Real example
See four answers become one
An example of how Synero answers this kind of question: four independent answers, then one combined answer. In the app you pick 1 to 4 of 20 models, and the Agreement panel shows where they agree and where they split.
Venture-Backed Startup Failure Rates: Statistical Analysis
The Commonly Cited Statistic
The most frequently cited claim is that 90% of startups fail. A related claim is that 75% of venture-backed startups fail within five years. Let us examine what these numbers actually mean and how reliable they are.
What the Data Actually Shows
Source Analysis
| Source | Claim | Methodology | Reliability |
|---|---|---|---|
| Shikhar Ghosh (Harvard, 2012) | 75% of VC-backed startups fail to return investor capital | Analyzed 2,000 VC-backed companies (2004-2010) | Moderate-High |
| CB Insights | 70% of tech startups fail (broadly) | Aggregated post-mortem analysis | Moderate (selection bias) |
| Startup Genome (2019) | 90% of startups fail | Self-reported survey data | Low-Moderate (definitional issues) |
| Cambridge Associates | ~60% of VC investments fail to return 1x capital | LP returns data, large dataset | High (financial data-based) |
| Bureau of Labor Statistics | ~50% of businesses survive 5 years | All businesses, not VC-specific | High (comprehensive data) |
Defining "Failure"
The reported failure rate changes dramatically based on definition:
| Definition of Failure | Approximate Rate (5 years) |
|---|---|
| Company ceases to exist | ~50-60% |
| Fails to return investor capital (1x) | ~60-75% |
| Fails to achieve target return (3x+) | ~80-85% |
| Does not achieve a successful exit (IPO/acquisition) | ~85-90% |
Critical insight: The "90% fail" statistic typically includes companies that return some capital, operate as modest businesses, or are acqui-hired, outcomes that are "failures" by VC return standards but not necessarily by entrepreneurial or economic standards.
The Most Reliable Estimate
The most methodologically rigorous data comes from Shikhar Ghosh's Harvard research and Cambridge Associates' LP data:
- ~60% of VC-backed startups fail to return 1x investor capital within 5 years
- ~75% fail to return the target return expected by VCs
- ~30-40% result in total loss of invested capital
Factors Affecting Reliability
- Survivorship bias in reporting: Failed companies are less likely to be documented
- Definition inconsistency: "Failure" means different things across different studies
- Vintage variation: Failure rates vary significantly by economic cycle (companies started in 2006 vs. 2009 had very different outcomes)
- Stage variation: Seed-stage startups fail at higher rates than Series B+ companies
- Sector variation: Biotech, hardware, and capital-intensive startups have different failure profiles than software
Conclusion
The commonly cited "90% failure rate" is an oversimplification that combines all startups (not just VC-backed) with the broadest possible definition of failure. For venture-backed startups specifically, approximately 60-75% fail to return invested capital within five years, with 30-40% resulting in total capital loss. These figures are directionally reliable but should be cited with their specific definitions and methodological context.
Startup Failure Statistics: What We Know, What We Think We Know, and What We Are Guessing
This question sits at an interesting intersection of data and narrative, and I want to be transparent about the epistemological quality of what passes for "startup failure statistics."
The Honest Starting Point
The commonly cited claim, "90% of startups fail", has become so ubiquitous that most people treat it as established fact. It is not. It is a rough estimate built on inconsistent definitions, self-selected samples, and methodological shortcuts. The real picture is more nuanced and less certain than the clean number suggests.
What the Research Actually Shows
The Ghosh Study (Most Frequently Cited for VC-Backed)
Shikhar Ghosh at Harvard Business School analyzed approximately 2,000 venture-backed companies between 2004 and 2010. His finding: about 75% of VC-backed startups fail to return investor capital. This is the most widely cited figure for venture-backed failure specifically.
But the finding is more nuanced than the headline:
- Ghosh found that the failure rate ranged from 30% to 95% depending on how "failure" was defined
- At the strictest definition (company ceases to exist), the rate was closer to 30-40%
- At the VC-specific definition (fails to return invested capital), it was ~75%
- At the aspirational definition (fails to meet projected returns), it was ~95%
This range matters enormously. A company that returns 80 cents on the dollar is a "failure" by VC standards but an outcome many entrepreneurs would not characterize that way.
The Bureau of Labor Statistics (Broadest Data)
The BLS tracks all business survival, not just startups. Their data shows approximately 50% of all businesses survive five years. This is often conflated with the startup-specific statistic despite measuring a fundamentally different population.
Cambridge Associates (Best Financial Data)
Cambridge Associates tracks actual VC fund returns. Their data suggests approximately 60% of individual VC investments fail to return 1x capital. This is probably the most reliable figure because it is based on actual financial outcomes rather than surveys or self-reported data.
Why the Commonly Cited Statistics Are Unreliable
-
Definition problem: There is no consensus definition of "failure." Ceasing operations, failing to return capital, failing to exit, and failing to meet projections are vastly different outcomes grouped under one word.
-
Sampling bias: Most studies either analyze post-mortems (which overrepresent dramatic failures) or use self-reported data (which introduces reporting bias in both directions).
-
Vintage effects: Startups founded during boom periods have different failure profiles than those founded during downturns. Aggregating across economic cycles obscures this variation.
-
Stage effects: A pre-seed company and a Series C company have radically different failure probabilities, but they often get aggregated in the same statistics.
My Assessment of the Statistics' Reliability
| Statistic | Reliability | Why |
|---|---|---|
| "90% of startups fail" | Low | No consistent source, combines definitions |
| "75% of VC-backed startups fail" (Ghosh) | Moderate | Reasonable methodology, specific to 2004-2010 vintage |
| "~60% fail to return 1x" (Cambridge) | Moderate-High | Based on financial data, not surveys |
| "50% of businesses survive 5 years" (BLS) | High | Comprehensive data, but measures all businesses |
The Bottom Line
The honest answer is: we do not know the failure rate of venture-backed startups with precision. The best estimate is that 60-75% fail to return investor capital within five years, but this range depends on definitions, economic conditions, and stage of investment. The "90%" figure should be cited with significant caveats, if at all.
The Startup Failure Rate: Archaeology of a Statistic
The "90% of startups fail" statistic is one of the most cited numbers in entrepreneurship, and one of the least examined. Let me trace its origins and assess its reliability through an archaeological lens.
Layer 1: The Sedimentary Origin
Like geological sediment, the "90%" figure has accumulated over time from multiple, inconsistent sources:
- Bureau of Labor Statistics: ~50% of all businesses fail within 5 years. But "all businesses" includes restaurants, dry cleaners, and sole proprietorships, not the same population as venture-backed tech startups.
- Shikhar Ghosh (Harvard, 2012): 75% of VC-backed startups fail to return investor capital. This is the closest to a rigorous estimate for the VC-specific question.
- Startup Genome Report: 90% of startups fail. Based on self-reported survey data with significant selection bias.
The "90%" appears to be a composite fossil, formed by blending these different sources across different definitions and populations.
Layer 2: The Definition Fault Line
The most fundamental problem is that "failure" is not one thing. It sits on a spectrum:
| "Failure" Definition | Rate (VC-Backed) | What It Actually Means |
|---|---|---|
| Company dies completely | ~30-40% | Total loss, nothing recovered |
| Fails to return 1x capital | ~60-75% | Investors lose money |
| Fails to meet projected returns | ~80-90% | Underperforms VC expectations |
| No IPO or major acquisition | ~85-95% | Did not produce the headline exit |
A company that returns 50% of invested capital, employs 20 people for 5 years, and gets acqui-hired is a "failure" by every VC metric. Is it actually a failure? That depends on whose lens you use.
Layer 3: The Survivor Island Problem
Imagine you are studying the population of a volcanic island. You only survey the animals that survived the last eruption. Your data will dramatically underestimate the island's danger.
Startup failure statistics face the same problem in reverse: the failures are harder to study than the successes. Failed companies do not publish post-mortems. Their founders move on. Their data evaporates. What we study are the failures dramatic enough to leave traces (CB Insights post-mortems) and the successes visible enough to generate data.
The "middle band", startups that quietly wound down, returned partial capital, or pivoted into modest businesses, is largely invisible to research.
Layer 4: The Vintage Effect
Startup outcomes follow economic weather patterns. A seed-stage company founded in 2006 (pre-financial crisis) had a fundamentally different survival probability than one founded in 2009 (post-crisis, beginning of a long bull market) or 2021 (peak of the zero-interest-rate era).
Aggregating across these vintages produces a number that is technically an average but practically misleading, like saying the average weather in a city is "comfortable" when it ranges from blizzards to heat waves.
The Honest Assessment
The most defensible estimate for venture-backed startups failing to return investor capital within five years is 60-75%, based primarily on Cambridge Associates financial data and Ghosh's Harvard research. The "90%" figure is poorly sourced and conflates definitions. The true rate depends on how you define failure, the economic vintage, the stage of investment, and the sector.
The commonly cited statistics are directionally correct (most VC-backed startups do not return capital) but precision is an illusion, the exact number is unknowable with current data.
The "90% of Startups Fail" Claim: A Statistic in Search of a Source
This is one of the most repeated statistics in business and one of the least examined. Everyone cites it. Almost nobody can tell you where it comes from.
Where Does the 90% Come From?
Short answer: nowhere definitive. It is a zombie statistic, undead, unsourced, and unkillable.
The number appears to be a frankenstein composite of:
- Bureau of Labor Statistics data showing ~50% of all businesses fail in 5 years (not VC-backed startups specifically)
- The Startup Genome Report claiming 90% (based on self-reported surveys with massive selection bias)
- Various venture capitalists saying "9 out of 10" in interviews (anecdotal, not data)
What the Actual Data Shows
The most credible number: ~60-75% of VC-backed startups fail to return investor capital within 5 years.
Sources:
- Shikhar Ghosh (Harvard): Analyzed ~2,000 VC-backed companies. Found 75% fail to return capital. Found 30-40% lose all invested capital.
- Cambridge Associates: ~60% of individual VC investments fail to return 1x. This is based on actual fund data, not surveys.
But here is the thing everyone ignores: the failure rate depends entirely on what you call failure.
- Company dies: ~30-40%
- Investors lose money: ~60-75%
- Did not achieve projected returns: ~80-90%
- No unicorn exit: ~95%+
The "90%" number only works if your definition of success is "became a billion-dollar company." By that standard, 99% of everything fails.
Why the 90% Number Persists
Two reasons:
-
It serves VC narrative: VCs use it to justify their model. "Most startups fail, which is why you need us, and why our winners need to return 100x." The high failure rate justifies high return expectations.
-
It serves media narrative: "90% fail" makes startups sound dramatic and heroic. It feeds the mythology of the entrepreneur as risk-taking maverick.
Both groups benefit from an inflated number. Neither has much incentive to correct it.
The Honest Assessment
| Claim | Reliability |
|---|---|
| "90% of startups fail" | Low, composite from inconsistent sources |
| "75% of VC-backed startups fail to return capital" | Moderate, based on Ghosh's research |
| "60% of VC investments fail to return 1x" | Moderate-High, based on financial data |
| "50% of all businesses fail in 5 years" | High, BLS comprehensive data |
The bottom line: most VC-backed startups do fail to return capital. The exact number is somewhere around 60-75%. The "90%" figure is sloppy and should be retired, but it will not be, because it is too useful as a narrative device.
Confidence in the commonly cited statistics: Moderate at best. The directional claim (most fail) is right. The specific number (90%) is poorly sourced.
Combined answer
Venture-Backed Startup Failure Rates: Synthesized Analysis
The Commonly Cited Claim
The assertion that "90% of startups fail" or that "75% of venture-backed startups fail within five years" has become ubiquitous in entrepreneurship discourse. These figures are directionally correct, most venture-backed startups do fail to return investor capital, but the specific numbers are less reliable than their confident citation suggests.
What the Data Actually Shows
The Most Reliable Sources
| Source | Finding | Methodology | Reliability |
|---|---|---|---|
| Cambridge Associates | ~60% of VC investments fail to return 1x capital | Actual fund returns data across large portfolio | Highest |
| Shikhar Ghosh (Harvard, 2012) | ~75% fail to return investor capital | Analysis of ~2,000 VC-backed companies (2004-2010) | Moderate-High |
| Bureau of Labor Statistics | ~50% of all businesses survive 5 years | Comprehensive government data, all business types | High (but different population) |
| Startup Genome Report | 90% of startups fail | Self-reported survey data | Low-Moderate |
The most defensible estimate for venture-backed startups specifically: approximately 60-75% fail to return invested capital within five years, with 30-40% resulting in total loss of invested capital.
The Definition Problem
The reported failure rate changes dramatically based on what "failure" means:
| Definition | Approximate Rate |
|---|---|
| Company ceases to exist entirely | ~30-40% |
| Fails to return 1x invested capital | ~60-75% |
| Fails to achieve target VC returns (3x+) | ~80-85% |
| No IPO or major acquisition | ~85-95% |
A company that returns 80 cents on the dollar, employs 20 people for five years, and is eventually acqui-hired is a "failure" by every VC metric. Whether it constitutes actual failure depends on whose perspective you adopt. The "90%" figure typically applies only when success is defined as achieving the venture-scale outcome (major acquisition or IPO) that VCs require to make their fund model work.
Why the Statistics Are Unreliable
Sampling bias
Most studies either analyze post-mortems (which overrepresent dramatic failures) or rely on self-reported surveys (which introduce reporting bias). The "middle band" of startups, those that quietly wound down, returned partial capital, or pivoted into modest businesses, is largely invisible to research.
Vintage effects
Startup outcomes vary significantly by economic cycle. Companies founded in 2006, 2009, and 2021 had fundamentally different survival probabilities due to macroeconomic conditions. Aggregating across vintages produces an average that may not represent any particular cohort.
Stage effects
Pre-seed companies and Series C companies have radically different failure probabilities, but they are often aggregated in the same statistics.
Definition inconsistency
No consensus definition of "failure" exists across studies, making direct comparisons unreliable.
Why the 90% Figure Persists
The inflated number serves multiple narrative purposes:
- For VCs: It justifies the portfolio model (many failures, few massive winners) and return expectations
- For media: The dramatic failure rate feeds entrepreneurship mythology
- For founders: It serves as both cautionary wisdom and badge of courage
None of these groups has strong incentive to replace a compelling round number with a nuanced range.
The Calibrated Assessment
The honest answer: we do not know the failure rate of venture-backed startups with precision. The best estimate is that 60-75% fail to return investor capital within five years, based primarily on Cambridge Associates financial data and Ghosh's Harvard research. The "90%" figure should be cited with significant caveats if at all, it conflates definitions, populations, and time horizons in ways that sacrifice accuracy for memorability.
The directional claim (most venture-backed startups fail to return capital) is well-supported. The specific commonly cited numbers are poorly sourced and precision is largely illusory.
How Synero works
1Ask your question
Your question goes to GPT, Claude, Gemini and Grok. Each answers on its own.
2Get one answer
Synero writes one combined answer from the best of all four.
3See where they split
Agreement is a good sign. Disagreement tells you what to check.
Benefits
How multi-model analysis creates natural confidence scoring
Natural Confidence Signal
When four independent AI models reach the same conclusion using different reasoning, that's a strong signal of reliability. When they disagree, that's an equally valuable signal that the answer involves genuine uncertainty.
Agreement Mapping
The synthesis explicitly identifies which parts of the answer all four models agree on, which parts have majority agreement, and which parts are genuinely contested. This granularity helps you know exactly which claims to trust.
Uncertainty Transparency
Single AI models hide their uncertainty behind confident-sounding language. Synero makes uncertainty visible, showing you exactly where knowledge is solid and where the models are effectively guessing, so you can allocate your verification effort wisely.
Calibrated Decision-Making
With confidence information, you can make appropriately calibrated decisions: act decisively on high-consensus answers, invest in additional research for contested ones, and avoid over-relying on AI for genuinely uncertain topics.
See it in 30 seconds
FAQ
Questions about AI confidence and reliability
Does Synero actually produce a numerical confidence score?
Synero provides something more useful than a number: it shows you where independent AI models agree and disagree. Unanimous agreement across different architectures and training sets is a strong confidence signal. The agreement panel explicitly maps areas of consensus and divergence, giving you a richer understanding of reliability than any single number could.
Why can't a single AI model tell me how confident it is?
A single AI model's 'confidence' is poorly calibrated: models tend to sound equally certain whether they're correct or hallucinating. They can't meaningfully distinguish between 'I know this well' and 'I'm generating plausible-sounding text.' Multi-model agreement provides an external check that the model's own self-assessment cannot.
What does it mean when the models disagree?
Disagreement between models is valuable information. It can mean the topic is genuinely debated in the field, the question involves recent developments that some models know about and others don't, the answer depends on definitions or assumptions that the models interpret differently, or one or more models are hallucinating. The synthesis helps you determine which explanation applies.
Is this approach validated by research?
The principle that aggregating multiple independent predictions outperforms individual ones is well-established in statistics and machine learning: it's the foundation of ensemble methods. Applied to language models, research shows that multi-model approaches reduce error rates and improve factual accuracy compared to single-model outputs.
Know when to trust AI, and when not to
Get built-in confidence signals from four independent AI models.
Start free3 free questions, no card. Then $10/month with $10 of usage included, or pay as you go with credit packs from $5, no subscription. Private by default, never used for training. Pricing · How it works · More use cases