What percentage of startups fail? What the data actually says

There is no single percentage of startups that fail – the honest answer depends on how you define “startup”, how you define “fail” and over what period – and the famous “90% of startups fail” figure is untraceable to any primary source.

That is a less satisfying opening than a big round number, which is exactly why the big round number persists. But if you are a founder planning runway, an employee weighing an offer, or a journalist reaching for a statistic, the difference between folklore and data matters. This piece takes the “90%” claim apart, lays out the two datasets actually worth citing – one European, one American, each with serious caveats – and explains why high attrition is a designed-in feature of the venture model rather than a scandal.

The “90% of startups fail” claim: where it falls apart

Try to trace “90% of startups fail” to a primary source and you will not find one. The figure circulates through blog posts citing news articles citing other blog posts, and somewhere near the bottom of most citation chains sits either a misread of US small-business survival data or no study at all. It survives because it is memorable, flattering to survivors and useful to people selling things to worried founders.

Beyond the missing source, the claim has three structural problems. First, it blends company types: statistics about all new businesses – cafés, plumbing firms, consultancies – get relabelled as “startup” statistics, though a venture-scale technology company and a corner shop have almost nothing in common as risk profiles. Second, it blends timeframes: failure within two years and failure ever are wildly different claims, and the quoted figure floats between them as needed. Third, it never defines failure: bankruptcy, acquisition below the last valuation, a pivot to a profitable lifestyle business and a quiet shutdown all get flattened into one word. A number that specifies neither the population, the period nor the event it measures is not a statistic. It is a vibe.

What general business data actually shows – with both caveats

The most robust survival data anywhere comes from the US Bureau of Labor Statistics, which has tracked new business survival for decades. Two caveats before the numbers, and both matter: this is US data, and it covers all new businesses of every type – not startups in the venture sense. With that firmly attached, the BLS pattern is remarkably stable across cohorts and decades:

Time from founding	Share of new US businesses that have failed (BLS, all business types)
2 years	~20%
5 years	~45%
10 years	~65%

Notice what this does and does not say. Even over a full decade, and even counting every fragile new business in the economy, the failure rate reaches about 65% – not 90%. If anything, the folklore number overstates the bleakness of general entrepreneurship. What the BLS data cannot tell you is anything specific about venture-backed technology startups, which fail differently: less often from the slow insolvency that kills small businesses, more often from the deliberate, binary logic of venture funding – which brings us to the European data.

The honest European benchmark: the Dealroom seed cohort

The best published European evidence on startup progression comes from Dealroom, which followed a cohort of 3,075 European companies that raised seed funding in 2016-18 and measured how many had raised a Series A over time. The data is dated – the cohort predates the 2021 boom and the correction that followed – but it is a real cohort with a defined population and a defined event, which already puts it in rare company.

Time from seed round	Share that had raised a Series A (Dealroom, 3,075 European companies, 2016-18 cohorts)
12 months	6%
24 months	18%
36 months	27%
48 months	31%

Roughly one seed-funded European company in four reached a Series A within three years, and fewer than one in three within four – after which the curve has largely flattened. That is the closest thing Europe has to an honest progression statistic, and it is worth quoting precisely because of what it does not claim: it measures fundraising, not survival.

Not raising a Series A is not the same as failing

It is tempting to subtract and declare that 69% of European seed startups “fail”. Resist it. The non-graduates in a cohort like Dealroom’s include several very different fates: companies that reached profitability and never needed another round; companies acquired before an A, at outcomes ranging from painful to excellent; companies that downshifted into sustainable businesses outside the venture track; zombie companies alive but not growing; and, yes, genuine shutdowns. The cohort data cannot separate these, and neither can anyone quoting it. What you can honestly say is that the seed round is, statistically, the last venture round most venture-backed companies ever raise – a fact that should shape how founders plan, but which is not a death rate.

Why attrition is priced in: the portfolio maths

The deeper reason startup failure rates are high is not incompetence – it is that the venture model is built to tolerate, even require, them. A fund’s returns follow a power law: in a portfolio of thirty seed investments, the expectation is that a large share return little or nothing, a middle band returns modest multiples, and one or two outliers return more than everything else combined. The fund’s economics live or die on the outliers.

This has two consequences worth internalising. First, investors are not trying to minimise the failure rate of their portfolios; they are trying to maximise exposure to the outliers, and a portfolio with no failures is usually a portfolio that took too little risk. Second, the seed-to-A filter – three in four not passing – is the mechanism operating as designed: many cheap experiments, few expensive scale-ups. The system runs on attrition the way an evolutionary system does. Understanding this will not make failure pleasant, but it dissolves the moral panic around the numbers: high attrition among venture-backed startups is what the funding model looks like when it is working. The venture economy this filter feeds is not small – Atomico’s State of European Tech 2025 puts the European tech economy at roughly $4 trillion.

The useful question: what has to be true

Aggregate failure rates are almost useless for deciding anything about a specific company. Your startup will not 27% raise a Series A; it will raise one or it will not, based on facts that are largely knowable in advance. So the productive move is to replace “what percentage fail?” with “what has to be true for this company to work?” – a question that produces a checkable list: people must have this problem, they must pay this much, we must acquire them below this cost, churn must stay under this line, this technology must work at this scale.

Lists like that do two things averages cannot. They tell you where the risk actually is – usually concentrated in one or two assumptions, not spread evenly. And they tell you what evidence would change the picture, which is what each funding stage is for: pre-seed tests the problem, seed tests willingness to pay, Series A tests the machine. Companies fail when a load-bearing assumption turns out false and the team learns it too late. The failure-rate question, asked properly, is a question about how quickly you find out.

Frequently asked questions

Is it true that 90% of startups fail?

The figure is untraceable to any primary source and blends different company types, timeframes and definitions of failure. The strongest broad data – US BLS figures covering all new businesses, not just startups – shows about 65% failing within ten years, and no rigorous dataset produces a clean 90%.

What percentage of new businesses fail in the first years?

Per US BLS data – US, all business types, not only startups – roughly 20% of new businesses fail within two years and about 45% within five. Comparable pan-European cohort statistics of that rigour do not really exist, which is itself worth knowing.

How many startups make it to Series A?

In Dealroom’s cohort of 3,075 European companies that raised seed funding in 2016-18, 27% raised a Series A within 36 months and 31% within 48. The cohort is dated but remains the best published European benchmark – and not raising an A is not the same as shutting down.

Why do startups actually fail?

The recurring patterns are unglamorous: building something too few people want badly enough to pay for, running out of money before finding repeatable demand, founder conflict, and being out-executed once a market proves real. Most trace back to a load-bearing assumption that was never tested early enough.

Keep going

For the definitions underneath the data, start with what is a startup. Then see what each stage filter actually tests in what is a seed round and what is Series A funding, and what the survivors become in what is a scaleup. For the European round data we track continuously, see the fundraising data hub.