What is the fourth moat question?

If you were sold tomorrow, would your most valuable data show up in the data room as an asset or as a liability? It extends Deepak Mukunthu's three-question test (replication time, whether users improve the product or only add cost, and supplier risk). If you cannot answer asset with a straight face, you have downside exposure with a growth curve attached.

How does Falkster.AI answer the network-effect question for Heidi?

Honestly, it fails it on purpose. Heidi is multi-tenant with strict tenant isolation, and pooling learning across tenants was ruled out because it collides with that promise. What compounds is each tenant's own company brain and the record behind every agent action. Weaker than a network effect, stronger on the fourth question, because the data belongs to the customer.

Why Did Character.AI Go Bankrupt?

Why did Character.AI go bankrupt with millions of users?

THE SHORT ANSWER

Because users were a cost, not a moat. Court filings show Character.AI spending $13.8 million a month on compute in 2024 against revenue that peaked at $5.3 million a month, so two hours a day of engagement was an inference bill. Its edge left in 2024 when Google licensed the technology for about $2.7 billion and rehired the founders, and its most valuable asset, millions of intimate conversations, became the thing state attorneys general intervened to keep from a buyer. A data flywheel has a reverse gear: when trust breaks, every interaction adds liability instead of advantage.

Character.AI filed for Chapter 11 on October 12, 2026 with about 12 million monthly active users, down from roughly 20 million at the peak. Users were not the problem and they were not the moat. Compute cost $13.8 million a month in 2024 against revenue that never passed $5.3 million a month, so the two hours a day users spent on the platform was a bill, not an asset.

The lesson I took from it is the data. Character.AI held conversations nobody else had, and in bankruptcy that data is what state attorneys general are fighting to protect from a buyer. That is why I added a fourth question to Deepak Mukunthu's moat test: is your most valuable data an asset or a liability in a sale? I ran it on Heidi, and the honest answer to the network-effect question is that we do not have one, by design.

The full case, question by question, is in The Data Flywheel Has a Reverse Gear. The context argument behind Heidi's answer is Context Is King. Written Context Is Rented.

Last reviewed 2026-10-19 · 1 min read

Why did Character.AI go bankrupt with millions of users?

SOURCES

THE LONG VERSION

RELATED ANSWERS