TL;DR: A product with network effects becomes more valuable as more people use it. But not all network effects are the same kind of thing — and the social variety can actually reverse. Knowing which kind you have determines whether your moat compounds or unwinds.

What it means

Network Economies are one of the seven-powers: the Benefit is that each new user adds value to all existing users. The Barrier is that a challenger has to somehow reach critical mass to offer comparable value, which is enormously hard once you’ve already crossed it. In Helmer’s framework, network effects emerge during the takeoff phase — there’s a singular window to establish them, and missing it means losing the opportunity for good (seven-powers, power-progression).

The canonical examples — telephones, fax machines, social networks — make network effects feel straightforward and inevitable. They are not. The interesting distinctions determine whether your network effect is an asset or a liability dressed up to look like one.

The argument

Direct vs. indirect. Direct effects mean users benefit from other users directly (messaging, social). Indirect effects mean a growing user base attracts complements that benefit everyone (app stores, marketplaces, developer ecosystems). Indirect effects are generally more durable because they create a two-sided dependency that’s much harder to unwind in any single move.

Utility networks vs. social networks. This is the distinction most founders miss, and it’s the one that matters most. Utility networks — payments, communications infrastructure, developer platforms — get monotonically stronger with scale. More users always equals more value. Telephone numbers don’t get less useful when more people own phones. Social networks don’t follow this pattern at all (status-as-a-service).

Social networks are driven by status dynamics. Early users gain social capital from being first. As the platform grows, that capital gets diluted. The cool people leave to find somewhere more exclusive. The same network effect that built the platform now reverses: more users means less value for the users who matter most. This is why social networks are inherently cyclical (Friendster → MySpace → Facebook → Instagram → TikTok → next) while utility networks are not (Visa is forever).

You don’t always need everyone. The “billion-user startup” assumption — that network effects require massive, universal adoption — is wrong for most categories. Activity-based networks like dating apps, gaming lobbies, or niche professional tools need a much smaller critical mass. A dating app works if it has enough attractive matches in your specific city. A gaming network works if it has enough players to fill a lobby right now. The critical mass is defined by the activity, not by total addressable market (end-of-billion-user-startup).

This has strategic implications most founders ignore. If your network effect kicks in at 10,000 users in a metro area rather than 100 million globally, your go-to-market should be hyperlocal and your strategic horizon completely different. The moat is smaller but it’s also faster to build, which is often the better trade.

Status games create engagement but not durability. Platforms that lean into status mechanics — likes, followers, leaderboards — generate intense short-term engagement. But status is zero-sum. One person’s gain is another’s loss. Over time this creates toxicity, performativity, and exhaustion. The users who drove growth burn out or move on, and the network effect becomes a network reflex against the product (status-as-a-service).

The implication: if you’re building on network effects, know which kind you have. Build for utility if you want durability. Use status mechanics carefully and be prepared for them to decay on a timeline shorter than your fundraising horizon.

The AI twist

There’s a new category that didn’t exist five years ago and that the original network-effects literature doesn’t cover. Victor Stepanov calls them relationship effects (never-go-viral-ai): the memory, personalization, and trust that accumulate between an individual user and a model over time. This is not a classic network effect — it doesn’t get stronger with more users. It gets stronger with more time. Every conversation deposits context that competitors literally cannot replicate without time travel.

ChatGPT’s moat is built on this (chatgpt-pmf). After two years of conversations with ChatGPT, switching to Claude means starting a new relationship from scratch. The switching cost isn’t can I export my data (you can). It’s do I want to teach a new model what I’m like. Most users won’t.

Loose threads

  • For the mechanics of building network effects from zero, see atomic-network and cold-start-problem. The atomic network concept explains how to survive the anti-network effects that kill most new networks before they ever reach scale.
  • AI is enabling another new form: model improvement from usage data. More users → more data → better model → more users. This looks like a classic flywheel, but open-source models may erode data advantages faster than expected (google-no-moat). The real question is whether the data pipeline — not the dataset — is the durable data-moat (cheating-is-all-you-need).
  • counter-positioning against a platform with strong network effects is exceptionally hard. The incumbent’s users are the barrier, not their code or brand. You have to make the users themselves want to leave.

What links here