In 2019 MakerStock's first customers were the prototyping shops of university architecture, design and engineering programs. The business was simple: we bought plywood and acrylic by the pallet, cut it into sizes that fit the laser cutters those shops had installed, and shipped it. One nice thing about the segment was that we could assemble a list of every potential customer by reading university websites. Unfortunately, knowing who your customers are does not make them your customers. We acquired shops one or two a week, by email, by phone and sometimes by visiting. It was slow. Only about half of those shops became repeat customers. But two years in, we had a couple of hundred shops, enough for a steady weekly flow of orders, enough to buy material in economical quantities, and enough to justify better cutting machines. By then the customer base improved our results every week. A rival would have needed two years and a great deal of patience to match it. Performance enhancing and hard to acquire: by the two tests, we had an alpha asset that had not existed when we started.
My experience with MakerStock illustrates the basic physics of spinning up a flywheel from a standstill. The engine is a value proposition offered to a customer with a job to be done. Because a flywheel accumulates energy incrementally, a small push every day for years produces a store of advantage that no single effort could. So the most fundamental variable is time. Firms build advantage by being there every day, acquiring and keeping customers. That chugging engine in turn spins the other wheels: volume, learning, quality and brand.
The hard part is the first push. When the wheel is stone cold, what value proposition can move it? In most cases, only a solution that is dramatically better than the alternatives for some customers, on some dimension they care about. Peter Thiel, in his book Zero to One, puts a number on it: a startup's technology should be ten times better than the nearest substitute. I take the ten as rhetoric rather than a measurement, but the underlying claim is right. When some disequilibrium in technology or in taste opens a gap, an organization that moves fast can offer a product far better than what existed. For a while that product is its only alpha asset. At the start of the pandemic, Zoom offered video calls that required no registration, no download and no special equipment; against a field of a dozen established products it took the market, almost entirely because its product worked when the others did not. Within two years Microsoft had Teams at parity on most features and ahead on some, and Google had rebuilt Meet. Zoom remains a large company, though its alpha assets now are its brand and its customer base; the product is table stakes.
A few firms skip the cold start. An established company can hitch a new business to an existing flywheel, as Microsoft hitched Teams to Office. A new firm with a famous founder starts with a brand, as the wine label LVE started with the singer John Legend. Chapter 2 used both cases as examples of endowments. For the rest of us the sequence is fixed. For a while the better product is the alpha asset; I call that period the window. The job during the window is to spin up the flywheels that will be the assets when it closes.
The Window
The window is short. I think it is the most important period in the life of a business, since what a firm builds during it determines whether the firm has anything left when the product advantage is gone. Founders naturally want to keep improving the product; that is usually the least useful thing they can do with the window. The firms that use it well do something that looks, from the inside, like premature optimization: they invest in brand, in customer relationships, in cost and in the organization's own capabilities at a time when the product is selling itself and those investments seem unnecessary.
Consider Keith Belling, who founded the snack brand Popchips and sold it. He then founded Right Rice, a rice-shaped pasta with a nutritional profile that many health-conscious eaters prefer to white rice. From Popchips he knew the playbook. Before the product was in a single store, he hired the design firm IDEO to create the packaging and the identity, because brand would be the asset when the product was copied. He built distribution in the proven sequence (direct to consumer, then specialty retailers, then the mainstream chains), since each channel is a customer network that the next one does not replace. He could move many times faster than General Mills, but every move was toward a machine with flywheels that would matter once the business reached its sustaining phase.
Chapter 4 made the same point about a person: an early success is a window. Like a better product, a young professional's early success (a project that worked, a paper that was noticed, a client who was pleased) will be imitated or forgotten. What they have when the attention moves on is what they built while they had it: the relationships, a reputation for a specific kind of problem, a body of work that becomes a method.
The Advantage Thesis
The founders who use the window well can say, before they have anything to show, what their advantage is now, what it will be in five years, and what they will do in the next two to get from one to the other. I ask every student in my entrepreneurship course to write this down as what I call an advantage thesis. It has three parts.
The first is the advantage now: why anyone would pay today for what the venture offers, which for a new firm is nearly always a solution that is better than the alternatives for some customers, plus the speed to reach them before anyone else does. The second is the advantage at year five, if the firm survives. By then the solution will have been copied, so the question is which of the five flywheels will have spun up far enough to be hard for a rival to acquire: a customer base, a brand, scale, a cost position, a capability. The third is the work of the next twelve to twenty-four months that converts the first into the second. The thesis is the logic that connects the three. I predict that a venture with a good product and no thesis will be copied at the moment it becomes worth copying.
The thesis is a prediction rather than a pitch. It is written before the result, the first of chapter 3's three habits. Camuffo's trials (later in this chapter) found that the habit improves decisions. Revenue in the first two years is noise, but the thesis can be checked on the assets rather than on the outcomes. Its measurements are the number of repeat customers, the cost per unit at the current volume, the share of new customers who arrive by referral, and whether the capability the firm claims to be building shows up in any observable practice. All of these move months before the financial results do. A founder who tracks the assets learns whether the thesis is working while there is still time to change it; a founder who tracks revenue learns at the end.
Here is MakerStock's thesis as I would have written it in 2019. Advantage now: nobody else sells plywood and acrylic cut to the sizes that fit a university shop's laser cutters, and we can ship this week. Advantage at year five: a customer base of a few hundred shops, acquired one at a time, that a rival would need years to match; purchasing scale that lowers our cost per sheet; and a name that makers recognize. Work for the next two years: acquire shops one at a time, build the catalog around what the shops actually cut, and keep every repeat customer. The thesis was mostly right about the customer base and the scale. It was silent about the capability, the part that turned out to be hardest to build and most worth having.
The fifth flywheel belongs in the thesis too. The version of this exercise I assign asks founders to apply two tests to it. First, the capability must be credibly rooted in the founders; a culture of customer obsession declared by founders who dislike customers will not survive the first bad quarter. Second, the capability must reinforce one of the mechanistic flywheels, since, as chapter 5 said, a culture that does not feed cost, quality, brand or the customer base is a club. A thesis that passes both tests names one capability, says which flywheel it feeds, and says what the founders will do in the first year, visibly and at some cost to themselves, to build it.
The Levers
Four of the five flywheels respond to ordinary management: sell more, learn faster, fix the product, feed the brand. The fifth, organizational capability, needs particular levers. The firms that have built a durable capability have pulled the same ones.
The example I use is DuPont and safety. My father, who worked as a consultant to DuPont, came home one day with a story: someone at the laboratory had cut a finger on a box cutter, and the unit had stopped to analyze the root cause and plan against a recurrence. For more than a century DuPont has listed safety first among its values. It has even turned the capability into a product, a consulting business that sells safety practice to other companies. The man who ran that business put the connection to the firm's performance in one sentence: good safety means control of operations; control of operations is what drives improvement in everything else.
The first lever is visible, costly action aligned with the capability. Employees are not stupid. Because it costs something, stopping a laboratory to investigate a cut finger persuades everybody; a slogan at the plant entrance persuades nobody. When I became Wharton's vice dean for innovation, the dean set aside about half a percent of the school's revenue for exploration with no specified return, appointed a senior faculty member to lead it, and announced the budget. The money was the signal; without it the title would have been decoration.
The second lever is an origin story. DuPont employees will tell you that the company began in 1802 with one product, gunpowder; that explosions were frequent; and that the founder built his own house within the blast radius of the mill. Amazon employees will tell you that the first desks were doors on sawhorses. Whether every detail is true is less important than whether the stories are told and retold, each one reinforcing a single capability: safety at DuPont, frugality at Amazon.
The third lever is codification: DuPont writes safety into performance reviews and has structured processes for risk assessment that it has refined for decades and now sells. Processes and systems aligned with a capability reduce the friction on the wheel, while processes at cross purposes with it stop the wheel entirely, whatever the slogans say.
The fourth lever is a manifesto, such as the slide deck on its culture that Netflix published in 2009 and that has since been viewed tens of millions of times. Two of its lines say exactly what capability the company wanted and what it would pay to keep it: "A great workplace is stunning colleagues," and "Adequate performance gets a generous severance package." (Whoa. Got it.) A company that says something that specific, in public, has committed itself in a way that a mission statement about excellence never does.
The fifth lever is training. DuPont implemented its safety system with multi-day programs for plant managers and supervisors, coaching for every leader, a half-day for every worker on the floor, and recurring workshops after that. The Indian textile plants in the Bloom experiment that chapter 5 described were trained for months and improved, but over nine years without reinforcement they let some of the practices slide. A capability lasts as long as people keep teaching it.
Learning When the Score Is Noisy
The levers above build capabilities whose effect can be seen. The harder problem is building an advantage when the outcomes that would tell you whether you are improving are too noisy to read, which chapter 3 argued is the normal condition in most consequential domains. Three findings from research on expertise and on young firms suggest what to do.
The first finding is about the conditions under which expertise develops. The psychologists Daniel Kahneman and Gary Klein, who disagreed about most things, agreed on this: intuition becomes reliable only in an environment regular enough to have patterns, and only when the learner gets prompt, unambiguous feedback. Chess, firefighting and anesthesiology qualify; picking stocks, forecasting politics and hiring executives do not. In those domains confidence grows with experience while accuracy stays put. So build the feedback where the environment does not supply it: write down the prediction, with its reasoning, before the result, and keep a record of decisions and their expected outcomes, and read it. A decision journal turns an environment with no feedback into one with some.
The second finding is about the kind of practice that builds expertise. The psychologist Anders Ericsson's research on expert performance showed that the kind of hours, not the number, is what produces expertise: deliberate practice, at the edge of current ability, with immediate feedback and repeated correction. The popular version, ten thousand hours of anything, is not what he found. Later work found that the share of performance that practice explains varies a great deal by domain and is far smaller in the noisy ones. Still, the structure of deliberate practice transfers. A salesperson who records every call and reviews it with a colleague is practicing deliberately; one who makes a thousand calls and counts the orders is accumulating experience.
The third finding comes from experiments with young firms. Arnaldo Camuffo and his colleagues ran four randomized trials in which several hundred young firms were trained in what they called a scientific approach: state the theory behind the business, derive what must be true for it to work, test it, and revise. The treated firms were not more successful by every measure. They were more likely to shut down projects that were not working and more focused in how they changed direction, and on average they brought in more revenue over the period. So better process led to more abandonment, not more persistence. In a noisy environment, the deliberate builder of an advantage specifies in advance what would make them stop, and then they stop.
An AI model can supply feedback where none existed, in domains where its output can be checked: code that runs or does not, a draft against a standard, an analysis against a known answer. In those domains it is a practice partner of a kind no previous generation had. Where the model produces the work instead of checking the person's, it builds no capability in the person. I suspect a career spent asking the model for answers is a career spent practicing nothing. Ericsson drew the same line between deliberate practice and experience.
One more distinction heads off a common mistake. The population findings that chapter 3 recommended for evaluation (a process feature that predicts outcomes across hundreds of firms or managers) are evidence for prediction but weaker evidence for prescription. For instance, active share, the fraction of a fund's holdings that differs from its benchmark index, predicts fund returns across the population of funds. It does not follow that a closet indexer (a manager who charges for active management while holding something close to the index) will earn more by raising their active share, because the funds that chose high active share may differ in other ways. The textile experiment is the exception: because the practices were assigned at random, we know that adopting them helps. For most process features no such experiment exists. A manager who adopts a population-validated practice is making a reasonable bet. They should call it that and watch for the moment the practice becomes a target and stops being a measure.
Bets That Did Not Pay
An account of deliberate building that reports only the wins is the halo that chapter 3 described, the result written back into its explanation. I have more failures than I have room for.
At Xootr, our scooter company, we spent two years and a meaningful fraction of our capital developing a product line adjacent to the scooter, on the theory that our brand and our manufacturing capability would carry it. The brand did not reach that far. Once the volumes were different, the manufacturing capability was not the asset we thought it was. We shut the line down. At MakerStock we assumed for a year that the shops we had won in universities would lead us to the professional fabrication shops that use the same machines, and that the customer network would extend itself. It did not; the professional shops bought on price and lead time from suppliers we could not match, and the university relationships gave us no entry. In both cases we treated an asset that was performance enhancing and hard to acquire in one setting as if it would be the same in the next. And in both cases we had not written down, before we started, what would make us stop, so we stopped late.
Notes
a startup's technology should be ten times better Peter Thiel with Blake Masters, Zero to One: Notes on Startups, or How to Build the Future (New York: Crown Business, 2014). The rule of thumb is stated in the chapter on building a monopoly.
Daniel Kahneman and Gary Klein, who disagreed about most things, agreed on this Daniel Kahneman and Gary Klein, "Conditions for Intuitive Expertise: A Failure to Disagree," American Psychologist 64, no. 6 (2009): 515–526.
Anders Ericsson's research on expert performance K. Anders Ericsson, Ralf Th. Krampe, and Clemens Tesch-Römer, "The Role of Deliberate Practice in the Acquisition of Expert Performance," Psychological Review 100, no. 3 (1993): 363–406; and Anders Ericsson and Robert Pool, Peak: Secrets from the New Science of Expertise (Boston: Houghton Mifflin Harcourt, 2016). The finding that the variance explained by deliberate practice differs widely by domain is Brooke N. Macnamara, David Z. Hambrick, and Frederick L. Oswald, "Deliberate Practice and Performance in Music, Games, Sports, Education, and Professions: A Meta-Analysis," Psychological Science 25, no. 8 (2014): 1608–1618.
Arnaldo Camuffo and his colleagues ran four randomized trials Arnaldo Camuffo, Alfonso Gambardella, Danilo Messinese, Elena Novelli, Emilio Paolucci, and Chiara Spina, "A Scientific Approach to Entrepreneurial Decision-Making: Large-Scale Replication and Extension," Strategic Management Journal 45, no. 6 (2024): 1209–1237. Treated firms terminated more projects and showed higher average cumulative revenue over the observation period; the simple claim that the treatment increased pivoting did not replicate consistently.