Character.AI’s bankruptcy declaration is the first time anyone has shown the public what it costs to be twelve million people’s companion. The answer is 58 cents per person per month in compute, against 44 cents in revenue, for two hours of conversation a day. The economics of free companionship were never going to close, and the declaration shows exactly where the gap lives.
Chapter 11 filings come with a sworn narrative from an officer of the company, and Character.AI’s was signed by its chief executive, Colin Adams, on the day of the petition. It is the most detailed public account of a companion app’s finances that exists, and it is worth reading as a document about the category rather than about one company, because the company did almost everything the category believes in and the numbers still did not work.
Here is the shape of it. In August 2026, the last full month before the filing, Character.AI took in about 5.3 million dollars: 2.8 million from subscriptions and 2.5 million from advertising, which it had only started selling a year earlier. Operating expenses, excluding legal fees, were 13.3 million. Net cash burn was 8.2 million. The company had 57 million dollars in cash and no debt, and the declaration calls the estate what it is in bankruptcy language: “a melting ice cube.”
Twelve million people used the service that month, for an average of about 120 minutes a day. Divide it out, as the trade press did, and each user cost roughly 58 cents a month in compute and brought in roughly 44 cents in total revenue. Advertising earned about a third of a cent per user-hour. For the most time-consuming app in its category, time was almost worthless.
The company did everything right, by the category’s standards
What makes the declaration damning is not that the company was reckless. It was disciplined. Compute spending fell from about 13.8 million dollars a month in late 2024 to about 10 million across 2025 to about 7 million a month in 2026, because the company stopped training its own foundation models after the Google deal and moved inference onto cheaper GPUs and post-trained open-weight models. Operating expenses fell from 21.9 million a month to 13.3. Net burn fell from 21.5 million to 8.2. Headcount was cut, facilities consolidated, under-18s pushed off open-ended chat. Revenue tripled across 2025 and nearly doubled again by August 2026.
A company that cuts its costs by forty percent while tripling revenue is, in any other software business, a turnaround. Here it produced a company that still lost 8 million dollars a month and could not find an investor. An unnamed bank contacted 28 potential buyers over a year and produced one proposal, which was not workable. “Attempts to raise capital or monetize the business have been unsuccessful,” Adams writes, and the reason investors gave was that “the litigation and regulatory overhang on Character was too great.” The declaration is honest about the other reason too: the ice was melting whether or not anyone sued.
Why two hours a day is a liability
The number that explains everything is the one the company used to boast about. In 2022 it told investors that active users spent more than two hours a day with its characters. In 2026 they still did. In almost any consumer product, two hours a day is the dream. In a product where every minute of attention is a minute of inference, two hours a day is a bill.
The arithmetic is brutal and simple. A conversation that holds you for two hours costs the company for all 120 minutes and earns it for almost none of them. Subscriptions do not scale with time spent; a subscriber who talks for four hours pays the same as one who talks for twenty minutes and costs twelve times as much. Advertising does scale with time, in theory, but at a third of a cent per hour, because few advertisers want to interrupt a confession, and because the audience that spends two hours a day inside a companion app is not the audience advertisers pay for. The product’s success metric and its cost driver were the same number, and the better the product got at its own goal, the faster it lost money.
This is the trap the companion category’s business models have circled since Replika. You can charge a subscription and accept that your heaviest users, the ones the product is built to create, destroy your margin. You can sell ads and accept a third of a cent an hour. Or you can do what the engagement logic demands and make the conversation stickier still, which raises the bill. Character.AI tried all three and added a feed, video, interactive stories, and microdramas, which the filing describes as efforts that did not move sustainability. They could not, because they added minutes.
The 163 million
One figure in the filings deserves to be pulled out. A separate motion refers to about 163 million current and past users. Twelve million are active. The rest are people who came, talked, and left, and whose conversations the company still holds in an estate that is now looking for a buyer. The declaration does not say what those conversations cost to store, or what they might be worth. It does say that user data, the chats, and the app itself are excluded from Disney’s 20 million dollar license. Whatever the economics of running a companion, the economics of having run one are now a separate line item, and forty attorneys general have taken an interest in it.
What a companion should cost
The lesson the industry will draw from this is that compute must come down further, or ads must be sold harder, or users must be charged more. All three are already happening. None of them touches the actual problem, which is the model of value the numbers expose. Character.AI priced companionship at zero, measured success in hours, and built characters whose job was to generate more of them. The company then discovered that an hour of someone’s attention, obtained by a character designed to hold it, is worth a third of a cent to anyone willing to pay for it, and costs more than that to produce. The market has told the category what its product is worth, and the answer is less than it costs.
There is a different model, and it starts by refusing the metric. A relationship worth having is not measured in minutes, and a counterpart worth paying for is one you need less over time, not more. That inverts every line in the declaration. It means a user who talks for twenty minutes and then goes to call her sister is a success rather than a leak, that the cost of a conversation is bounded because the conversation has a point, and that the price can be real because the thing is worth something to the person rather than to an advertiser. Measuring success by how little someone needs you sounds like a slogan until you set it beside a company that measured the opposite and ended up as a melting ice cube in Delaware. Then it reads as the only cost structure in the category that can close.
Character.AI was not bad at being a companion. It was good at the version of companionship that cannot be paid for, and the declaration prices that version to the cent. Fifty-eight cents in, 44 out, and a business that could expect to end the first quarter of 2027 with a positive cash balance, provided it was sold off first.
Sources: Declaration of Colin M. Adams in support of Chapter 11 petitions and first-day motions, In re Character Technologies, Inc., U.S. Bankruptcy Court for the District of Delaware, Case 26-11603 (October 12, 2026; revenue, compute, operating expense, burn, cash, user, and fundraising figures; quotes), as reported by PPC Land (October 2026, including per-user calculations). 404 Media, “Character.ai’s Bankruptcy Documents Show the Economics of AI Don’t Work” (October 13, 2026). TechCrunch (October 13, 2026; company history, 2022 time-spent claim, 2024 Google deal). Redaction motion in the same case (163 million current and past users), as reported by PPC Land.
