is about one point on a base of ten, and the direction reverses in 8 of 18 launch months. Whatever it is, it is roughly twenty times smaller than the funded-versus-unfunded gap. Crude and month-adjusted figures agree (NaN% vs NaN% crude), so launch age is not doing the work.

Finding 3 · the one that needs no time series

The public record and the infrastructure disagree, in both directions

Y Combinator publishes a status for each company. We observed each domain independently on 2026-09-04. Holding both signals at once is the whole finding, and it was available on day one.

Companies Y Combinator lists as inactive

10 of 12 still serve a live site. In the other direction, 4 of the 1,259 companies listed as active have no address record at all.

LIMIT This measures the gap between a public record and infrastructure. It is not a claim that any company here is operating or has stopped - a serving page proves neither, which is the point. Both quantities are small: 12 and 4. They are reported because nobody publishes either direction, and they are reported with their denominators so nobody has to take the proportion on trust.

Finding 4 · two federal records, opposite directions

Applications to start a technology business are up sharply while the number raising money is down

The US Census Business Formation Statistics count applications for an employer identification number - every new business needs one, funded or not. The SEC’s Form D counts companies raising disclosed private capital, which every one of them is legally required to file. Same window, same country, opposite directions.

2019 → 2025	2019	2025
Publishing, incl. software publishers	15,310	37,130	+143%
Computing infrastructure, data processing, hosting	3,810	6,490	+70%
Professional, scientific and technical services	457,160	764,250	+67%
Telecommunications	9,670	9,660	-0%
Web search portals and other information services	1,520	1,100	-28%
Tech companies raising disclosed capital (SEC Form D)	4,789	4,186	-13%

Applications naming software publishing have more than doubled. The number of technology companies raising disclosed capital is below every year of the 2010s - the 6-year baseline never fell under 4,787, and 2024 came in at 4,090. Over the same period the number of investment funds filing rose from 19,191 to 36,654.

LIMIT An EIN application is an intent to operate, not a going concern, and the professional-and-technical category is broad enough to hold consultancies alongside software firms. These are applications, never “startups founded”. Checked for the obvious artifact: if companies were drifting into Form D’s catch-all category the decline would be reclassification rather than reality - that category is flat across the window. Both readings of the divergence are a story, and we do not know which is true: either the boom is thinner than it sounds in the formal economy, or it is happening almost entirely outside the funding system.

Finding 5 · the other half

Of the companies that do raise, 47% come back for more

Every US technology company that filed a Form D in the window, with its full filing history read from the SEC directly: 4,798 of 4,798 resolved, 0 failed. A filing is documentary, dated proof a company was alive on a specific day - the strongest signal in this record by some distance.

2,541raised once and have not filed again

2,257raised more than once

146reached a public-markets track, 3.0% of the cohort

LIMIT Presence of a filing is strong evidence of life on a dated day. Its absence is weak evidence of anything - most companies raise once and a company that stops raising has not necessarily stopped operating. “Raised once” is never an outcome. And survival is not success: this record cannot see revenue, and it can never evidence that anyone got rich.

Against ourselves

Five things we got wrong, and how each was caught

These are published because a record whose errors are invisible cannot be checked. Each was caught before this page existed, and each changed a number we had already written down.

An empty response is not a dead site. Every EMPTY_RESPONSE reading in the record turned out to be amazon.com, amazon.de and amazon.co.uk returning an empty body to our probe. Nought of three true positives. The signal is now recorded as inconclusive and excluded from every rate; the original rows stay in the ledger unedited.
A partial pull is a skewed sample, not a small one. An over-eager run against the SEC was rate-limited into a 3% subsample - whichever requests won the race - and it said 63.1% of companies raise more than once. The complete run says 47.0%. It was wrong by a third and looked entirely plausible. The instrument now refuses to write any distribution below 90% resolved.
A search cap looks exactly like a flat line. Monthly queries for launches returned precisely 1,000 every month and reported a single page of results, so the request looked successful. Adaptive windows recovered 22,120 domains against 8,304 - the cap had been costing three fifths of the cohort.
A second signal is worth different amounts in different populations. Of the 53 vendor domains read as no longer serving, 25 still run working mail servers - 47.2%, which more than halved that cohort’s figure. The same check moved the launch cohort by only 11.4%, because an indie project domain frequently never had mail to lose. “No mail” means something different in each.
A redirect is not a disappearance. Of the 74 redirects in the vendor cohort, at least 39.2% are the same company changing address, not a company being absorbed or sold; in the launch cohort at least 25.4%. Both are lower bounds - the test compares domain names and cannot see an abbreviation, so a company shortening its name still counts as a move.

Every correction made the numbers smaller. That is the direction corrections run when the instrument is being checked rather than defended.

Method

How to check any of this

Two structural signals, never one alone: an address-record lookup and a fetch, plus a mail-server lookup as a cross-check. A fetch that returns a bot policy - a 403, a 429, a refused connection, an empty body from a healthy host - is recorded as inconclusive and excluded from every denominator, never counted as an absence. We did not spoof a friendlier user agent to get past those walls, which would have raised the observable count and cost the record the only thing that makes it worth reading.

Every observation is one line in an append-only, hash-chained ledger: domain, date, signal, and the raw evidence behind it. Corrections are supersessions, never edits - a domain that goes quiet and returns keeps both transitions visibly, and a wrong reading is superseded by a later one that cites what it replaces. Nothing is deleted, which is why the errors above are still in the file.

Dark readings were re-observed the same day: 2,082 of 2,088 reproduced (99.71%). That rules out a momentary resolver blip. It does not rule out a multi-day migration, and it is a reproduction rate, never a survival measurement.

Sources, all public and free: US Census Business Formation Statistics · SEC Form D quarterly datasets and filing histories · the Y Combinator public company directory · the Hacker News search API · UK Companies House bulk data. No paid database is used anywhere in this record.

The record itself

Download everything this page was derived from

Every figure above comes from these files, and nothing else. Open them, re-derive a number, find the mistake we missed. Licence CC-BY 4.0.

ledger.jsonl - the append-only, hash-chained observation ledger, 26,980 observations, one line each with the raw evidence. Also as observations.csv.
_cohort-b-inception.json and _cohort-c-unfunded.json - the funded and unfunded cohorts, with launch dates and, for YC, the declared status we compared against.
_formation-series.json, _bfs-series.json, _trajectories.json - the SEC and Census arms, and every funded company’s filing history.
_publication-figures.json - every number on this page with its caveat attached, plus _h2-arm-c.json, _survival-curve-c.json and the cross-checks. README.

What this cannot tell you

The questions this record does not answer yet

Whether the 2025-26 cohort fails faster than earlier ones. Nothing in it has had time to fail under observation. That needs rounds, not analysis.
Whether AI companies fail faster than the rest. The record can compare launches that say AI in the title against those that don’t (above); it cannot see what a company is actually building, and the funded cohort has too few dark domains to test at all.
Whether any specific company is alive. The record states what a domain did on a date. It never states what a company is.
What happens to the ninety-nine per cent. Fewer than one in a hundred technology business applications raises disclosed capital. Every existing tracker watches the funded slice. The unfunded cohort here is the first attempt at the rest, and it is skewed toward the technically-inclined end of it.