No. 01 Foundations
Most network effects aren’t
A network effect means each new user makes the product better for the users already there. Most products have something else instead.
When we hear the word network, most of us think of mobile operators. But does a mobile network become more valuable to the people already on it each time someone new joins?
In February 2017, Reliance Jio announced that it had signed up 100 million subscribers in 170 days. Its press release compared the pace with Facebook, WhatsApp and Skype, and on speed alone the comparison was fair.1 Very few consumer businesses anywhere have drawn a curve like that. About three years after launch, Jio had more mobile subscribers than any other operator in India.
There was no network effect in it.
A Jio SIM could call any phone in the country, on any network, free. What it was worth to you did not depend on how many other people had one. If half of Jio’s subscribers had vanished in the middle of 2017, the other half would have lost nothing that mattered, and their data would probably have gotten faster, since they’d be sharing the same towers with fewer people. The growth came from price and from an enormous network build, helped along by an Aadhaar-based sign-up that removed most of the paperwork. Those are real advantages. They are advantages of scale, execution and Mukesh Ambani’s sheer tenacity.
The one place a network effect could have lived in that market was the wiring between operators. If calls from Jio to the incumbents’ customers kept failing, a Jio connection would be worth less, and the incumbents’ size would start to count for something. The first big fight of the launch was over exactly that. In October 2016 the regulator recommended a combined penalty of ₹3,050 crore on Airtel, Vodafone and Idea for denying Jio enough points of interconnection.2 When networks fight about interconnection, they are fighting about who owns the network effect. In Indian telecom the answer was everybody, which means nobody in particular.
I start with Jio because it’s the cleanest case I know of a business that grew faster than anyone imagined without having a network effect. It did have something else: the free-data offer spread by word of mouth, which is virality, not a network effect (more of that difference later).
The halving test
A network effect exists when one more user makes the product more valuable for the users who are already there. Economists have worked on this at least since Jeffrey Rohlfs’s 1974 paper on the demand for communications services, and Michael Katz and Carl Shapiro gave it its standard treatment in 1985.3 The definition fits in one line. It is surprisingly easy to lose once a growth chart is on the screen.
The test I keep coming back to is simple. Picture half of a product’s users gone tomorrow morning, picked at random. Keep everything else exactly as it is. Then ask whether the product is worse for the users who remain.
For WhatsApp the answer arrives in minutes. Half your family group goes silent. For a well-built SaaS tool with a strong brand, the remaining customers notice nothing; the software does on Monday what it did on Friday. Netflix is the interesting one. Next day nothing changes, but a year later the service is worse as half the revenue is gone and the content budget with it.
Now ask how this damage reaches the users who stayed. If it travels through the company’s income statement (less revenue, so a smaller content budget), you are looking at economies of scale. If it travels directly from one user to another (fewer people to message, fewer recruiters to hear from), you are looking at a network effect. Netflix gets better with size because it spreads a fixed content bill over more subscribers. Your neighbor subscribing does nothing for your evening. There is a social effect from engagement perspective, since people want to watch what their friends are talking about, but it attaches to the show more than to the service, and it fades.
Scale is a perfectly good moat, with different rules. It rewards being big rather than dense, anyone with large enough check can buy it, and it doesn’t compound the way a true network effect does.
Things that get called network effects
Scale is the thing most often mislabeled as a network effect. Four others come up all the time. They just deserve a different framework.
A brand gets mistaken for one because both make customers pick you without being sold to. But a brand is a belief each customer holds on their own, and it would come through the halving test untouched.
Switching costs look similar from the outside, since both keep customers from leaving. Think of the bank account that receives your salary and pays out your EMIs and a handful of UPI autopay mandates. Moving it would cost you a weekend of forms, and that keeps you there whether or not anyone else banks with it. Switching costs hold on to the users you have. They don’t make the product better for the next one.
Data is the trickiest of the four, because the loop sounds identical: more users, more data, a better product, more users. Sometimes that loop is real. More often the learning curve flattens long before people stop citing it, and the data mostly improves the product for the person who generated it, which is personalization, a close cousin of switching costs. I’ve written separately about when data network effects hold.
Virality is the confusion that costs the most. A viral product spreads because users bring in other users. A network-effect product gets better because users bring in other users. The first is about what it costs to acquire the next customer; the second is about what each customer gets. Prisma, the app that turned photos into something like paintings, was downloaded more than seven million times in the five weeks after its June 2016 launch. Its makers did it again in late 2022 with Lensa’s AI avatars. Neither app was any better for you because your friends had it. Your friends were the distribution, and that was all they were.
You can have a growth loop without a value loop, and the reverse. Some of the strongest network effects grow slowly, because each new user is valuable only to the small set of people who want to reach them, and assembling that set takes years. On a matrimony platform, ten thousand active profiles within one community, language and city take patience to build. Once they’re there, that platform can be a far better product for the people in it than a much larger site where the same people are spread thin.
The trouble with n squared
The pitch-deck version of all this is Metcalfe’s law: a network’s value grows with the square of its users, because n users can form roughly n² connections. It started in the early 1980s as a slide Robert Metcalfe used to sell Ethernet, and George Gilder gave it the name in 1993.4 It’s a lovely story for a fundraise. Double your users and your value quadruples.
The weak point is an assumption buried inside it, that every possible connection is worth the same. In 2006 Bob Briscoe, Andrew Odlyzko and Benjamin Tilly argued in IEEE Spectrum that value grows more like n log n.5 You can check their reasoning on your own phone. Rank your contacts by how much they matter to you. Your mother is near the top. Your two-hundredth contact is someone you once bought a second-hand scooter from. If the k-th person on that list is worth about 1/k of the first, the total value of the network to you grows like a logarithm: it keeps rising, slowly, and each person added later adds less than the one before. (Metcalfe has defended his law, and in 2013 fitted it to Facebook’s revenue. I suspect much of the disagreement is about which part of the curve you happen to be standing on.)
For a business the useful version is simpler. Value depends on who connects with whom, and headcount is a poor proxy for that. A chartered accountant in Indore with eight years of experience gets nothing from a million software jobs in Bengaluru. They need enough relevant openings in their city and experience band, and once they have twenty to choose from, the twenty-first adds much less than the third did. That is what most real network effects look like: local, and saturating. The local part is the subject of a later essay, Density beats size.
Run it twice
The halving test tells you whether you have a network effect. It doesn’t tell you whether it’s still doing anything for you. For that I run the same thought experiment the other way: double the users overnight, and ask whether the product gets better for the ones already there.
Put the two answers together and you can tell roughly where you stand on the curve. If halving hurts and doubling helps, you’re on the steep part, and growth compounds: each new user makes the product better, which brings in the next one. If halving hurts and doubling does nothing, you’re on the shoulder. The network effect is real and it protects you from anyone smaller, but more users won’t make the product better. That’s a moat, not an engine, and it should be valued like one. If neither test moves anything, you are so far out on the flat part that a rival with half your users would offer the same experience, and the network effect has stopped protecting you as well.
Doubling can even make things worse. Jun Li and Serguei Netessine studied a holiday rental platform that took in a large batch of listings migrated from other sites in one go. They estimated that doubling the size of the market cut the rate at which travelers got a confirmed booking by about 15%, and hosts’ occupancy fell by a similar amount. Travelers searched about 18% more, and fewer of them ended up booking.6 More choice made matching harder. Anyone who has worked on a matrimony inbox, or on a job posting with three thousand applications, has seen this from the inside. When it goes far enough the loop starts running backwards, which is what Networks unravel, too is about.
In a two-sided market, run both tests on each side separately, because the answers rarely match. WhatsApp’s network effect is same-side (economists also say direct): more users make it better for users like them. A jobs platform’s is mostly cross-side. Recruiters come for candidates and candidates come for jobs, and the two pulls are not equally strong. Halve the recruiters and job seekers feel it within days, because the openings they were waiting for stop appearing. Halve the job seekers and a recruiter hiring for a popular role might not notice for a month; they already had more applications than they could read. A recruiter looking for a rare skill notices by lunch.
Same platform, same morning, three answers. When someone asks me whether a business has a network effect, that is roughly the form an honest answer takes: yes, for this side, in these cities and categories, up to about this size.
Notes
Reliance Jio Infocomm, media release, “Jio family 100 million strong in 170 days”, Mumbai, 21 February 2017. In TRAI’s subscriber data for November 2019, Jio passed Vodafone Idea to become the largest mobile operator in India by subscribers. ↑
The recommendation went from TRAI to the Department of Telecommunications: ₹50 crore per licensed service area, which came to about ₹1,050 crore each for Airtel and Vodafone and ₹950 crore for Idea. The department never cut the figure. In 2021 it sent demand notices for the full amount, which Airtel and Vodafone Idea challenged at the telecom tribunal. Briscoe, Odlyzko and Tilly (cited below) tell older versions of the same fight, between the early Bell System and the independent telephone companies, and among CompuServe, Prodigy and AOL. ↑
Jeffrey Rohlfs, “A Theory of Interdependent Demand for a Communications Service”, Bell Journal of Economics 5(1), 1974, pp. 16-37. Michael L. Katz and Carl Shapiro, “Network Externalities, Competition, and Compatibility”, American Economic Review 75(3), 1985, pp. 424-440. For a business reader, Carl Shapiro and Hal Varian’s Information Rules (1998) is still the most readable book on the subject. ↑
Bob Metcalfe tells the origin story, and fits the law to Facebook’s revenue, in “Metcalfe’s Law after 40 Years of Ethernet”, IEEE Computer 46(12), 2013, pp. 26-31. Gilder named it in Forbes ASAP in September 1993. ↑
Bob Briscoe, Andrew Odlyzko and Benjamin Tilly, “Metcalfe’s Law is Wrong”, IEEE Spectrum, July 2006. ↑
Jun Li and Serguei Netessine, “Higher Market Thickness Reduces Matching Rate in Online Platforms: Evidence from a Quasiexperiment”, Management Science 66(1), 2020, pp. 271-289. ↑