Network Effects for Startups: The Math of Defensibility
Most founders talk about viral growth. Far fewer design for network effects, and that is where the durable value sits. Virality brings users in. A network effect is the reason they stay, because each new user makes the product more valuable for everyone already there. This guide covers the maths behind that, which types of network effect matter for AI products, and a 30-day plan to test whether you really have one.
Virality grows you. Network effects keep you.
Virality (acquisition)
- One user brings in another
- Referrals, invites, sharing loops, incentives
- Lowers the cost of getting customers
- Stops working the moment you stop feeding it
Network effects (retention)
- Every new user adds value for existing users
- Friends, liquidity, shared data, an ecosystem
- Makes leaving feel like a loss
- Keeps compounding after growth slows
A simple test: if growth stopped tomorrow, would the product get worse for the people already using it? If not, you have virality without defensibility. You are in the “fresh produce” business, constantly shipping new content or features to give people a reason to come back.
The four defensibilities left in a digital market
Digital distribution has eroded the old moats of geography, shelf space and capital. Four remain: scale (cost advantages at volume), embedding (hard to rip out of workflows), brand (trust and preference) and network effects. In the NFX data, the companies built on embedding, scale and brand typically topped out at $1–2 billion. The outliers were network businesses, because their value can grow faster than their costs.
The network maths, drawn
Four models describe how value scales with the number of users (n). Treat them as lenses, not laws.
| Model | Value grows with | At 10 users | At 1,000 users | Typical product |
|---|---|---|---|---|
| Sarnoff | n | 10 | 1,000 | Broadcast and media audiences |
| Odlyzko–Tilly | n log n | 23 | 6,908 | Most real networks, where a few links matter most |
| Metcalfe | n² | 100 | 1,000,000 | Direct one-to-one connections, such as messaging |
| Reed | 2ⁿ | 1,024 | Astronomical | Communities where users form groups, teams and rooms |
The founder insight: do not put Metcalfe or Reed into a financial model. Plan with something closer to n log n and let the upside surprise you. What the curves do tell you is where to design: products where users form groups (Reed) or direct links (Metcalfe) have far more headroom than products where users only consume (Sarnoff).
The network effects worth designing for
| Type | How it works | Founder move |
|---|---|---|
| Direct | Personal, utility, protocol and physical networks. Value comes straight from who else is on it. | Make leaving costly in relationships, identity or daily coordination. |
| Marketplaces and platforms | Two-sided marketplaces, market networks and developer platforms. Liquidity decides the winner. Some markets plateau once service is “good enough”. | Win liquidity in one niche before adding features. Watch for the asymptote. |
| Data and performance | Usage improves speed or quality for everyone. Critical for AI products. | Build a loop where each user’s activity improves the next user’s outcome. |
| Social and psychological | Language (“just Google it”), bandwagon and belief effects. | Use them to accelerate adoption, but only on top of a core network effect. |
“We have a data moat”: three things founders confuse
| Level | What it means | How to spot it |
|---|---|---|
| Data network effect | Usage automatically improves the product for all users, often in real time. Value grows geometrically. | Users say “it’s good because everyone uses it”. |
| Data scaling | More data makes the product better over time, with diminishing returns. Value grows linearly at best. | The 1,000th data point helps less than the 10th. |
| Data embedding | You hold critical operational data, so switching is painful. | Retention comes from migration cost, not rising value. |
Embedding is robust but it is not a network effect. The strongest AI defensibility combines a real data network effect with deep workflow embedding. To check for the real thing, use the Waze test: is data captured automatically as people use the product, does it improve outcomes quickly, does value keep growing rather than plateau, is the data central to the product, and can users feel that collective usage makes it better?
How to start a network from zero
- Single-player mode. Be useful to the very first user on their own. Tool first, network second.
- Subsidise the hard side. In marketplaces, supply is usually harder to win than demand. Secure it with credits, guarantees or done-for-you onboarding.
- Use the technology window. Generative AI is resetting workflows. Use the reset to deliver a step change in speed or outcomes that pulls people onto a new network.
- Move beyond fresh produce. If your agent generates content, shift towards user-to-user interaction, shared data loops, templates and integrations that others build on.
Your 30-day network effect plan
- Week 1: Score yourself. Answer the seven questions below with your team. Two or fewer yeses means you have growth tactics, not defensibility.
- Week 2: Map your nodes. List who joins (users, suppliers, developers, data sources), what each gives the others, and which side is hardest to win.
- Week 3: Design one loop. Pick a single mechanism, such as single-player value, a hard-side subsidy or a shared data loop, and ship the smallest version.
- Week 4: Measure density, not just growth. Track the share of users with at least one active connection, fill rate or time-to-match, and whether retention rises as networks get denser.
The seven-question self-test
- Do users stay because other users are there?
- Does value rise for existing users when new users join?
- Does leaving mean losing connections, liquidity or collective intelligence?
- Does usage create data that improves outcomes for everyone, quickly?
- Does value keep growing past “good enough”, or does it plateau?
- Are we embedded in workflows so removal is painful?
- Can others build on top of us while we sleep?
Virality can ignite growth. Network effects, paired with embedding and data loops, are what make it last. For agentic products the goal is not to generate more. It is to design interactions so the system gets more valuable as the network grows.
Defensibility is a portfolio decision. Strategy Execution turns it into first moves you can start now, each with the outcomes and metrics that prove it. Explore Strategy Portfolio →
We help founders map their nodes, pick the right loop and measure whether it is compounding.
Talk to usSources: NFX, 70% of value in tech is driven by network effects; NFX, The Network Effects Manual; Briscoe, Odlyzko and Tilly, Metcalfe’s Law is Wrong (IEEE Spectrum, 2006).
