Network effects

A single telephone is useless; the millionth telephone makes every other telephone more useful. That’s the whole idea. Each new user adds value for the existing users, which attracts more users, which adds more value — a demand-side flywheel.

The economics are brutal once it tips: the largest network is the most valuable, so users keep joining the largest network, so it gets more valuable still. Network-effect markets tend toward winner-take-most. Meta is the canonical example — you’re on the platform because everyone you know is on the platform.

• Direct: more users of the same type make the product better for each other — phones, messaging, social graphs.
• Indirect / two-sided: more users on one side attract more on the other — marketplaces (buyers ↔ sellers), app stores (developers ↔ users), payment networks.

Both create the same self-reinforcing pull. Both are demand-side: the moat is the other users, which is why it’s so hard for a rival to dislodge — they have to replicate the network, not just the software.

“Network effects” gets used as a synonym for “moat,” which muddies thinking. Two other forces are routinely mislabeled as network effects — each is its own idea:

• More usage making the product better via a data/ML loop is not (usually) a network effect — that’s a data flywheel, a form of scale economies. The value comes back to you through a better product, not to users through each other.
• Being hard to leave is switching costs, not a network effect. Lock-in keeps users; it doesn’t make the product more valuable as the network grows.

Keep them separate and you can actually diagnose a business instead of waving “network effects” at it.

Mostly no, and this is the surprising part. Your use of a model does not make the model more valuable to another user the way your presence on a social network does. There is no user-to-user network being assembled.

The “data flywheel” — more usage produces more interaction data, which trains a better model — is real but it’s a scale economy, not a network effect, and it’s weaker than claimed: frontier capability comes mostly from pretraining scale, algorithms, and synthetic data, not from your chat logs. Genuine network effects, when they appear in AI, sit at the product layer — an agent marketplace, shared tools, a developer ecosystem — not in the weights.

That absence is load-bearing. A market without network effects doesn’t tip to one winner on demand-side gravity; it stays contestable, which is half the story of why the model layer is commoditizing. (The other half: substitutes set a price floor.)