The Atomic Network: Andrew Chen on the Smallest Network That Works

Chen spent years at Uber and then at Andreessen Horowitz watching network products catch or die, and what bothered him was how random the winners looked. Snapchat starts in high schools. Tinder starts in colleges. Slack starts inside one small team. He argues that is not luck. Those products all had to clear a threshold, and small dense places are where it is cheapest to clear.

Read that definition slowly, because two clauses in it are doing work that the popular version drops.

At the same time. Not five thousand accounts created since January. The people have to overlap. A collaboration tool with ten thousand users spread one-per-company is a product nobody can use, and it will churn while the dashboard says growth.

Doing the same thing. Not merely present, but pointed at the same activity. Chen's test is whether the product is valuable, which only happens when the other people who showed up want the thing you want.

On the Tim Ferriss show he gives the version that names the goal directly.

Smallest, and it has to retain. Those pull against each other, and the tension is the whole design problem. Go smaller than the threshold and the network dissolves on its own. Aim much bigger and you have set yourself a launch you cannot afford.

Generic launch advice converges on a small round figure. Get ten users who love you. Chen's threshold is not a round figure, it is a property of your product's mechanics, and the spread between products is enormous.

Zoom needs two or three, because a call with one other person is already the whole product. Slack sits higher, and Chen gives the number from the person who would know, having interviewed Stewart Butterfield for the book.

Three people in one company, not three people. The unit matters as much as the count.

Then there is Tinder, where the number jumps two orders of magnitude, and Chen is careful to bound it from both directions.

He derives it rather than guessing: you can swipe a couple dozen profiles a day quite easily, so a few dozen people is a network you exhaust by Tuesday. That reasoning is the transferable part. Estimate how much of your network one active user consumes per session, then work backwards to the number that leaves them somewhere to go next week.

Which is why Tinder's launch looks like an event rather than a campaign. The team sponsored a birthday party near the USC campus with a bouncer who made you install the app to get in. A couple hundred people installed it that night, and nobody used it at the party. The next morning they woke up to a set of faces they had actually been in a room with, and from there Tinder took the campus. What they bought was not a couple hundred downloads. It was one functioning atomic network, in a single evening.

The honest limitation is that Chen is reasoning from a set of famous winners, and the thresholds he quotes are estimates rather than measurements. He says as much about Tinder, hedging to probably hundreds. Use the number as a target to design against, then correct it with your own retention data.

It sits directly on top of Paul Graham's do things that don't scale, and sharpens it. Graham tells you to recruit users by hand. Chen tells you when you can stop: when one unit holds together on its own and you have built the second one. Against Hormozi's Core Four, the difference is what a channel is for. Core Four optimizes volume of contacts. Chen would spend the same effort concentrating a smaller number of people into one place, on the grounds that a thin spread across a big market never crosses the threshold anywhere.