The cold start problem

The term came out of recommender systems, where it described the problem of having no data about a new user or a new item. Andrew Chen took it somewhere larger. He led rider growth at Uber through the years the platform went from fifteen million active users to a hundred million, then became a general partner at Andreessen Horowitz, and in 2021 published The Cold Start Problem, built on that experience and on interviews with the founding teams of LinkedIn, Twitch, Zoom, Dropbox, Tinder, Uber, Airbnb and Pinterest.

The problem he describes is structural rather than a marketing failure. A networked product is worth nothing to its first user, because the value is other users. A marketplace with no sellers is useless to buyers, and with no buyers it is useless to sellers. A messaging tool with one person on it is a notepad. Each side is waiting for the other, and neither has a reason to arrive first.

Two ideas do most of the work in his account. The first is the hard side. The two sides of a network are almost never equal, and one of them is consistently scarcer, does more work and quits faster. On a marketplace it is the sellers. On a content platform it is the creators. On a ride-hailing service it is the drivers. Chen was on the driver growth team at Uber for exactly this reason, and his argument is that the product should be designed around whichever side is hard, because that side sets the ceiling for everyone else.

The second is the atomic network, and it is the part that changes what you do. It is the smallest group for which the product is genuinely useful with nobody else on it at all. The test is not size, it is whether the group survives on its own: if everyone outside it disappeared, would the people inside keep using the thing. For a workplace chat tool that can be three colleagues on one team. For a dating app it was one campus, or in Tinder's case a single party. For Uber it was enough drivers in one city that a rider opening the app finds a car nearby. Facebook launched into one university.

What makes the threshold matter is the force running underneath it, which Chen calls anti-network effects. Below the atomic size a network does not simply grow slowly. It destroys itself. Empty rooms make people leave, which empties the room further, which drives out the people who were still checking. So a thousand users spread across a hundred cities is not one percent of the way to a hundred thousand. It is a hundred separate networks, each of them dying.

The honest limit on all of this is the evidence base. The underlying economics of network effects has a long research literature behind it, going back to work on standards and telecommunications in the 1980s. Chen's framing is not that literature. It is a synthesis assembled from the companies that made it, told largely by the people who ran them, which means the failures that used the same playbook are not in the sample. The mechanism is sound and the specific prescriptions are argued from survivors, and both things are true at once. See survivorship bias.

This applies to more than apps. Any internal system whose value depends on other people using it has the same first day: the shared documentation nobody writes because nobody reads it, the referral scheme with no referrals in it, the internal marketplace for skills or equipment, the community a company builds around its product, the partner programme with two partners in it.

Which reframes the usual launch. The instinct with an internal tool is to roll it out to everyone at once, on the reasoning that adoption is a numbers problem and a company-wide launch gives it the best chance. That produces the hundred empty rooms. Nobody in any single team has enough colleagues using it for it to be worth their while, everyone concludes it does not work, and the failure gets attributed to the tool.

The alternative is to pick one team and make it work completely there. Not a pilot in the sense of a light trial across several departments, but total saturation of one group small enough to reach and large enough to be self-sufficient. When that group would keep using it even if the rest of the company stopped, it is real, and the next group has something to join rather than something to start.

Two questions do the work. Which side is hard here, meaning which participants are scarcest and quit most easily, and does the design serve them first. And is there any single group, however small, where this already works without help from anyone outside it. If the answer to the second is no, the number of total users is not information about progress. It is a description of how many separate rooms are currently emptying.

The related mechanism worth holding beside this one is the Matthew effect, since a network that clears the threshold accumulates advantage from that point onward, and one that does not clear it loses ground for the same reason.

Source: Chen, The Cold Start Problem: How to Start and Scale Network Effects, Harper Business, 2021. Chen's account of his own route to the argument, including the years running rider growth at Uber, is on his site at andrewchen.com. The term originates in recommender systems research, where it describes having no prior data on a new user or item.

The Cold Start Problem

Andrew Chen, 2021

The case laid out in full, with the five stages a network passes through and interviews with the teams who took products through them. Read it for the atomic network idea, which is the part that changes what you actually do on Monday.