Naive prefix-matching assigned 789 unrelated fund filings to startups named "Galaxy" and "Stellar." The fix: filer name must equal the company name exactly, or name + corporate-suffix tokens only. Result: 428 clean filings.
startup_
graveyard
a census of yc outcomes · 2021–2026 · receipts included
What actually happens to startups after the launch tweet? This is a census, not a highlight reel: every Y Combinator company from the 2021–2026 batches — all 3,566 of them — tracked from demo day to whatever came next, using five free public sources and a pipeline where every fact carries its source URL. It is a lifecycle study, not a failure study: the cohort holds every outcome from acquisition to wind-down, and it's that full denominator that makes the rates mean anything. Survival curves come with confidence bands, funding gets tested against death, and silence gets recorded as silence — because most startups never write a goodbye post.
select outcome, evidence from yc_companies where batch >= '2021'
recreate the analysisschema, ingesters, tests, and a citable v1.0.0 data release with sha-256 checksums.
3,566companies tracked
300shut down
197acquired
0ipos
00 / window
three regimes, one cohort
Five years is not an arbitrary slice. The 2021+ window catches three funding climates inside a single comparable population — same accelerator, same selection, same directory, only the weather changed. That is what makes cohort-to-cohort comparison legitimate here and illegitimate almost everywhere else.
the ai-wave cohorts are the largest ever and the least dead — but they are also the youngest, and a young cohort is an unfinished observation, not a healthy one. only the first two regimes have had enough time on the clock to produce a rate worth quoting.
01 / verdict
six numbers that survived the analysis
Every number below carries its denominator, its cohort window, and — where it's a rate — a Wilson 95% confidence interval. The most interesting finding is a null result, which is exactly why it's publishable.
f1
17.2%2021 cohort · 125 of 727
of the ZIRP-peak 2021 batch is gone five years later. The correction was real — but quieter than the discourse suggested.
f2
26.6% vs 7.5%consumer vs healthcare, 2021–23
What you build matters more than when. Consumer startups die at 3.5× the healthcare rate, and the confidence intervals don't touch.
f3
no differencefunding vs death · overlapping CIs
Companies with no Form D on file die at 14.2%; companies that filed $5M+ die at 11.1%. The intervals overlap. Money postpones the question; it doesn't answer it.
f4
~6 of 300shutdowns with any public post-mortem
Most startups die silently. No founder letter, no HN thread, no news story — just a website that stops answering. The failure discourse is written by the loudest 2%.
f5
3quiet acquisitions found
Homepage-notice mining caught companies listed as "inactive" that were actually acquired (Fabius, Flike, Launcher Labs). Directory labels lag reality; receipts don't.
f6
$6.3b428 form d filings · 251 companies
Fundraising events pulled straight from SEC EDGAR filings — dollar amounts from the filings themselves, not press-release rounding.
02 / survival
how cohorts actually decay
Kaplan-Meier survival per cohort year, with Greenwood 95% bands. Companies enter at their batch start (winter = jan, spring = apr, summer = jun, fall = sep); death is proxied by the last month a company's website returned HTTP 200 to the Wayback Machine; active, acquired, and public companies are right-censored at the export date — an acquisition is not a death, and acquisition dates aren't public. The 2021 cohort — funded at the peak — sits at 81.2% survival at month 66.
y-axis starts at 70% — no cohort's curve goes below it. survival at last observed month: 2021 → 81.2%, 2022 → 86.1%, 2023 → 87.1%.
| company | announced | proxy says | delta | evidence |
|---|
the proxy runs late: websites outlive companies (median +3 months; moxion power's site zombied for 20). read km lifespans as upper bounds — that's a documented bias, not a footnote.
03 / mortality
what you build decides how you die
Shutdown rate by industry, 2021–23 cohorts only (later cohorts haven't had time to die yet). Whiskers are Wilson 95% intervals — small industries get honest, wide ones.
consumer's interval [20.5%, 33.8%] clears healthcare's [4.8%, 11.7%] with room to spare — one of the few differences in this dataset that survives its own error bars.
inactive
acquired
still operating
2024–26 cohorts look immortal because they're young, not because they're better. this is right-censoring wearing a flattering outfit.
04 / money
funding doesn't buy survival
Every company matched against SEC EDGAR Form D filings — exact-name matching, ambiguity dropped, amounts from the filings. Then the obvious question: do funded companies die less? Within this window, not detectably.
all three intervals overlap. and the honest caveats: "no form d found" ≠ "raised nothing" (only us exempt offerings get filed), and selection effects — better companies raise more — likely make money's causal effect even smaller than the observed 3pp gap.
05 / the build
five sources, one duckdb file, $0
The dataset is the product. Everything below is free public data — the price is politeness delays, entity-resolution rigor, and recording every check so nothing runs twice. Any step can be re-run safely; every fact carries its source URL.
ingest (per source) ─→ resolve (canonical id + guards) ─→ enrich (llm, receipts) ─→ export ─→ dashboard │ duckdb · data/lifecycle.duckdb · 8 tables · idempotent upserts on natural keys
3,566companies · yc directory, the ground-truth denominator
7,603wayback snapshots · monthly liveness heartbeat, 283 companies
428form d filings · $6.3b, parsed from primary_doc.xml
278hn mentions · surviving the homonym filter
269final homepages read · 7 shutdown notices found
4,003timeline events · one dated stream, any source
10llm-enriched shutdowns · quote + source url mandatory
~2.5hfull pipeline runtime · politeness delays included
the sources
| source | provides | politeness |
|---|---|---|
| yc directory | The cohort itself with labeled outcomes — the denominator every rate in this report divides by. | static json |
| wayback cdx | Monthly website-liveness timeline per company; last live snapshot is the death-date proxy. | 1s delay |
| wayback snapshots | Final-homepage shutdown notices — founder-letter-grade evidence, mined and human-reviewed. | 3s + 30s backoff |
| hn (algolia) | Launch posts, shutdown threads, community post-mortems — comments say what founder letters won't. | 0.3s delay |
| sec edgar | Form D fundraising events with dollar amounts, straight from the filings. | 0.5s + declared ua |
considered and rejected: crunchbase (five figures), linkedin (tos + litigation), x (api cost). the price of $0 data is patience.
the schema
| table | contents | natural key |
|---|---|---|
| companies | Canonical company — slug, name, apex domain, batch, cohort year, status, industry, team size. | company_id |
| company_sources | Per-source identity plus the raw record: the entity-resolution ledger. | company_id + source |
| funding_rounds | Form D filings with amount_usd (totalAmountSold, falling back to totalOfferingAmount) and filing URL. | source_url |
| timeline_events | One dated stream of everything — yc_listed, funding, layoff, shutdown_notice. A company page is an ORDER BY. | — |
| wayback_snapshots | One row per company-month with its HTTP status: the liveness heartbeat. | company_id + month |
| wayback_notices | Final-snapshot checks; a null phrase means "checked, found nothing" — which is what makes the miner resumable. | company_id |
| hn_mentions | Hacker News stories that survived the homonym filter. | company_id + story_id |
| shutdown_enrichment | LLM extraction: reasons, category, confidence, and a verbatim quote + source URL, both NOT NULL. | company_id |
every ingester is idempotent — insert-or-replace on the natural key — so any step can be re-run without duplicating a single row.
rebuild it yourself
git clone https://github.com/rohitvpgawli/why-startups-die && cd why-startups-die python3 -m venv .venv && .venv/bin/pip install -r requirements.txt # duckdb, httpx, anthropic lifecycle init-db apply schema.sql lifecycle ingest-yc ~30s · 3,566 companies lifecycle ingest-wayback --status Inactive ~30min · liveness timelines lifecycle ingest-hn --status Inactive ~10min · hn mentions lifecycle ingest-edgar ~75min · form d, full cohort lifecycle ingest-wayback-notices ~20min · resumable; re-run to retry errors lifecycle enrich needs ANTHROPIC_API_KEY; skips enriched lifecycle export && dashboard/build.py export.json → self-contained dashboard
scripts/refresh.sh chains the whole thing. re-running later grows the dataset — new deaths arrive as yc flips a directory status, which is the one part of this pipeline nobody controls.
the actual hard part: entity resolution
Companies named June, Cache, Fable, and Nexus match endless unrelated news. A story counts only if the company's domain is in the URL, or the exact name appears in the title alongside startup-context tokens ("YC", "Show HN", "shut down", "raises").
A company matching more than two distinct SEC CIKs is dropped entirely. A filing matching two different YC companies (there are two "Hyper"s) is dropped for both. Recall is sacrificed for precision, on purpose, everywhere.
"Acquired by" on a homepage only counts if it's about this company — team-bio brags about other companies' exits were rejected on manual review. Human review of snippets is a pipeline step, not a nice-to-have.
Every resolution rule is pinned to the concrete failure that motivated it — Galaxy-fund pollution, Abel/June homonyms, month-name subjects. Writing the tests caught two further live filter leaks.
Claude Opus 4.8 extracts shutdown reasons under structured output, into an eleven-term controlled vocabulary — but confidence is capped by source quality (founder letter ≤ 0.9, community ≤ 0.7, dead-website inference ≤ 0.4), and a verbatim quote + source URL is mandatory. No receipt, no row.
citable artifact · releases/v1.0.0 — one csv per table, analysis json with every ci, proxy-validation sample, sha-256 checksums. the duckdb working file stays local; the release is the thing you cite.
licence · code (src, dashboard, scripts, schema, tests) under mit · data compilation under cc by 4.0. both permit commercial use; the data asks for attribution.
gawli, r. (2026). the startup lifecycle project: outcomes of y combinator startups, 2021–2026 batches. data release v1.0.0.
licence · code (src, dashboard, scripts, schema, tests) under mit · data compilation under cc by 4.0. both permit commercial use; the data asks for attribution.
gawli, r. (2026). the startup lifecycle project: outcomes of y combinator startups, 2021–2026 batches. data release v1.0.0.
06 / registry
the graveyard, row by row
All 300 companies the YC directory marks inactive — each with its lifespan, last sign of life, Form D filings, HN record, and (where the public record supports one) an evidence-linked shutdown reason. Companies, not people; outcomes, not morality plays.
07 / limitations
disclose these when citing
A dataset that hides its weaknesses is marketing. These are the load-bearing caveats.
form d ≠ funding. Only US exempt offerings actually filed appear; "no Form D found" does not mean "raised nothing."
correlation only. The funding-vs-death comparison is correlational; selection effects likely shrink money's causal role further.
labels lag. YC's own statuses trail reality — "inactive" conflates shutdown, zombie, and pivoted-elsewhere, and we caught quiet acquisitions hiding in it.
enrichment is honest, not complete. ~10 evidence-backed shutdown reasons out of 300 deaths, because that is all the public record supports.
precision over recall. Common-word company names are dropped from EDGAR rather than guessed.
the proxy runs late. Websites outlive companies (median +3 months, one 20-month zombie); KM lifespans are upper bounds.
the archive throttles. archive.org rate-limits sustained crawls, so coverage is a function of patience. Every check is recorded — including the ones that found nothing — so a re-run only visits what it hasn't seen.