Network effects
A product has a network effect when each new user makes it more valuable for existing users. The five types, how to spot fake ones, and the cold-start problem.
A product has a network effect when each new user makes it more valuable for the users who are already there. The one-line test: would an existing customer get more value, or be less likely to leave, because an unrelated customer joined?
A phone is useless if you are the only owner and essential when everyone you know has one. For software founders, network effects matter because they are one of the few advantages a better-funded competitor, or a customer with an AI coding agent, can't copy by writing code. They are also one of the most over-claimed ideas in startups. This guide covers how network effects work, the five kinds you will meet, how to tell a real one from a lookalike, the cold-start problem, what all this means for B2B and bootstrapped SaaS, a worked example with numbers, and how founders and investors were talking about network effects in September and October 2026.
How network effects work
NFX, the venture firm that has written the most on the subject, uses a simple definition: network effects occur when a company's product or service becomes more valuable as usage increases. It traces the first description to 1908, when AT&T chairman Theodore Vail noticed how hard it was for other phone companies to compete once AT&T had more customers in a given area. A technically better phone couldn't make up for the missing network.
The math behind the idea is about connections. Robert Metcalfe, who created the Ethernet standard, proposed that a network's value grows with the square of its users, now called Metcalfe's law. The intuition is easy to check: with n users there are n × (n − 1) ÷ 2 possible connections. Ten users have 45. A hundred have 4,950. A thousand have 499,500. Users grow by 10x each step, while possible connections grow by about 100x.
Possible connections aren't real ones, though. NFX's own Network Effects Bible stresses that the effect comes from use, and Andrew Chen, author of The Cold Start Problem, says that when he writes "more users" he means more density and interconnection, "Not scale effects". A network of 50,000 people who never interact is weaker than one of 500 who use it every day.
NFX also claims that network effects account for 70% of the value created by tech companies since 1994. Treat that figure as NFX's own study, not a law. Writing in Lenny Rachitsky's newsletter, Dan Hockenmaier lists network effects as one of only four moats that compound, alongside switching costs, scale economies and data.
The five kinds of network effect
NFX's manual catalogs 16 types in five broad groups, and the list keeps growing. For a SaaS founder, five kinds cover almost every real case.
Direct
Each new user can reach every existing user, so value rises with the number of people you can connect with. Phones, messaging apps and shared file formats work this way. Inside a company, a team chat tool gets more useful with every colleague who joins it.
Two-sided (indirect)
Two different groups attract each other. Buyers come because sellers are there, and sellers come because buyers are. Marketplaces like eBay and Craigslist are the classic case. NFX calls the same pattern a two-sided platform when the supply side builds things that only work on that platform, such as apps for an operating system. In SaaS, that's the template gallery, plugin store or integration directory that brings users because builders are there, and brings builders because users are.
Data
Usage produces data, and the data makes the product better for everyone. NFX's best example is Waze: nearly every driver who reads traffic data also adds to it, and because the data is used in real time it has to be updated all the time. NFX also warns that data network effects "tend to be weaker" than many investors believe. More data often stops helping after a point, and if usage doesn't create the data, you have a scale advantage, not a network effect. Hockenmaier treats data as its own compounding advantage, and so does the learning loop idea from Rob Walling and Einar Vollset.
Local
Only the users near you, or in your circle, add value. Chen, who worked on growth at Uber, describes it as "a vast worldwide network of smaller networks", one per city. A rider in San Diego gets nothing from a new driver in Boston. In B2B, the same thing happens inside an industry or a supply chain: a restaurant cares about its own suppliers, not every supplier in the country. Local networks are easier to start, because you only need density in one place, and easier to attack for the same reason.
Community
Members help, teach and make things for each other: forums, template galleries, meetups and Q&A threads. NFX files related patterns under "social" network effects (tribal, belief and bandwagon). Pieter Levels lists this moat as "Network effects / community: users make it more valuable for each other". Community is the kind most open to small companies, because it doesn't require millions of users or two sides. It does need a core of active contributors, which Chen calls the "hard side" of the network.
The types overlap, and strong companies usually stack several. NFX's point is that once one network effect is working, it becomes easier to build others on top of it.
How to tell a real network effect from a lookalike
Chen opens The Cold Start Problem by describing pitch meetings where "network effects" was the answer to every hard question, with nothing to measure it. Four questions, asked in order, catch most false claims.
- Does a customer get more value when an unrelated customer joins? If customer 1,000 gets nothing from customer 1,001, you have something else. Most B2B tools are used by one company on its own. Your invoicing software doesn't improve when a stranger signs up.
- Does value grow with usage and connections, not just sign-ups? A big but inactive network is hollow. NFX: "Networks without effective retention have weak network effects."
- Does value keep rising as the network grows? Some networks flatten early. NFX's example is ride-hailing: shorter waits matter a lot up to around four minutes, then barely at all. Past that point, a rival with half as many drivers can offer the same service.
- Is it costly to use a rival network at the same time? NFX calls this multi-tenanting. Drivers run Uber and Lyft side by side, and sellers list on eBay and Etsy. The easier it is, the weaker the moat.
A "no" doesn't mean you have nothing. It means you have one of these lookalikes, which are worth having but work differently:
- Virality. Users bring in more users, measured by the viral coefficient and built with a viral loop. NFX is blunt: "Network effects are not viral effects." Virality is about getting users for free. Network effects are about keeping them. A product can go viral with no network effect, and a marketplace can build a network effect with paid ads and zero virality.
- Scale. Bigger companies get lower unit costs. NFX notes that scale and network effects are often confused because both get stronger with more users.
- Switching costs. A customer stays because leaving is expensive, not because the other users are valuable to them. See switching costs and vendor lock-in.
- Loops and flywheels. A growth loop or flywheel describes how growth feeds itself. Hockenmaier calls every compounding advantage a self-reinforcing flywheel, but a flywheel only involves a network effect if the users themselves create value for each other.
Once a product passes all four questions, measure it. Chen suggests tying network effects to ordinary metrics instead of treating them as a vague idea. More users should mean lower CAC and more sign-ups, higher retention and usage, and higher revenue per user. If your densest segments don't beat your sparsest ones on those numbers, the network effect isn't doing much yet.
The cold-start problem
Network effects work against you at the start. A video app with few videos, or a marketplace with few sellers, gives early users little reason to stay, and every user who leaves makes the product worse for the rest. Chen calls these "anti-network effects", and the framework of his book starts from the observation that "Most networks fail at the beginning" (Andrew Chen, 2021).
NFX defines critical mass as the point where the network's value beats the product on its own and beats competing products. OpenTable is the classic delayed case: it needed many restaurants in a city before diners saw any value, and then the effect became strong. Until you reach that point, a good single-user tool beats your network.
The standard playbook, drawn from Chen and NFX:
- Build the smallest network that works. Chen's "atomic network" is the smallest network that is stable and can grow on its own. A Zoom call works with two people. Airbnb needs hundreds of active listings in a market. Pick a niche small enough that you can make it dense, the same logic as a beachhead market and crossing the chasm.
- Win the hard side first. Chen's "hard side" is the minority of users who create most of the value and are hardest to attract: creators on a video platform, sellers in a marketplace, the project managers who set up a workplace app. At Uber, he writes, "Power Drivers" made up 20% of supply but created 60% of trips. Solve a real problem for that group and the other side follows.
- Come for the tool, stay for the network. NFX says it knows at least 19 tactics for the chicken-and-egg problem, and several start with a single-player mode: a product that is useful to one user before anyone else arrives. For SaaS this is often the natural path. Build a tool people pay for alone, then add the shared layer.
- Grow one network at a time. Uber's weekly reviews looked at each city as its own network. Chen writes that the team regarded company-wide totals as "mostly meaningless". Do the same with your industries, regions or communities.
Network effects in B2B SaaS
Most B2B SaaS doesn't have a network effect between customers, and that's fine. The ones that do tend to fall into a few patterns:
- Inside each account. Team chat, docs and project tools get more useful as more colleagues join. This makes the account sticky, though it rarely helps you win the next account.
- Across companies. Products that sit between businesses, such as ordering, invoicing with a shared portal, scheduling with clients, or e-signature, can link companies to each other. A supplier on your platform makes it more useful to every restaurant that buys from that supplier.
- Expertise. NFX describes an "expertise" network effect for professional tools such as QuickBooks, Salesforce, HubSpot, Excel, WordPress and Figma. The more professionals who know the tool, the easier it is to hire for, and the more companies pick it.
- Ecosystems. Integrations, templates and plugins that third parties build for your product work like a two-sided platform.
Hockenmaier describes how differently network businesses are run. At Faire, a marketplace connecting independent retailers and brands, most of the roadmap went to getting more retailers, getting more brands and their products, improving discovery, and lowering retailers' risk with free shipping and payment terms.
Every decision gets filtered through one lens: "Does this strengthen the network?"
Dan Hockenmaier / How to tell if a startup has a moat, Lenny's Newsletter (October 13, 2026)
He goes as far as saying that in network businesses most product features, and even the UI, matter very little. He points to Craigslist, which still gets heavy traffic despite its dated design. A feature-led SaaS company should not assume the same: if you don't have a network, features and workflow are still what keep customers.
Network effects for small and bootstrapped products
You rarely need a true network effect to build a good small business, and chasing a two-sided marketplace without money is one of the hardest things a bootstrapper can try. In episode 836 of Startups for the Rest of Us (June 9, 2026), Walling says the odds are close to zero unless you already have access to at least one side of the market.
I’ve known maybe two people who’ve bootstrapped successful two-sided marketplaces
Rob Walling / Startups for the Rest of Us, episode 836 (June 9, 2026), from 18:55
His exceptions are founders who already have access to one or both sides, such as Dynamite Jobs, which connects companies hiring with remote job seekers. Vollset adds the other half of the trade-off: a marketplace is hard to build, but hard to undo once it works. He now sees a marketplace component next to a core product as a possible moat, because as the marketplace grows it makes the software stickier.
The smaller forms are more realistic and still add up:
- Community. Levels made the nomads.com community nearly free (from $99 to $1, kept only to stop spam) and wrote "the more people, the more network effect". He also runs in-person meetups. A forum or Slack group where customers answer each other's questions gets better with each active member, and each member has another reason to stay.
- Templates and shared content. A gallery of user-generated templates, recipes or workflows makes the product more useful to each new customer, and a clone starts with an empty gallery.
- Integrations. When other tools build to your API, each integration makes your product more useful and makes you the default in the next customer's stack.
- Benchmarks. Pooled, anonymised data that shows each customer how they compare is a small data network effect, and the start of a learning loop.
- Multi-user accounts. If a product only one person uses becomes something a team uses, leaving becomes a group decision, which is the "everyone's already in it" moat at small scale.
Each of these supports a wider moat without needing millions of users.
Worked example: testing for a network effect
Fictional example. The product, numbers and price are invented to show the method.
A small team sells an ordering tool to independent restaurants for $79 a month. Restaurants place orders with their suppliers in the app. Suppliers can join for free, and when a restaurant's supplier is on the platform, orders go straight through instead of by email. The founder thinks this is a network effect. Here is how to check.
Test 1: retention by density. Split the 600 restaurant accounts by how many of their suppliers are on the platform, then compare 12-month retention:
- 0 suppliers: 240 restaurants, 55% retained = 132
- 1–2 suppliers: 200 restaurants, 68% retained = 136
- 3+ suppliers: 160 restaurants, 85% retained = 136
- Overall: 404 of 600 = 67.3%
A 30-point gap looks convincing, but it could be selection: bigger, busier restaurants may simply have more suppliers and churn less anyway.
Test 2: before and after. Look at the same restaurants before and after their main supplier joined, next to similar restaurants whose supplier didn't join. The 50 restaurants whose supplier joined went from 18 to 27 orders a month: 9 ÷ 18 = +50%. The comparison group went from 18 to 19: 1 ÷ 18 = +5.6%. The difference, about 44 points, is the part you can credit to the network.
Test 3: the unrelated customer. When a supplier joins in Leeds, restaurants in Bristol that don't buy from it see no change. This is a local network effect, so the growth plan should be one city, or one supplier's customer list, at a time.
What it's worth. If the team gets 40 restaurants from zero connected suppliers to three or more, and they then retain like that group, 40 × (85% − 55%) = 12 more restaurants stay for the year. At $79 a month that's 12 × $79 × 12 = $11,376 of revenue kept, before counting new restaurants that suppliers invite. Last year 90 of 300 new restaurants (30%) arrived through a supplier invite, which is the acquisition side of the same effect.
Network effects in the AI era (September–October 2026)
As AI makes software cheaper to copy, network effects have moved back to the center of the moat discussion. Four recent takes:
"Everyone's already in it." In episode 855 of Startups for the Rest of Us (October 20, 2026), Walling and Vollset name eight SaaS moats that still work. Moat six is network effects, which Walling also calls "who has to move?" One person can switch tools in an afternoon. A team has to agree and move together. A whole company or community is a migration. Walling's example is Slack, which people keep using even though, as he puts it, "None of us really like Slack that much." Vollset's example is school parent apps.
The school doesn’t want to move a thousand students and 4,000 parents
Einar Vollset / Startups for the Rest of Us, episode 855 (October 20, 2026), from 36:08
His point is that his kid could vibe code something better in a couple of days, but the school won't ask a whole community to sign up for something new. The question for your own product: how many people would have to agree before a customer could leave? In the same episode, Vollset describes software that learns from its customers' use as "similar-ish to a network effect", which is the learning loop moat.
Levels' community moat. On October 12, 2026, Pieter Levels listed the moats he thinks are left, including "Network effects / community: users make it more valuable for each other". Four days later he described acting on it: making the nomads.com community nearly free and running more meetups.
Compounding advantages. Hockenmaier's October 13 framework counts network effects as one of four compounding advantages, and treats speed and distribution as useful but not compounding on their own. His advice for judging a network effect is to ask how strong it could get, noting that every marketplace has some network effect but many still fail.
Will AI agents have network effects? Chen's October 6 essay argues that general-purpose AI agents don't have them by default.
These agents are tools, not networks.
Andrew Chen / Will agents have network effects? (October 6, 2026)
His reasoning: your agent can talk to mine, just as Gmail can email Outlook, so there's no need to use the same one. He expects the fight to be over where the network effects live. An agent can try to pull identity, reputation, shared context and transactions inside its own walls, while existing networks try to stay open to every agent. For a SaaS founder, the practical reading is that a feature an agent can reproduce isn't a moat, but a network of people, data or trust that the agent has to plug into can be.
Common mistakes
- Calling a big user base a network effect when users never interact or benefit from each other.
- Confusing virality with a network effect. Invites bring users in, but they don't make the product better for those already there.
- Counting "we have lots of data" as a data network effect when usage doesn't produce new, useful data, or when more data stopped helping long ago.
- Measuring total users instead of density in each local network, industry or community.
- Launching "everywhere" instead of making one small network dense first.
- Trying to bootstrap a two-sided marketplace with no access to either side.
- Assuming a network can't be left. Cheap multi-tenanting, and agents that work across tools, can drain a network that looked locked in.
Listen, watch and read next
- Watch: The Eight SaaS Moats that Still Work in the Age of AI, Startups for the Rest of Us, episode 855, from 36:08. The "everyone's already in it" segment, with the Slack and school-app examples.
- Listen: The 5 AI Moats Acquirers Value Most, episode 836, from 17:38. Walling and Vollset on two-sided network effects and why bootstrapping a marketplace is so hard.
- Watch: Andrew Chen, The Cold Start Problem, Talks at Google (December 9, 2021). An hour on how networks start, tip and defend themselves.
- Book: The Cold Start Problem, Andrew Chen (Harper Business, 2021). Case studies from Slack, Zoom, Tinder, Uber, Airbnb and others, organized by stage.
- Read: The Network Effects Manual, NFX. The reference list of 16 types, with examples and the strength of each.
- Read: Will agents have network effects?, Andrew Chen (October 6, 2026). Where network effects might live once AI agents do more of the work.
Related terms
- Moat
- Switching costs
- Learning loop (data moat)
- Flywheel
- Growth loop
- Viral loop
- Viral coefficient (K-factor)
- Vendor lock-in
- Crossing the chasm
Sources
- The Network Effects Manual: 16 Network Effects (and counting), NFX
- The Network Effects Bible, NFX (updated 2024)
- The Cold Start Problem, Andrew Chen, Harper Business, 2021
- Previewing a full chapter of The Cold Start Problem, Andrew Chen, 2021
- Solve a Hard Problem (Tinder), chapter 8 of The Cold Start Problem, Andrew Chen, 2021
- What to look for when you're hiring a Head of Growth, Andrew Chen, 2021
- Will agents have network effects?, Andrew Chen, October 6, 2026
- Episode 855: The Eight SaaS Moats that Still Work in the Age of AI, Rob Walling and Einar Vollset, Startups for the Rest of Us, October 20, 2026
- Episode 836: The 5 AI Moats Acquirers Value Most, Rob Walling and Einar Vollset, Startups for the Rest of Us, June 9, 2026
- How to tell if a startup has a moat, Dan Hockenmaier, Lenny's Newsletter, October 13, 2026
- I'd expand the moats left to proprietary data, network effects, distribution, hardware, capital, and regulation, Pieter Levels, October 12, 2026
- I am now scrambling to chase any remaining moats, Pieter Levels, October 16, 2026
- Moats: Durable competitive advantage, A Smart Bear (Jason Cohen)
