📗 On Building a Vocabulary for Discussing Network Effects

A Review of The Cold Start Problem by Andrew Chen

I read The Cold Start Problem for the same reason I read and reviewed Certain to Win — I wanted to understand business better now that I work for a startup instead of a school system. Happily, The Cold Start Problem turned out to be one of the more useful books I read last year, and it comes up in discussion with friends regularly in all sorts of contexts. Sometimes we’re talking about corporate recruiting, sometimes it’s about forming friend groups for kindergarteners, sometimes we’re pondering the nature of our book club... but The Cold Start Problem applies to all of it.

Andrew Chen wrote The Cold Start Problem because he found his own understanding of network dynamics to be, in his words, “unforgivably shallow for something so core to the technology industry.” Chen was trying to build a shared vocabulary for discussing mechanics that most high-level tech folks understand intuitively, but struggle(d) to articulate.

Most people, whether they work in tech or use social media or not, have an intuition about network effects. Products get more valuable as more people use them, that marketplaces need both buyers and sellers, and getting the first core group of people to start showing up to a recurring party is the hardest part. But knowing that and having the words to talk about it after different things. Without a good vocabulary, it’s hard to improve your own mental models, or harder to remember things later. Jargon exists because it offers a useful shorthand.

I have found this vocabulary incredibly useful not only for communicating in professional contexts, but also in personal ones. For example: we like the school our kids attend, but’s the network effects that keep us there more so than the quality of instruction or the educational philosophy.

Let’s get to the vocabulary, then

Chen organizes the book around five stages of network growth: Cold Start, Tipping Point, Escape Velocity, Ceiling, and Moat. I am not going to talk about all of them because they are not all equally interesting to me personally, but this is the lens through which Chen examines how specific companies handled each transition. He added enough concrete examples to ground the abstractions in interesting stories that help contextualize and aid memory.

Along the way, he breaks the classical “network effect” concept down into three distinct forces. The Acquisition Effect lets products tap into the network for low-cost viral growth. The Engagement Effect increases interaction as networks fill in. The Economic Effect improves monetization and conversion as the network grows. These effects require different interventions and fail in different ways, which is why he thinks treating “network effect” as a single phenomenon leads to sloppy thinking.

The “cold start” is a zero state, when a user joins and none of their friends are on the service, or when a search returns no results. Every networked product that survives has found a way to make those first moments tolerable.

The “come for the tool, stay for the network” strategy, where a product offers the utility of the tool on its own first and layers on network effects later, is common precisely because solving the cold start problem head-on is so brutal.

Another useful concept in the book is the “atomic network”: the smallest self-sustaining unit of a network that can function independently.

Atomic networks: the minimum viable unit

The atomic network concept lets you focus on step one on the path to the moonshot goal (millions of customers, big successful parties, a great team to work with... whatever). It starts out as “how do I get this one small cluster working.” The smaller the atomic network, the quicker you can build it and then repeat the process. Messaging apps and video calls need only two people. A marketplace for rare sneakers needs a critical mass of collectors in a specific category. A social network for a particular fandom needs enough active members to generate daily content.

For Uber, it was a single city with enough drivers and riders to ensure reasonable wait times. For Slack, it was a single team that had exchanged 2,000 messages. (Stewart Butterfield found that regardless of any other factor, after 2,000 messages, 93% of teams were still using Slack.) For Facebook, a single college campus was enough. In my experience, a party needs at least four families to feel like a party instead of a dinner.

The thing is, a product (or a party plan…) can be perfectly designed and still fail if it launches in a market where the atomic network can’t form. This is one reason that Chen made such a point of focusing on “zeroes” and empty states, like when a user has no search results returned, no friends available on a given weekend, or no listings showing up in a category.

But the product (or rather: the tool) does matter, and successful networked tools “tend to sound like a meme” when described. Snapchat lets you send photos to friends. Uber lets you hit a button to get a ride. Dropbox is a magical folder that syncs your files… although I am still lookging for the magical app that lets me easily sync two folders, because Claude settings being stored in the users folder is really annoying!

That said, the simplicity limits complexity and cost, and it makes the product easy to describe to the next person you’re trying to recruit into the network. Curating who’s on the network, why they’re there, and how they interact matters as much as the product design (party venue, school) itself.

The hard side does most of the work

Every network has what Chen calls a “hard side”: the participants who do most of the work and are hardest to acquire. On Wikipedia, a tiny percentage of prolific editors wrote most of the content. About 5% of Uber’s users (the drivers) carried the entire marketplace on their backs. On YouTube and Instagram, a power-law distribution means the top 20% of creators generate the vast majority of engagement.

The earliest users of a network tend to have abnormally large contact lists — often in the thousands. Thousands sounds crazy until you realize my newsletter goes out to over 8,000 people. Lots of really famous newsletters (Matt Levine’s amazing Money Stuff for example) go out to orders of magnitude more than that.

Even a normal messaging app like Telegram or Signal might look like a peer-to-peer network where everyone is equal, but in practice there are active, extroverted users (👋) who initiate conversations and organize get-togethers, and there are passive users who respond. Nearly every network has this split, even my personal social groups. It feels weird to say it so bluntly, but I invite people to my house way more often than I am invited to parties and dinners, I am almost always the most active poster of the week in the Discord server where I spend most of my time, and I am very rarely a “lurker” in any social network I actually spend time on.

In retrospect I suppose this is one reason people tend to shovel responsibility my way :P

I’ll admit it’s a little strange to think of myself in such business-oriented terms, but the hard side of a network is also the hard side of any social group. Without the people willing to do the organizing, the group doesn’t cohere. A social network can’t exist without its content creators. A marketplace can’t exist without sellers. “Who is the hard side of your network, and how will they engage with it?” turns out to be one of the most important strategic questions any organizer of a networked {thing} can ask.

Chen suggests a diagnostic: “If a piece of content was created, and no one saw it, would the creator be disappointed?” If yes, the social feedback loop is central to the product’s value proposition. Users publish content, others engage with likes, shares, and comments, and positive feedback drives more creation. Without an audience, creators leave.

“Content creation” usually makes us think of the internet influencers, but you can apply this to beautiful decorations at a school dance, a church bake sale, a neighborhood yard sale… the applications are endless.

One of the reasons I am on Substack instead of Ghost (and Twitter instead of Bluesky, Discord instead of ...I dunno, does anything even compete with Discord?) is because the network effects mean I get answers to my questions. I like to write whether people read it or not (I have tons of words written that no one has ever read and that’s fine), but I do really value feedback and discussion and community and all the things that sharing my writing in public allows for.

Substack understood this dynamic well enough to pay prominent writers like Scott Alexander to bring their existing audiences onto the platform, solving the hard side problem with direct financial investment. “If you have a chicken and egg problem,” as the old wisdom goes, “buy the chicken.” Chen is totally up front that some companies just outright “buy” users.

This doesn’t really mean that a network is “bad” or “doomed” or whatever. Sometimes you can’t grow a network with positive unit economics from day one, and pretending otherwise is a good way to not have a network at all. Some products just spend money.

PayPal embedded itself into eBay’s existing transaction flow, acquiring users who already had the behavior of buying and selling online. The $5 referral fee that lit up the original PayPal network is one of the most cited growth hacks in Silicon Valley history.

Subsidizing driver earnings, offering free first books, fronting huge piles of money to established writers to get them to join your platform...

Sometimes this works, sometimes it doesn’t, because...

Retention matters too! Teamwork and camaraderie create bonds that keep users engaged over months and years, which is harder to manufacture than it sounds. The network has to generate real social value, beyond just the utility of the tool.

Enshittification by another name

One of the ways companies “spend” money to encourage growth is by skipping the “charging users so you can be profitable” stage. Charging users directly adds friction that slows network growth, and every paywall is a barrier to the next person joining.

Mature networks shift from growth to extraction. Chen doesn’t use the word “enshittification,” but this framework describes the same phenomenon. I checked and as far as I can tell, this book came out in January 2022, and Cory Doctorow coined the term in an essay published in January 2023.

Any name you want to give it, tho, the product experience often degrades because the switching costs do the retention work that the product used to do... and companies do eventually want to start making money.

When networks grow, they sort of seed their own destruction. Different forces end up pulling toward expansion or fragmentation. Platform dependence becomes a trap for some parts of the network, for example if you integrate too closely with a preexisting network, allowing them to control your distribution, engagement, and business model, you become just a feature of their network. And you risk getting cut off if you’re costing them money instead of making it for them.

Basically, when large networked products reach immense scale, they become networks of networks with diverse needs, and that diversity creates opportunities for smaller, more focused competitors to peel off subnetworks and serve them better.

This is how Airbnb unbundled vacation rentals from Craigslist. “Unbundled” here means that specialized services broke away from general-purpose platforms by serving a specific subnetwork better than the generalist could. Craigslist could have incorporated Airbnb’s features, but it would have been hard for a small team to respond to one particular subnetwork being carved off when dating, real estate, gig work, and everything else were all being unbundled simultaneously.

The seemingly undesirable niche segments are where the next atomic networks form, and you can kind of see this in a high school context: small groups of nerds are ill-served by mainstream school environments, break off and then end up seeding niche educational entrepreneurship adventures like Alpha School.

The idea is that the incentives stop being aligned and things start to get adversarial (do you really want to pay for a tool you had gotten free?). I spend a lot of time thinking about coordination problems (for example, how cooperation and competition interact at scale) and Chen’s framework explains why quality breaks down. The network itself becomes the “moat” protecting the organization from competition, and the platform’s interests diverge from its users’ interests.

And if network effects describe how value grows as networks expand, Ponzi schemes are network effects in reverse: each new participant adds less value while extracting more, until the whole thing collapses.

An anthology of clever solutions

The scrappy plays are pretty fun, tho. Chen distinguishes between paying with money vs. paying with time and effort — smaller companies that can’t write big checks instead do high-touch work, like building custom functionality for partners.

Paul Graham famously argued that entrepreneurs should do things that don’t scale, and Chen seems to agree that many atomic networks require unsustainable effort to get moving. Brief moments of opportunity like a conference, a campus, or a wave of new technology (hellloooo dumb phones, internet, smart phones, ai), can tip a market if you’re ready with the right idea. Twitter launched during SXSW 2007, when a critical mass of early adopters were attending. Airbnb targeted major local events where hotel inventory was scarce.

The moment matters as much as the product.

For example: it’s really hard to build a new friend group, most of the time. Yet it’s surprisingly easy if you happen to like the cohort your kids start attending school with! About 70% of my real-life social life is hanging out with folks I met via my son’s childcare providers between ages 3 and 5. The remaining 25% is mostly my husband’s college friends. The remainder are neighbors and work friends.

But more than the moment or the product are the high-effort tactics that get the ball rolling.

Many early social networks grew by scraping email contacts from Hotmail, Yahoo Mail, and other clients, using libraries like Octazen (later acquired by Facebook). At the time, these new social networks didn’t look like direct threats to email providers, so the API access was freely available.

Tinder (which is heavily reliant on smart phones) seeded its early networks by throwing college parties. Sean Rad and his team would go to a sorority, get all the women to download the app, then go to the corresponding fraternity and show the men that all these women were already on it (pro tip: on dating apps, women are the hard side). They built one atomic network, then figured out how to replicate it: throw another party at another school. From 4,000 downloads, they hit 15,000 within a month, then 500,000 a month after that.

Uber built an internal tool called “Starcraft,” named after the real-time strategy game, that let general managers click on a group of cars on a map and text them “Go to the train station, lots of riders!” They were literally directing the network in real time¹.

Reddit’s founder scraped news websites and posted them with made-up usernames to make the homepage look like an active community. When he went camping for a month after launch and stopped submitting links, the homepage went blank. I was kinda shocked to discover that the bot problem was baked into Reddit from day one by the developers themselves, although I suppose in retrospect I shouldn’t have been. Lots of modern-day tech has shady backstories like this.

Airbnb reverse-engineered Craigslist, not using Craigslist’s (either irrelevant or nonexistent, I can’t remember) APIs, but by building bots to automatically cross-post listings. When a host finished setting up their listing on Airbnb, they could publish it to Craigslist with photos, details, and a link that drove Craigslist users back to Airbnb.

Networks are everywhere

The book is ostensibly about tech companies, but the conceptual framework extends far beyond software. Chen notes, almost in passing, that “money is a network. Religion is a network. A corporation is a network. Roads are a network. Electricity is a network.” And then: “Networks must be organized according to rules. They require Rulers to enforce these rules. Against cheaters. And the Rulers of these networks become the most powerful people in society.”

Religion-wise, the spread of Christianity through the Roman Empire looks different through a network-effects lens. If one wants to be academically secular about it, the early church faced a classic cold start problem: how do you build a community of believers when much of the value of belonging to the community depends on other people already being in it? Jesus and the apostles started with people on the margins: criminals, low-class women. The solution involved atomic networks (individual congregations), a hard side (clergy and missionaries who carried the whole operation), growth hacks (miracles, martyrdom as social proof), and a tipping point (Constantine’s conversion making it the state religion). One can model the rise of evangelicalism this way, with megachurches as atomic networks and charismatic pastors as the hard side.

We can go even further back. The kula ring of Melanesia — a ceremonial exchange network where shell ornaments circulated between island communities — is basically a prehistoric atomic network. The elites who participated in kula exchange built the prestige relationships that held the network together, making them the “hard side.” But the kula ring also enabled a parallel layer of ordinary trade: while the high-status partners exchanged ceremonial valuables, everyone else traded food, tools, and raw materials alongside them. That two-layer structure — prestige relationships enabling utilitarian exchange — maps surprisingly well onto how modern platforms work, where a small number of highly engaged users create the conditions for everyone to get value from the network.

The merchant class emerged because the structural conditions (trade routes, legal frameworks, surplus production) created an environment where merchant networks could sustain themselves. The atomic network for a Phoenician trading colony was a single port with enough resident merchants and enough regular cargo, and also trading partners in the surrounding region willing to exchange goods.

Looking back at the “dense networks are exponentially valuable” piece, consider how a tight-knit guild of fifty glassblowers (or weavers, or…) in a single city exerts more economic influence than five hundred scattered weavers who never interact.

On spherical cows

Metcalfe’s Law holds that the value of a network grows proportionally to the square of connected nodes. Originally formulated in the 1980s by Robert Metcalfe for selling Ethernet, it was later repackaged as the value of a website growing non-linearly as it added users. Basically if a network doubles from 100 to 200 users, its value quadruples rather than doubling.

Chen says that this is basically a spherical cow, which if you’re not familiar with those it’s a physics joke about oversimplified models. “Assume a spherical cow of uniform density.” Physicists use it to mean a model that strips away so much real-world complexity it barely resembles the thing it describes, but still captures a useful directional truth. Whenever I argue with economists, I like to point out that humans (with all our lovely, irrational behaviors) are not spherical cows.

Metcalfe’s Law assumes every connection is equally valuable, which is obviously false. Your relationship with your best friend is worth more than your connection to the person you met once at a conference. Real networks are messy, with clusters of dense connections and vast stretches of indifference. The law is directionally useful because more users generally means more value, but the specifics of how and where the valuable bits end up being are more complicated than power laws, especially in the beginning stages of the network before statistics start to matter.

Ok but was it good?

In addition to being unusually well-written and engaging, this book is broadly useful in a way that’s rare for business books. The standard audience is obviously anyone building a marketplace, platform, or product. Beyond that, it’s useful for understanding why dominant platforms behave the way they do — why they degrade over time, why they resist interoperability, why they acquire competitors rather than competing with them. This book gives you the structural vocabulary to discuss those dynamics instead of vaguely gesturing at “monopoly” or “greed.”

If you think about coordination problems, collective action, or institutional design, the network-effects framework maps onto those domains in ways Chen doesn’t explicitly explore but that become obvious once you have the vocabulary. The hard side of a network is the hard side of any collective action problem: the small number of participants who do the work and without whom the whole thing falls apart.

The concepts are more durable than the examples, which are mostly success stories. Normally I would be wary of survivorship bias, but the vocabulary was more useful than the how-to guide in some ways so the precise samples used doesn’t matter that much. I have long since forgotten what growth hacks Dropbox used. It might have been fun if there were more matched comparisons companies with similar starting positions that made different choices and diverged in outcomes. But that is a different book, and a less readable one (too much algebra)... 7 Powers: The Foundations of Business Strategy by Hamilton Helmer :P

What I got from this book is that I’ve really started to think in terms of “growth hacks” in general, cold start problems, atomic networks, how I can’t abandon a shoddy tool because of the value of its network, institutional infrastructure, etc. The vocabulary of The Cold Start Problem has improved my understanding of everything from the 1177 BC era Bronze Age Collapse to the real value of how well my son’s kindergarten classmates are already integrated into my social group. I highly recommend checking it out, or at least Andrew Chen’s articles about why it’s hard to evaluate new social products, how the supply side is king, and what the power user curve really means.

1. Interestingly, people are doing something similar to direct AI in a RTS-style interface.