The Data Flywheel Has a Reverse Gear

Character.AI filed for Chapter 11 on October 12. Five years old, about $193 million raised, a $1 billion valuation at its peak, and an opening bid from Disney of $20 million for a non-exclusive license to some of its IP.

Deepak Mukunthu, who leads the Agentforce AI platform at Salesforce, used the news to make the right point: audience isn't a moat. He offers a three-question test you can run on your own product today. It's a good test. So I ran Character.AI through it using the numbers from the court filings. Then I ran my own company through it.

The short version

Character.AI's bankruptcy is the cleanest AI moat case study of the year. Run through Deepak Mukunthu's three-question moat test, it fails all three: Google licensed its technology and rehired its founders in 2024, every hour of engagement was an inference bill ($13.8 million a month in compute against peak revenue of $5.3 million), and Disney, the IP owner it once had to remove characters for, is now the stalking-horse bidder. The lesson the test misses is the data. Character.AI held a data asset nobody else had, and in bankruptcy court it is the thing state attorneys general are trying to keep away from a buyer. A data flywheel has a reverse gear, so the fourth question is whether your most valuable data is an asset or a liability in a sale. I ran all four questions on Heidi, the company brain Falkster.AI builds, and my weakest answer is the network-effect question, by design.

Question 1: If a competitor raised $100M tomorrow, how long until they replicate your edge?

For Character.AI the honest answer is worse than "months." The edge was never the app. It was the people and the model, and in August 2024 Google paid roughly $2.7 billion to license the technology while Noam Shazeer, Daniel De Freitas, and part of the research team went back to Google.

The edge didn't need replicating. It walked out the front door with a license attached.

The users stayed. And the users were never the moat.

Question 2: Does every new user make the product better for everyone else, or just add server costs?

This is where the filings get specific. Character.AI spent $13.8 million a month on compute in 2024 and about $10 million a month in 2025. Revenue from all sources peaked at $5.3 million a month. The company was losing $8 to $9 million a month when it filed.

Active users averaged around two hours a day. That sounds like the dream until you price it. Every hour of engagement was an inference bill, and the subscription and ad revenue never caught up.

One note on scale: Deepak's post cites 163 million users. The court filings put monthly actives near 12 million, down from roughly 20 million at the peak. Either number makes his point. Time spent is a vanity metric when the marginal user is marginally unprofitable. I made the same argument about feature ranking in Rank AI Features by Downside Exposure, Not Engagement, and Character.AI is that argument at company scale.

Question 3: If your biggest supplier or distributor became a competitor overnight, what breaks?

Character.AI trained its own models, so on paper it looked safe here. Now look at who shows up in the bankruptcy. Google holds a license to the technology and employs the founders. Disney, which sent a cease-and-desist over its characters in 2025, is the stalking-horse bidder. The most recognizable content on the platform belonged to other people.

When the IP owner decides to play, what breaks is the catalog.

Three questions, three failing answers. Deepak's test works.

The part the test misses

Character.AI did have the thing every moat framework puts at the top of the list. It had a data asset nobody else had: millions of people, many of them young, talking to its characters for hours a day about the most personal parts of their lives. On a whiteboard that's the textbook data flywheel.

In bankruptcy court it's the opposite. Dozens of state attorneys general have intervened to demand privacy and consumer protections in any sale, warning that users' conversations could end up with a new owner. The company has also faced lawsuits from families alleging its chatbots contributed to teen self-harm and suicide, and settled several of those cases earlier this year. The most valuable data Character.AI holds is the data a buyer can least afford to touch.

That's the reverse gear. A data flywheel spins in both directions. Every conversation that could have made the product smarter also made the company's exposure bigger. When trust broke, the flywheel didn't stop. It reversed, and every turn added liability instead of advantage.

So I'd add a fourth question to Deepak's three:

4. If you were sold tomorrow, would your most valuable data show up in the data room as an asset or as a liability?

If you can't say "asset" with a straight face, you don't have a data moat. You have downside exposure with a growth curve attached.

Trust is on Deepak's list of five moats. I'd move it off the list and under it. Data flywheels, workflow entrenchment, distribution, and scarce inputs all compound only while trust holds. Trust isn't the fifth moat. It's the ground the other four stand on.

Running the test on Heidi

I'm building Heidi, the company brain Falkster.AI makes for AI-native businesses, so I owe the same honesty about my own answers. This is the Building a Company in Public part: the answers before the outcome is known.

Question 1. A well-funded competitor can replicate Heidi's features. Features aren't the edge. Models are rented, and they get better and cheaper every quarter for everyone, including us. That's why the line we build against is "rent the model, own the company brain."

Question 2. This is the uncomfortable one. Heidi is multi-tenant with strict tenant isolation, and we ruled out pooling learning across tenants because it collides with that promise. So a new customer does not make Heidi better for every other customer. No classic network effect.

What compounds instead is per tenant. Every week a company runs on Heidi, its Cortex holds more of its people, process, and decisions, and the record behind every agent action gets richer. That's workflow entrenchment plus a private flywheel the customer owns. Weaker than a network effect on paper. Much stronger on question four, because the data was never ours to sell.

Question 3. Heidi runs on foundation models from companies that also want to own the workspace. If a model provider decides to build the company brain itself, anything we built on borrowed ground breaks. What doesn't break is the tenant's observation record. That is the argument in Context Is King. Written Context Is Rented.: written context ports, and the earned record of what actually happened does not.

Question 4. Tenant isolation, receipts on agent actions, and a guard layer around everything an agent does aren't compliance overhead. They're how the data stays an asset.

My weakest answer is question two. I'd rather know that now than read it in a court filing.

Run it this week

This takes an afternoon if you use real numbers instead of vibes.

• Question 2: compute contribution per active user with inference included. If it's negative, engagement growth is burn growth.
• Question 3: list your top three dependencies (model, distribution, content) and write down what each one could take from you.
• Question 4: imagine the data room. Which tables would a buyer want, and which would a regulator ask about? If they're the same tables, design for that now, while it's cheap.

Character.AI had two hours a day of its users' attention and drew a $20 million opening bid. The gap between those two numbers is the whole lesson.

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

Sources: Deepak Mukunthu, three-question moat test, LinkedIn, October 2026. TechCrunch, "Just 5 years in, interactive chatbot platform Character.AI files for bankruptcy," October 13, 2026. NBC News, "Character.AI files for bankruptcy, will sell some assets to Disney for $20 million." SAN, "Character.AI's bankruptcy leaves millions of users' chats in limbo," for the court-filed compute and revenue figures. KQED, "Your Chatbot Just Went Bankrupt. What Happens Now to Your User Data?" PYMNTS, "Google Reportedly Spent $2.7 Billion to Rehire Character.AI Founder." 36Kr, "Character.AI bankruptcy," for the 12 million monthly actives in the filing.