Start Smaller Than the Market

The Core Insight

On September 18, 1958, Bank of America mailed 60,000 credit cards into Fresno, unsolicited, with no application to fill in. More than 300 merchants had signed up first. Thirteen months later the network held 2 million cards and 20,000 merchants.

Andrew Chen was inside Uber while it grew past 100 million active riders, 800 markets, and 50 billion dollars of gross revenue. He is now a general partner at Andreessen Horowitz. The book comes out of that seat and more than a hundred interviews.

A networked product is worth nothing to its first user, so the effect everyone wants runs backwards at the start. Chen calls that anti-network effects. Users arrive, find nobody, leave, and each departure makes the next arrival worse.

Most operators treat a network effect as one thing that shows up with scale. Chen argues it is three separate forces with three sets of metrics, and all three run in reverse below a threshold. The unit of planning is the atomic network, the smallest group that holds together without help.

The Framework

The book is one S-curve with a droop at the end, cut into five stages.

• The Cold Start Problem. A sub-scale network destroys itself, so the first job is one atomic network.
• The Tipping Point. Launching becomes repeatable, and each new network tips faster than the one before it.
• Escape Velocity. Acquisition, engagement, and economics become three separate programs.
• Hitting the Ceiling. Saturation, channel decay, hard-side revolt, spam, and overcrowding arrive one after another.
• The Moat. Network fights network, and the winner takes one network at a time.

Chen retires Metcalfe's Law for this job. Metcalfe says value grows as the square of the number of nodes, which says nothing about the first hundred users. Warder Clyde Allee found in the 1930s that an animal colony below a threshold spirals to zero. Above the threshold it compounds to a carrying capacity. Allee's threshold becomes the tipping point, and carrying capacity becomes saturation.

Every network splits into a hard side and an easy side, and the hard side does the work. Below the threshold the product is a ghost town, and above it the product delivers a magic moment.

Key Ideas

The Atomic Network Has a Number

An atomic network is the smallest network that stands on its own. Chen finds the number by plotting network size against an engagement metric and looking for the kink in the curve.

The thresholds span two orders of magnitude. Zoom works at two people. Slack takes three, and 93 percent of customers who send 2,000 messages stay. Airbnb needs 300 listings in a city, 100 of them reviewed. Uber needs 15 to 20 cars online and an average pickup time under three minutes. Tinder counted 20,000 users in one market as escape velocity, and a subreddit needs about a thousand subscribers.

A higher threshold makes the launch harder and the finished network harder to attack. Ten people on one team beat ten people scattered across a large company. Fresno worked because 45 percent of its families already banked with Bank of America.

The measurable inverse of a magic moment is a zero. At Uber a zero was a rider opening the app with intent and finding no cars. One more driver does not clear zeroes, because the network has to be built out and active. Chen tracks the share of users who see a zero, network by network.

The Hard Side Holds the Power

Every network has a minority that does the work and holds the power. The list runs from creators and sellers to developers and the manager who opens a workspace and invites a team.

Wikipedia draws 500 million unique visitors a month and runs on about 100,000 active contributors, which is 0.02 percent of the viewer pool. About 4,000 people make more than 100 edits a month. Steven Pruitt made nearly 3 million edits and wrote 35,000 articles, unpaid, at more than three hours a day.

Drivers are about 5 percent of Uber's users, and power drivers are 20 percent of supply producing 60 percent of trips. A new active rider cost 20 to 50 dollars. An active driver cost more than ten times that, and in supply-constrained San Francisco it reached 1,000 or 2,000 dollars.

Concentration shows up in software too. Slack's S-1 reported under 1 percent of customers producing 40 percent of revenue. Zoom took 30 percent of revenue from 344 accounts. Vine's top creators, all eighteen of them, asked for 1.2 million dollars each plus product changes in return for twelve posts a month. Vine refused and shut down a few years later.

Tipping Points Copy and Paste

Once one atomic network works, the job is to run it again faster. Chen catalogs four moves that hold a network up while density builds.

Invites work as a copy-and-paste feature. LinkedIn spent its first week with no sign-up form on the site, and every invited user arrived connected to someone. Gmail went invite-only in 2004 because it ran on three hundred old Pentium III machines nobody else at Google wanted.

The second move ships a single-player tool and hangs a network off it. Six months after Instagram launched, 65 percent of users followed nobody, while 2.2 million users uploaded 3.6 million photos a week. Facebook bought it eighteen months after launch for a billion dollars.

The third move is money, and his maxim for a chicken and egg market is to buy the chicken. Uber posted Craigslist listings guaranteeing 30 dollars an hour whether trips came or not. Referrals of give 200 dollars, get 200 dollars, plus word of mouth, brought almost two-thirds of its drivers. Money accelerates a tipping point and buys nothing during a cold start.

The fourth move puts manual labor where the product is missing, and Reddit's founders posted the front page themselves under dozens of dummy accounts. Turn that off once real users arrive, or the fakes soak up the status that motivates the real ones.

Escape Velocity Runs on Three Separate Effects

The most reusable idea splits the network effect into three forces with separate metrics.

Acquisition is product-driven viral growth. PayPal paid 10 dollars to any user who invited a friend and 10 dollars into the new account. Growth ran from fewer than 10,000 users to 100,000 within months, then a million, then 5 million within a year. A viral factor of 0.5 doubles the base, 0.7 multiplies it by 3.3, and 0.95 multiplies it by 20.

Engagement is retention, measured by cohort and segment. Of the people who install an app, 70 percent are gone the next day, and 96 percent by three months. The a16z bar is 60 percent at day one, 30 percent at day seven, and 15 percent at day 30. Churned users inside a live network are dark nodes that a colleague's shared folder wakes up.

Economics is the business model improving with size, and the Uber math is the cleanest demonstration. Take a guarantee of 25 dollars an hour against an average fare of 10 dollars. A driver doing one trip an hour earns 10 dollars, so the burn is 15 dollars per trip. At two trips an hour the burn falls to 2.50 dollars per trip. The larger network then pays bigger incentives, takes the drivers, and cuts rider prices.

Every Ceiling Arrives With a Named Cause

Growth stops for five reasons, and the book names each one. Market saturation caps the number of people, and network saturation caps the value of new connections. A Snapchat memo put a user's top friend in a week at 25 percent of send volume. By eighteen friends, each additional friend adds under 1 percent.

The answer to saturation is the adjacent user, the person who tried the product and failed to become engaged. Bangaly Kaba joined Instagram in 2016 at over 400 million users growing in a straight line. His team worked through about eight groups over three years. The first was women aged 35 to 45 in the United States, and the last was women in Jakarta on older 3G Android phones. Instagram passed a billion users.

Channels decay on their own. The first banner ads on Hotwired in 1994 drew 78 percent clickthrough, and banner clickthrough now sits between 0.3 and 1 percent. A 50 percent drop in invite conversion produces an 80 percent drop in total new users.

The remaining causes are the hard side revolting as it professionalizes, context collapse, and overcrowding. Usenet was created in 1980, flooded in September 1993 when AOL mailed out millions of CD-ROMs, and gone by 2000.

The Moat Is the Shape of the Network

A moat around a networked product is the difficulty of cloning the network. Wimdu launched against Airbnb with 90 million dollars and hired over 400 people inside a hundred days. It reached 50,000 listings. Airbnb was then two and a half years old, with 40 employees and 7 million dollars raised. Wimdu went to zero, because its best 10 percent of inventory sat at the bottom 10 percent of Airbnb's.

The shape of the moat follows the shape of the network. Uber's effects were local, so winning New York did nothing in San Diego. Airbnb's travel network is global, so an attacker has to replace travelers from everywhere to take one city. Its tipping point in a new city is over 300 listings with 100 reviews.

The upstart move is cherry picking. Craigslist is a network of networks, so an attacker needs one entry point while Craigslist defends all of them. Airbnb took room rentals and pushed its listings back onto Craigslist with a link home. Share also moves double, because two players at 50/50 who trade 20 points end up at 60/40.

The incumbent move is the wide launch, and Google+ is the case against it. It launched in June 2011, pushed from Google.com, YouTube and Photos, and announced more than 90 million users within months. Comscore measured desktop visitors at about three minutes a month, against six to seven hours a month on Facebook. Google+ claimed 300 million active users at its peak and shut down in 2019.

Practical Applications

Write your atomic network as a place and a time before writing a plan. Uber's real unit was five in the evening at the Caltrain station at Fifth and King.

Find the kink in your own curve. Plot network size against your engagement metric and find the size where retention stops sagging. That number becomes the launch target for every network after the first.

Count zeroes weekly, split by segment. Users who hit a zero churn and conclude the service is unreliable, and one more supplier does not clear them.

Subsidize the hard side, and only after the product works without subsidy. Chen's order for a marketplace is supply, demand, supply, supply, supply. When you attack a larger network, take the one sub-network where their density is thinnest.

Split growth reporting into the three effects before the next planning cycle. Acquisition owns viral factor and cost per user. Engagement owns retention curves by segment. Economics owns conversion, pricing, and burn per transaction.

Who This Is For

Founders building anything two-sided get the most from it: marketplaces, social products, developer platforms, workplace tools that spread by invitation. Operators inside a large company get the ceiling chapters, which name five reasons growth stalls.

Skip it if you sell a single-player product to one buyer at a time. The machinery assumes the value comes from other users.

The vantage point sits inside the case selection. Chen writes from a general partner's chair at Andreessen Horowitz. He put more than 400 million dollars into over two dozen startups there in three years. He led the firm's Series A in Clubhouse, the book's magic moment case, when it had two employees. Survivors crowd the case list: Slack, Zoom, Airbnb, Tinder, Instagram, PayPal, Uber. The thresholds come from operator interviews rather than audited data, so treat each number as a hypothesis for your own funnel.

The Decision

Two answers settle whether this book applies to your week. Name the atomic network you are building, as a group of people in one place at one time. Then name the smallest size at which it holds together without you pushing.

If you cannot state both, you are planning against a market, and a market retains nobody.

Plot your engagement metric against network size, find the kink, and write that number down. Launch the next network to that number and nothing wider.