The Cold Start Problem

The chicken-and-egg bind facing every network product: it has no value until it has users, and cannot attract users until it has value.

A new messaging app with no one on it is worthless to its first user, who has no one to message. A marketplace with no sellers is worthless to the first buyer, who finds an empty store; with no buyers, it is worthless to the first seller, who makes no sales. The product only becomes valuable once enough people are already using it, but no one wants to be early to something empty. This is the wall that defeats most products that would otherwise have ridden a network effect to a dominant position, and “just launch and see” is the strategy that walks straight into it.

What It Is

The cold start problem is the bind that faces any product whose value to each user depends on other users already being present. Andrew Chen, a general partner at Andreessen Horowitz who spent years on growth at Uber, gave the problem its name and its fullest treatment in The Cold Start Problem (2021). The bind is circular. The product is only valuable once a network exists, and the network forms only once the product is valuable. A launch that simply opens the doors to everyone produces an empty room that the first arrivals immediately leave.

The defining move in Chen’s framework is to abandon the idea of launching to a whole market at once. The goal instead is to assemble a single atomic network: the smallest group of users for which the product is genuinely useful on its own, with no one else present. For a workplace chat tool, an atomic network is one team inside one company, not “the company” and certainly not “the market.” For a ride-hailing service, it is enough drivers and riders in one neighborhood at one time of day that a rider opening the app reliably finds a car. The atomic network is the unit of progress. A product crosses the cold start not by acquiring users in general but by completing one self-sustaining network, then another, then another.

Chen frames the full arc as five stages a network product moves through, each with its own dominant problem:

The first stage is the one that kills companies. It is where the product has nothing to offer and the usual growth tactics have nothing to amplify. The later stages are problems of a working network; the cold start is the problem of having no network at all. Most of the framework’s strategic content lives in how a team manufactures that first atomic network when the product, by itself, gives a lone user no reason to stay.

Why It Matters

The cold start is the practical counterpart to the network effect. The network effect is the prize; the cold start is the gauntlet standing between a founder and the prize. A network effect is the moat investors rank highest, but it does not exist until a network does. The period before the network is large enough to be valuable on its own is where the great majority of would-be network businesses die. Naming the stage precisely turns “we’ll have network effects at scale” from an aspiration into a question with a concrete first milestone: have you completed even one atomic network yet?

The founder building a marketplace, a social product, a communications tool, or a platform reads the cold start as the central sequencing problem of the early lifecycle. The instinct to launch broadly and let the network find itself is exactly wrong here. Broad, thin acquisition spreads the early users so far apart that no atomic network ever reaches the density that makes the product useful. The discipline the framework imposes is to pick a single network narrow enough to actually complete, dominate it, and only then move to the next. That’s the same logic as a beachhead, applied to a product whose value is its users.

The investor evaluating a network business reads the cold start as a diligence question that separates two pitches that look identical on a slide. A founder claiming network effects has to answer one thing: is a single atomic network self-sustaining today, or is the company still propping up its early users with subsidies, hand-seeding, and concierge effort that will not survive contact with scale? A marketplace whose first city works without the company manually filling both sides is across the first stage. One whose every market needs the same expensive priming is still in the cold start, however large the aggregate user count looks.

The talent reader weighing an offer from a network-effect startup reads the cold start as a stage gate on the risk. A company that has not yet completed one self-sustaining network is a far earlier, riskier bet than one already replicating networks reliably. The equity should be read against which it actually is. The headline user number does not answer the question; whether the networks stand up on their own does.

How to Recognize It

The cold start is visible in whether the product is useful to a single user dropped into it today, with no special seeding, and in how the early networks behave when the company stops pushing.

• The empty-room test. Open the product as a brand-new user with no existing connections. If there is nothing to do, no one to reach, and no reason to return until others arrive, the product is squarely in the cold start, and acquisition spend will leak straight back out.
• Early users that only stay while subsidized. When retention holds only as long as the company is paying for supply, manually matching both sides, or running the network by hand, what looks like traction is the company standing in for the network. The cold start is solved when an atomic network sustains itself after the props come down.
• Density that is too thin to be useful. A thousand users spread across a thousand cities is a thousand empty rooms; a thousand users in one neighborhood may be a working network. Read whether users are concentrated enough that a typical user finds the value the product promises, not just whether the total count is rising.
• Both sides waiting on each other. In a two-sided market, sellers cite the lack of buyers and buyers cite the lack of sellers, each refusing to be early. That mutual standoff is the signature of the bind, and the only way through is to over-supply one side deliberately until the other has a reason to show up.

How It Plays Out

Tinder’s launch is the textbook case of solving the cold start through atomic networks rather than broad acquisition. A dating app is the purest form of the bind: it is worthless to a user who finds no one nearby to match with, and no one wants to join an empty one. Rather than launch to the public and hope for liquidity, the team seeded the product one university at a time, throwing parties on college campuses where the price of entry was installing the app. Each campus was an atomic network, a population dense enough and socially connected enough that a student opening Tinder found real matches immediately. Once one campus tipped into self-sustaining use, the same playbook moved to the next. The product did not try to be valuable everywhere at once; it became valuable in one bounded network, then replicated.

Uber faced the same problem in physically local form. A rider values the app only if a car arrives in a few minutes, which requires enough drivers in that area at that time; drivers stay only if there are enough riders to keep them earning. The company solved it city by city, and within a city neighborhood by neighborhood and hour by hour, spending heavily on driver guarantees and rider incentives to manufacture the density that made the core promise hold. The subsidies were not a permanent business model; they were the cost of priming each atomic network until its own liquidity made them unnecessary. A rider in one city benefits only from drivers in that city. That is why this network effect is largely local, and why each new market presented its own fresh cold start rather than inheriting liquidity from the last.

The instructive failures are the products that skipped the atomic-network discipline. Google Plus launched in 2011 to an enormous existing user base, an advantage that seemed to make the cold start irrelevant. Yet it never assembled the dense, self-sustaining social circles that make a social product engaging. Users created accounts and found their networks empty of the people they actually wanted to interact with, so they did not return. Vast top-of-funnel reach did not substitute for atomic-network density. A network product cannot borrow liquidity from an adjacent product; it has to manufacture its own, one self-sustaining cell at a time.

Consequences

Holding the cold start as a named stage with a concrete first milestone changes how a team sequences its early spending, and it carries real costs of its own.

Benefits. A founder who frames the early problem as completing one atomic network stops burning acquisition budget on thin, scattered users. Resources concentrate where density can actually be reached, the move most likely to produce a network that survives. An investor with the atomic-network test can separate a marketplace whose first market is genuinely self-sustaining from one whose every market is propped up by subsidy: the difference between a company across the first stage and one still inside it. And all three readers gain a checkable question, “is even one network self-sustaining without the company holding it up?”, in place of an aggregate user count that hides whether any of those users have a reason to stay.

Liabilities. The framework can be over-applied. Not every product has a network effect, and a team that reaches for atomic networks when its product is valuable to a lone user from day one (most straightforward software-as-a-service) imposes a sequencing constraint it does not need. The concierge, do-things-that-don’t-scale effort that primes an atomic network is genuinely expensive and genuinely unscalable. The hard judgment is when to stop subsidizing a network that should by now sustain itself; props left up too long disguise a network that never actually tipped. And the local-versus-global distinction is decisive and easy to get wrong. A founder who completes one local atomic network and assumes the rest of the market will follow cheaply discovers, as ride-hailing did, that each new geography is its own cold start to be paid for again. The stages are a map of a real terrain, not a guarantee that crossing the first one makes the rest free.

Related Articles

Sources

• Andrew Chen, The Cold Start Problem: How to Start and Scale Network Effects (2021) — the founding treatment that named the problem, the atomic-network concept, and the five-stage arc from Cold Start through the Moat.
• The do-things-that-don’t-scale principle that underlies concierge network-seeding was articulated by Paul Graham in his 2013 essay of that name, written for Y Combinator founders, and is the canonical statement of why manual, unscalable early effort is the right way to prime a network.
• The technology-adoption-lifecycle and network-economics vocabulary the cold start sits inside — atomic networks as bounded beachheads, local versus global effects, liquidity in two-sided markets — emerged from the venture community’s writing on marketplaces and network effects through the 2010s and 2020s, and is treated here as field vocabulary rather than the contribution of any single source.