Direct answer: sometimes a number in that range can be calculated, but there is no single timeless “startup failure rate.” For US employer establishments, federal data imply that about 50.8% were no longer operating after five years, 66.2% after ten years and 74.4% after fifteen years. Those figures change with the cohort, and closure is not automatically the same as failure.
The familiar statistic sounds useful because it compresses uncertainty into one clean warning. It also removes the information you need to make a decision: six to eight out of ten which businesses, over what period, and using what definition of failure?
BYBI verdict
Misleading unless qualified. A statement such as “six to eight out of ten startups fail” should not be used without naming the population, geography, cohort, time horizon and outcome. It can be directionally compatible with some long-horizon closure data, but it is not a universal probability for a new founder.
Confidence: high that the unqualified version is not decision-safe; moderate when transferring US employer-establishment data to other populations such as solo businesses, venture-backed technology companies or businesses outside the United States.
What the strongest US data actually say
The US Small Business Administration Office of Advocacy reports the following average survival rates for new employer establishments from 1994 to 2021:
| Time since start | Survived | Implied not surviving |
|---|---|---|
| At least 2 years | 67.9% | 32.1% |
| At least 5 years | 49.2% | 50.8% |
| At least 10 years | 33.8% | 66.2% |
| At least 15 years | 25.6% | 74.4% |
The right-hand column is simple subtraction from 100%. It is not a separate federal “failure” measure.
The Bureau of Labor Statistics also shows that five-year survival changes by birth cohort. Its published examples range from 49.8% for establishments born in 2006 to 57.3% for those born in 2018. Economic conditions matter; there is no single fixed five-year constant.
The safest summary is therefore:
In US employer-establishment data, roughly half do not survive five years, and roughly two-thirds do not survive ten years. The rate varies by cohort, and survival is not the same outcome as founder or investor success.
Why “startup” changes the denominator
“Startup” can describe very different populations:
- A newly opened employer establishment.
- A new legal entity with no employees.
- A founder-funded small business.
- A venture-backed company designed for rapid scale.
- A side project or self-employment activity.
- A new location belonging to an older firm.
The federal survival series cited above tracks employer establishments. It should not be relabelled as a clean failure probability for every technology startup, every new company or every person considering self-employment.
This is not pedantry. Different populations have different financing, risk, closure incentives and success criteria. A venture-backed company may be called a failure if it returns less than investors expected even when it continues operating. A small business may close because the owner retires, sells assets or accepts employment. A project can stop without bankruptcy, and a surviving company can still be a poor outcome for its owner.
Closure, failure and investment loss are different events
A useful statistic needs a defined event. Common events include:
- The establishment stops operating.
- The legal entity dissolves.
- The business enters bankruptcy.
- The founder loses money.
- Investors fail to receive their target return.
- The company misses a growth target.
- The product or business model is abandoned.
These events overlap, but they are not interchangeable. Federal establishment survival data are strong evidence for operating survival. They do not automatically tell you whether the founder regretted the attempt, whether creditors were repaid or whether an investor earned an acceptable return.
Why famous “reasons startups fail” lists do not give a failure rate
Post-mortem collections can be useful for understanding mechanisms. They are usually poor denominators for estimating a population rate.
CB Insights, for example, maintains a curated collection of hundreds of notable startup failure post-mortems. That corpus can surface repeated themes such as cash, market demand, team conflict and competition. It does not represent every startup that began, so it cannot tell us what percentage of all startups fail.
This is a selection problem: the collection begins with known failures and then analyses their reported reasons. It does not begin with a representative cohort of all new businesses and follow each one to an agreed outcome date.
Run the BYBI Denominator Check
Before you repeat or act on a failure statistic, fill in all six fields.
| Field | Question | Example of an adequate answer |
|---|---|---|
| Population | Which businesses count? | US private-sector employer establishments born in 2018 |
| Geography | Where are they located? | United States |
| Cohort | When did they start? | 2018 birth cohort |
| Time horizon | How long were they followed? | Five years |
| Event | What counts as the outcome? | Establishment no longer operating |
| Source | Who measured it and how? | BLS Business Employment Dynamics |
If any field is missing, label the number as incomplete rather than turning it into a universal warning.
Five ways an apparently precise failure rate changes
1. Moving the observation date
Survival is a curve, not a single event. A business can survive two years and close before year five. That is why the same underlying cohort can support a low early non-survival share and a much higher long-horizon share. Quoting the ten- or fifteen-year result as if it were a near-term launch risk removes the calendar from the evidence.
2. Changing the cohort
Businesses born before or during a recession can face different conditions from those born in a recovery. The BLS cohort examples show meaningful five-year variation. A useful source therefore gives a birth year or period, not just a pooled statistic.
3. Changing the unit of analysis
An establishment is a physical or operational location. A firm can operate more than one establishment. The closure of one location is not necessarily the death of the firm, and the survival of a legal entity does not prove that its original product or strategy survived.
4. Changing the outcome
Operating survival, profitability, founder income, bankruptcy and investor return answer different questions. If an article switches among them, the percentage stops having a stable meaning.
5. Changing who is eligible
Employer-business datasets exclude some solo and informal activity. Venture datasets may begin only after a funding event. Accelerator cohorts include selected founders. Each eligibility rule changes the denominator before any outcome is measured.
Match the statistic to the decision
The right denominator depends on what you are deciding.
| Decision | Useful population | Useful outcome | Evidence to avoid |
|---|---|---|---|
| Should I leave a job to open a local employer business? | Comparable new employer firms in the same industry and region | Survival, owner income, cash loss and hours over the runway | Venture-capital portfolio anecdotes |
| Should a fund invest in an early-stage technology company? | Comparable funded companies by stage, vintage and sector | Exit, write-off, return multiple and time to liquidity | All-small-business survival alone |
| Should I test a side project? | Similar low-cost projects with the same time commitment | Validated demand, cash contribution and opportunity cost | Bankruptcy rates |
| Should a policymaker support new establishments? | Official establishment cohorts | Survival, employment, job creation and destruction | A selected founder post-mortem corpus |
This mapping prevents a true statistic from answering the wrong question.
A calculation audit anyone can reproduce
When a source reports a survival rate, calculate the corresponding non-survival share as:
100% − survival rate = share not surviving to that date
For the SBA five-year figure:
100% − 49.2% = 50.8%
That arithmetic is simple. The interpretation is not. “Not surviving as an employer establishment to five years” should remain attached to the result. Replacing it with “failed startup” adds assumptions that the calculation did not measure.
Also resist multiplying or averaging rates from different studies unless they use compatible populations, dates and events. A precise decimal produced from incompatible inputs is still a weak estimate.
A stronger way to write the claim
Weak:
Eight out of ten startups fail.
Stronger:
In SBA calculations using US employer-establishment data for 1994–2021, 49.2% survived five years and 33.8% survived ten years. These figures measure establishment survival, not every definition of startup success.
The stronger version is less dramatic and far more useful.
What this statistic can—and cannot—tell a founder
It can tell you that operating survival declines materially over time and that planning only for the launch year understates the risk. It can also justify questions about cash runway, demand validation, owner capacity and the conditions that would trigger a change in plan.
It cannot give you a personal probability without more information. Industry, financing, founder experience, starting resources, geography, cohort conditions and the chosen outcome all matter. A population average is a baseline, not a destiny.
Use the statistic to improve a decision process:
- Define what success must look like by a specific date.
- Define the loss, debt or time commitment you can tolerate.
- Track leading evidence—customer behaviour, contribution margin and cash—rather than waiting for a binary survival result.
- Specify stop, adapt and continue thresholds before sunk cost makes them harder to use.
The practical takeaway
Do not use a dramatic failure rate to decide whether entrepreneurship is “worth it.” Use a relevant base rate to design the downside: the runway you need, the evidence milestones you will demand, the losses you can absorb and the conditions under which you will stop or change course.
Next step: run the Denominator Check on the statistic that influenced your decision, then save the completed claim and source trail. Subscribe only if you want the evidence record when the official series updates.
What would change this verdict?
The verdict would change if a speaker or publisher supplied a clearly defined population-level dataset showing a stable six-to-eight-in-ten outcome for the exact startup population, time horizon and failure definition being discussed.
It would not change because another article repeats the statistic, because a post-mortem list contains many failures or because a successful founder says the risk is high. Repetition and authority do not repair a missing denominator.
Sources and update record
- US Small Business Administration Office of Advocacy, Frequently Asked Questions About Small Business, July 2024.
- US Bureau of Labor Statistics, Business Employment Dynamics twentieth anniversary.
- US Census Bureau, Business Dynamics Statistics.
- CB Insights, Startup failure post-mortems, used for qualitative mechanism context only.
Last evidence check: August 22, 2026. Review when the SBA, BLS or Census series is materially updated, or when a quoted speaker supplies a different denominator.
Frequently asked questions
What percentage of startups fail in the first year?
There is no universal percentage for every startup population. Use a source that names the cohort and survival event. Do not derive a first-year number from a long-horizon statistic.
Is business closure the same as failure?
No. Closure may accompany bankruptcy or loss, but it can also follow a sale, retirement, project change or voluntary exit. The source’s event definition must be retained.
Do 90% of startups fail?
That statement is not decision-safe without a defined population, period and outcome. It should not be treated as a universal fact.
Are venture-backed startups included in the SBA survival statistic?
The cited statistic covers US new employer establishments. It is not a clean venture-backed-only cohort, so it should not be presented as one.
What is a better question than “What is the startup failure rate?”
Ask: “For businesses like this one, in this market and funding model, what happens by the date that matters to my decision—and what loss or outcome counts as failure?”
