Ran out of capital
The final cause—not always the root cause.
IS A SKILL Evidence reviewed
A field note from the capability collapse
The model got better.Your moat didn’t.
431
DOCUMENTED VC-BACKED SHUTDOWNSSINCE 2023 · PUBLIC COHORT, NOT A CENSUS
They raised $17.5 billion. Most did not die because the code was impossible. They died because capital, timing, economics, and product–market fit failed to hold.
Now model releases are turning entire product categories into native capabilities. The company is temporary. The capability should be portable.
This is not “AI killed 431 startups.” It is evidence that fragile businesses were already failing—and thin AI products now face an unusually fast compression cycle.
01 / THE BODY COUNT
431 companies · $17.5B in equity · median $11M
Capital was the last cause. The business was the first.
Poor product–market fit
A product existed. A durable market did not.
Bad timing
The promised market failed to arrive on schedule.
Broken unit economics
Growth could not make the underlying math work.
READ THE FINE PRINT
CB Insights examined 431 public shutdowns since 2023 and identified causes for 385. Companies could cite more than one cause, so the percentages exceed 100%.
Open the dataset2025 SHUTDOWN COHORT15.9%
were AI companies
02 / THE FIRST FILTER
The wrappers went first.
In SimpleClosure’s sample of hundreds of 2025 shutdowns, AI was not collapsing. It was being selected.
THE PATTERN
Copilots, generators, assistants, and vertical SaaS whose differentiation was mostly an LLM front-end.
THE PRESSURE
Low switching costs, shifting API economics, and capabilities becoming widely available.
THE CAVEAT
AI’s share actually fell from 17.7% to 15.9%. This is a vendor cohort—not a market-wide census.
03 / MODEL CHURN
Selected frontier releases · February—July 2026
The product boundary moved nine times.
This is a selection, not every release. Each new model made yesterday’s missing capability cheaper, faster, or native.
FEB 05
Claude Opus 4.6
AnthropicFEB 19
Gemini 3.1 Pro
GoogleMAR 05
GPT-5.4
OpenAIAPR 23
GPT-5.5
OpenAIMAY 19
Gemini 3.5 Flash
GoogleMAY 28
Claude Opus 4.8
AnthropicJUN 30
Claude Sonnet 5
AnthropicJUL 08
Grok 4.5
xAIJUL 09
GPT-5.6 × 3
OpenAI04 / ACTUAL COMPANIES
Named, dated, and sourced
Not vibes. Postmortems.
Founder statements, formal filings, and reported wind-downs. Cause labels tell you what is confirmed, what is inferred, and what remains unknown.
01
MAR 2026
Yupp
SHUT DOWN$33M seed
1.3M reported users were not enough. The founders cited weak product–market fit and a model-capability landscape that had changed dramatically.
FOUNDER-CONFIRMED CAUSE02DEC 2025
Doohickey AI
SHUT DOWN5 years
Its CEO named the battle for product–market fit, technical selling, and keeping up with the emerging AI paradigm.
FOUNDER-CONFIRMED CAUSE03JUL 2025
Astra
SHUT DOWN2 enterprise clients
Long sales cycles, sensitive-data trust, confusion around AI agents, and cofounder differences overwhelmed an AI sales product.
FOUNDER-CONFIRMED CAUSE04MAY 2026
Huxe
SHUT DOWN$4.6M raised
The AI podcast app closed as similar generation features spread across larger platforms. That timing is evidence of pressure—not proof of cause.
CAUSE NOT DISCLOSED05JAN 2026
Entropy
WOUND DOWN$25M seed
Four years and multiple pivots did not produce a scalable business model. Remaining capital was returned to investors.
FOUNDER-CONFIRMED CAUSE06MAY 2026
Parker
CHAPTER 7$50–100M liabilities
The fintech filed for liquidation after acquisition talks reportedly failed. Its financing claims included a large credit facility, not only equity.
FORMAL BANKRUPTCY07FEB 2026
Orbex
ADMINISTRATIONCapital intensive
The UK launch company entered administration after failing to secure the scale-up funding its hardware roadmap required.
FORMAL LEGAL NOTICE08AUG 2025
Kyte
RECEIVERSHIPDebt financed fleet
The rental-car delivery startup defaulted on fleet debt and could not secure replacement financing. The lender liquidated the fleet.
FORMAL WIND-DOWN05 / THE THESIS
The startup is a temporary container. The capability is the thing.
Stop wrapping the model.
A login, a dashboard, and a subscription do not make a moat. If the core job collapses into one good instruction, let it collapse.
Encode the judgment.
Preserve the inputs, sequence, tools, examples, evaluation, edge cases, and the line where a human takes over.
Make it portable.
A skill should survive its first model, interface, employer, and founder. It should be readable, forkable, and testable.
“We do not need a graveyard of tiny SaaS. We need a living library of capability.”
THE CONTAINER SHIFT
Keep the outcome. Lose the overhead.
THE THIN STARTUP
A brittle container
- × Brand before behavior
- × Dashboard before decision
- × Subscription before repeat value
- × Runway attached to one interface
THE PORTABLE SKILL
A durable outcome
- + Clear instructions
- + Explicit tool access
- + Examples and evaluation
- + A result someone can inspect
06 / WHAT SURVIVES
The case for real companies
Build where the model stops.
The argument is not that every company becomes a prompt. A company earns its weight when it owns something the next release does not hand to everyone.
Distribution
A reason people find you before they open a model.
Proprietary data
A learning loop competitors cannot prompt into existence.
Trust
Permission to act where being almost right is not enough.
Workflow ownership
The place where a job begins, moves, and gets approved.
07 / TAKE A SKILL
Three products, compressed
Take the capability with you.
No signup. No fake package command. Copy the readable brief, adapt it to your tools, and make the evaluation criteria your own.
01
Compress a retention dashboard
churn-interceptor
Reads billing, product, and support signals. Returns a ranked risk list, the evidence behind it, and the next human move.
INPUTS
Stripe events / CRM notes / Support history
OUTPUT
A weekly retention brief with account-specific actions.
02
Compress a migration platform
model-scout
Tests every credible new model against your actual work—not a leaderboard—and tells you when switching is worth the disruption.
INPUTS
Golden task set / Cost ceiling / Latency target
OUTPUT
A stay, test, or migrate decision with receipts.
03
Compress a feedback repository
customer-voice
Turns calls, tickets, and lost deals into evidence. Groups recurring pain, keeps the original language, and refuses fake certainty.
INPUTS
Call transcripts / Tickets / Lost-deal notes
OUTPUT
A cited problem map for the next product decision.
YOUR STARTUP IS A SKILL