The previous article ended by noting that the product which earns a retention curve is built under deep uncertainty. This article is about that building, and about the risk that a venture fails not for want of a market but for want of reaching it before the money runs out. The central reframing is that a startup does not execute a known plan. It searches for one. The binding risk is therefore not whether the team can build the thing, which is the engineer’s question, but whether the team can discover what to build in the finite number of attempts that its cash allows. That number, and the odds it buys, can be written down, which is the work of this article. The treatment is general, and it is information rather than business advice.

A Brief History

The modern view of execution begins with a distinction drawn by Steve Blank, that a startup is not a smaller version of a large company. A large company executes a business model it already understands, while a startup searches for one it does not. The customer development method made the search systematic, and the lean-startup movement gave it a loop, build, measure, learn, turned as quickly and cheaply as possible. The minimum viable product was named as the smallest build that tests a hypothesis, the device that keeps each turn of the loop short. Paul Graham later compressed the survival question into a single pair of words, default alive or default dead, asking whether a company’s present trajectory reaches profitability or fit before its bank balance reaches zero. The empirical literature added a warning, that premature scaling, building and spending as though the search were already won, is among the most common ways that funded startups destroy themselves.

Two Senses of Execution

The word execution hides two different risks. The first is engineering execution, whether a team can build what it set out to build, on time and to a standard that works. This risk is real, and in deep-technology ventures, where the science itself may not cooperate, it can be the risk that matters most. For the typical software startup, however, it is rarely the risk that kills, because building a given specification is the part the founders usually know how to do. The second risk is search execution, whether a team can discover what is worth building at all, which specification will earn the product-market fit that the previous article measured as retention. This is the binding constraint for most ventures, and it is harder, because the target is unknown at the outset and is found only by trying. The rest of this article concerns the second sense, and treats the search as a sequence of attempts made against a finite budget.

The Build-Measure-Learn Loop

A single turn of the loop takes a hypothesis about what to build, builds the smallest version that tests it, puts it in front of real users, and reads the result as evidence for or against the hypothesis. Two quantities describe the turn. Its duration is the cycle time, written $\tau$, the calendar cost of one attempt. Its money cost is the burn incurred over that time. The whole purpose of the minimum viable product, and of the discipline around it, is to make $\tau$ small, so that more attempts fit into the same fixed runway. A team that ships a careful product once a quarter and a team that ships a rough test once every two weeks are not working at different speeds alone. They are buying different numbers of attempts at the same unknown.

Runway and the Number of Attempts

Let a venture hold cash $C$ and spend it at a burn rate $b$ per unit time. Its runway, the time before the cash is gone, is

\[T = \frac{C}{b}.\]

If one turn of the loop takes time $\tau$, the venture can complete

\[N = \frac{T}{\tau} = \frac{C}{b\,\tau} = \frac{C}{e}, \qquad e = b\,\tau,\]

attempts before the runway ends, where $e$ is the cost of a single attempt, the burn multiplied by the cycle time. The number of attempts is the cash divided by the cost of an attempt, and everything execution does well or badly shows up in that fraction. Cheaper experiments, shorter cycles, and a lower burn all raise $N$, and a venture that spends heavily on long cycles buys very few attempts at the only question that matters.

The Odds of Finding Fit

Treat each attempt as a trial that discovers a fitting product with probability $p$ and otherwise teaches something and fails. The chance of at least one success across $N$ independent attempts is the complement of failing every time,

\[P_{\text{fit}} = 1 - (1 - p)^{N},\]

which is the way a geometric process accumulates its chance of a first success. Put numbers to it. A venture with six hundred thousand in the bank, burning fifty thousand a month, has a runway of twelve months. At a cycle time of six weeks, it gets $N = 8$ attempts, and at a per-attempt success rate of $p = 0.15$, its odds of finding fit are

\[P_{\text{fit}} = 1 - (0.85)^{8} \approx 0.73.\]

Now move only the execution levers. Halve the cycle time to three weeks, and the same runway yields $N = 16$ attempts, lifting the odds to about $0.93$. Halve the burn instead, stretching the runway to twenty-four months, and the result is the same. But let the cycles drag to three months each, so that $N = 4$, and the odds fall to about $0.48$, a coin flip, from nothing but slow building. The market never changed. Only the number of attempts did.

Default Alive or Default Dead

The number of attempts a venture needs follows from the same trial. On average, fit first arrives after $1 / p$ attempts, the mean of the geometric process,

\[\frac{1}{p} = \frac{1}{0.15} \approx 6.7 \text{ attempts}.\]

Paul Graham’s question, whether a venture is default alive or default dead, becomes a comparison of two numbers. A venture is default alive when the attempts it can afford comfortably exceed the attempts it is likely to need, $N \gtrsim 1 / p$. The example venture, with $N = 8$ against a need near seven, is barely on the living side of that line, and the same venture building slowly, at $N = 4$, is on the dead side no matter how real its market.

The comparison also sizes a raise. To reach a chosen confidence $P^\ast$ of finding fit, a venture needs

\[N^\ast = \frac{\ln(1 - P^\ast)}{\ln(1 - p)}, \qquad C^\ast = e\,N^\ast,\]

attempts and the cash to fund them. For ninety percent confidence at $p = 0.15$, that is about fifteen attempts, and at seventy-five thousand dollars each, roughly $1.1$ million in the bank. The figure is only as trustworthy as the estimate of $p$, but it converts a vague unease about runway into a number that can be checked.

Two Levers, Not One

The odds depend on two things, the number of attempts $N$ and the quality of each attempt $p$, and they are not interchangeable. The value of one more attempt is the chance that fit first arrives on exactly that attempt,

\[P_{\text{fit}}(N) - P_{\text{fit}}(N - 1) = p\,(1 - p)^{N - 1},\]

which shrinks geometrically as $N$ grows. The eighth attempt in the example above adds about five percentage points, against the fifteen that the first attempt adds. Past some point, buying more attempts returns little, and the way forward is to raise $p$, the quality of each attempt, by talking to users, sharpening the hypothesis, and building the test that actually decides it. Speed multiplies attempts, and judgment improves them, and a venture needs both, since fast attempts that test nothing and sharp hypotheses tested once a year fail in opposite ways.

Failure Modes of the Build

Each common way that building goes wrong is a way of wasting attempts. Over-engineering spends a large $e$ on a single attempt, gilding a product whose premise is not yet confirmed. Premature scaling does the same at greater cost, building the machinery of a winner before the search is won, and is among the most reliable ways to convert a generous runway into none. Perfectionism lengthens $\tau$ in pursuit of a polish that the test did not require, trading attempts for shine. Building without measuring spends an attempt and reads no result, learning nothing for the money. The opposite error, shipping something so rough that it tests the wrong thing, spends an attempt and reads a false result, which is worse than learning nothing. Technical debt is the one item on the list that is often correct to incur, since code written to be discarded should be written quickly, and only the parts that survive contact with a real market deserve to be built well. The discipline is to spend cheaply while searching and to build durably only once there is something proven to keep.

Epistemic State

The model treats attempts as independent trials with a constant success probability $p$, and this is its largest simplification. In reality a team learns, so that $p$ should rise from one attempt to the next, which is the entire point of the loop and a reason the constant-rate odds are pessimistic for a team that learns and optimistic for one that does not. The boundary of an attempt is fuzzy, since real builds overlap and branch rather than proceeding one at a time, and burn is not perfectly constant. The single probability $p$ also compresses a great deal of skill, luck, and market structure into one number. What survives these simplifications is the shape of the result, that the odds of finding fit rise with the number of affordable attempts, that this number is cash divided by the cost of an attempt, and that execution is the work of making attempts many, cheap, and sharp. The arithmetic is a lens for that work, not a forecast of any particular venture. Throughout, this is general information, and it is not business advice.

Out of Scope

The craft of engineering management, the specific methods of agile development, and the mechanics of hiring a team that can build are operational subjects left to their own literatures. The deep-technology case, in which engineering execution is itself the dominant risk, is noted here but not developed. The unit economics that decide whether a found product can be sold at a profit belong to the next article, and the durable advantage that execution speed can become is a thread for the article on moats at the end of the series.

Conclusion

Execution risk, for most startups, is the risk of running out of money before discovering what to build. The discovery is a search, not the running of a known plan, and a search is a sequence of attempts bounded by the cash on hand. The number of attempts is the cash divided by the cost of one, $N = C / (b\tau)$, and the odds of success climb with that number and with the quality of each try. A venture improves its position by iterating quickly, spending little per attempt, and sharpening each hypothesis, which is to say by building to learn rather than building to impress. What it cannot do is change the base rate by wishing, or buy fit with a single expensive bet when it could have afforded many cheap ones. The funnel that began this stretch of the series loses most ventures in exactly this work, and the first article counted the bodies. The next article turns from building the product to reaching and charging the people who will pay for it.

References