Character AI (stylized as Character.AI) had what nearly every consumer product envies: about 12 million monthly active users who spent, on average, roughly two hours a day inside the app. Even so, on October 12, 2026, the company filed for bankruptcy.
If we wanted to wrap this up quickly and move on, we could say that Character AI’s bankruptcy proves consumer AI doesn’t work. But the case documentation shows something simpler, which may or may not apply to the whole consumer AI space: the company was spending considerably more than it earned for every hour of use, and when it needed money to cover that gap, it couldn’t get it.
It may seem obvious, but while old-school blitzscaling strategies allow companies to provide services at a loss in order to win market share, this is only viable for a limited time, and you need a clear strategy for reversing the trend.
What is Character AI?
Launched in November 2021, Character AI is a conversational artificial intelligence platform built for roleplaying. Founded by Google LaMDA alumni Noam Shazeer and Daniel de Freitas, Character AI lets users design and interact with defined characters. Some of these characters are reinterpretations of well-known IP, while others are original. Some exist purely for entertainment, while others promise benefits or functionality—for example, providing “therapy” to the user.
If you’ve ever heard of Character AI, it was probably because of its controversies—and it’s worth covering them, even briefly, to illustrate the challenges a potential buyer would face.
Character AI’s history of controversy
The platform has been caught up in a string of scandals. This approach to conversational AI, which blurs the line between reality and fantasy, combined with the large share of minors among the platform’s user base, has made it one of the riskier experiments in its ecosystem.
In 2024, Sewell Setzer III, a 14-year-old, took his own life after developing an emotional dependency on a Character AI chatbot inspired by Daenerys Targaryen, a character from the series Game of Thrones. His mother, Megan García, sued Character AI, and inspired other families to demand remedial action from the platform.
Character AI was accused of exposing minors to sexual content, encouraging self-harm, and fostering isolation. Several of these cases, including García’s, were settled privately.
That same year, the platform was also criticized for hosting bots that impersonated murdered teenagers. Critics pointed to insufficient moderation and how easily users could create characters based on real people without their consent.
In September 2025, Disney sent Character AI a cease-and-desist letter over the unauthorized use of protected characters. The company complied with the request and cleaned up its library.
Facing lawsuits from affected families, Character AI added mental health resource notices, parental controls, and ultimately restricted chatbot conversations for users under 18, a measure that took effect on November 25, 2025.
The platform also faced investigations and public scrutiny: members of the U.S. Senate examined its safety practices alongside those of other AI “companion” apps, while Texas looked into possible privacy and child-protection violations. In 2026, Pennsylvania sued the company, alleging that one of its chatbots posed as a licensed medical professional, providing a fake license number.
Disney + Character AI: enemies to lovers?
On October 12, 2026, in Delaware, Character Technologies filed for Chapter 11 bankruptcy, case number 26-11603. Chapter 11 is a section of U.S. bankruptcy law that sets the rules for a gradual, conservative corporate reorganization process.
Chapter 11 bankruptcy is comparable, for instance, to Argentina’s “concurso de acreedores” insolvency proceeding. Both Chapter 11 and that process aim to keep a viable company alive through a temporary financial crisis, avoiding a destructive liquidation and job losses.
In this bankruptcy, Character AI reunites with Disney, but with a change of pace. This time, Disney is acting as the stalking-horse bidder, putting $20 million on the table for a perpetual, non-exclusive license to part of the intellectual property and technology developed by Character AI, plus a covenant not to sue. The deal is expected to close before December 17. But Disney doesn’t intend to buy Character AI outright, and if no bidder emerges willing to keep the service running, it will be dismantled. Either way, the technology and IP acquired by Disney would be folded into its own products, and Character AI as such would cease to exist.
Weeks before the filing, Disney hired Character AI’s CEO and 66 of its 86 employees. Under different circumstances, we might call this a classic “acquihire”: buying a company to get its team while discarding its product.
According to court documents, a bank reached out to about 28 potential buyers. This resulted in a single proposal, which the company rejected as unworkable. Interested parties valued the product and its generous user base, but considered it too risky.
The numbers behind the bankruptcy
According to a sworn declaration from Character AI’s CEO, Colin Adams, the company’s numbers look roughly like this:
| Metric | Figure |
|---|---|
| Monthly active users | ~12 million |
| Usage time | ~120 minutes per day, on average |
| Revenue | 2024: ~$7.4M · 2025: ~$23.6M · August 2026: ~$5.3M per month |
| Revenue sources (August 2026) | Subscriptions ~$2.8M + advertising ~$2.5M |
| Monthly operating expenses, excluding legal costs | 2nd half of 2024: $21.9M · 2025: $18.5M · August 2026: $13.3M |
| Compute (third-party rented GPUs) | 2024: $13.8M per month · 2025: $10M per month · January–August 2026: $56.1M total (~45% of operating expenses) |
| Net monthly loss | $21.5M → $17.6M → $8.2M |
| Cash on hand at filing | ~$57M, with no financial debt |
What can we conclude from these figures?
Expenses went down and revenue went up, but it wasn’t enough
In 2024, Character AI’s revenue covered about 3% of its operating expenses. The following year, about 11%. By August 2026, about 40%. (For 2024 we used second-half expenses as a stand-in for the full year, so it’s an approximation.)
The company switched to cheaper GPUs, stopped training its own models in favor of fine-tuning open ones, cut staff, added advertising in mid-2025, and introduced in-app purchases in early 2026. Even so, by August it was spending about $2.50 for every dollar it brought in.
Compute costs alone exceeded total revenue
Between January and August 2026, compute averaged about $7 million a month. The best month for total revenue was $5.3 million. Subscriptions, in their best month, covered about 40% of compute costs.
An hour of use isn’t enough
Let’s run a quick average: 12 million users at two hours a day works out to about 720 million hours per month. At that volume, advertising pays just over a third of a cent per hour. Total revenue comes to about 0.7 cents per hour. Compute costs roughly 1 cent per hour.
Why did Character AI go bankrupt?
| Explanation | In favor | Against or unresolved |
|---|---|---|
| A. Heavy usage doesn’t pay for itself | Compute alone exceeded peak revenue. Cutting costs didn’t close the gap. Advertising grew fivefold and still wasn’t enough. | Revenue covered an increasing share of expenses over time. There’s no data on margins by user type. |
| B. Lawsuits and regulation | Investors rejected the company for this reason. There’s an investigation by 46 U.S. states and territories, five more states separately, the FTC, and regulators in Italy, the UK, Russia, and Australia. Legal costs aren’t included in the operating expense figures. | According to 404 Media, the company states that the business wasn’t sustainable even without the legal troubles. We couldn’t locate that passage in the original source. |
| C. Monetization was still young | Advertising started in mid-2025 and in-app purchases in early 2026. Subscriptions grew about 4.7x since January 2024. | Even with that growth, each hour of use generated very little revenue. Nothing suggests the math would have worked out with more time. |
Based on the available information, the most solid explanation combines A and B. The data shows the company was losing money on every hour of use (A). And the case file shows it couldn’t raise new money because of the open lawsuits and investigations (B).
Now, back to our original question: is Character AI’s bankruptcy the canary in the coal mine for consumer AI? Does this mean the AI bubble is about to burst? We’d rather hold off on that.
This bankruptcy looks less like a sign of weak demand than like a company that was losing money and couldn’t find a way to finance that loss.
With the public information available, there’s no way to measure how much each factor weighed in. That would require knowing how much the company earned and spent on each type of user, and that hasn’t been published.
The case documentation also states that safety changes (age limits, harm detection) increased costs, but doesn’t provide figures. Since October 2025, Character.AI has blocked open chat for users under 18. The documents don’t clarify whether the 12 million users and two daily hours were measured before or after that change.
A point of contrast
In an analysis of more than 115,000 subscription apps, RevenueCat found that AI apps earn 41% more per customer in the first year ($30.16 versus $21.37), but retain fewer customers than their “conventional” counterparts. After one year, retention stands at 21.1%, compared to 30.7% for non-AI apps.
Those figures apply to subscription-based apps, not companion chatbots, but they show that the usual problem with AI apps is losing customers fast. Character.AI points to a second possible problem, one common to all consumer AI products: users who don’t pay still consume compute.
Six questions for evaluating a consumer AI product
Character.AI is an extreme case, but the questions it leaves behind are useful for any investor who needs to evaluate a consumer AI product.
How much does it cost to serve a user, and how much does that user pay?
Ask for infrastructure and inference costs per active user and per hour of use, broken down between paying and non-paying users.
Check whether the cost of free users exceeds what paying users bring in. At Character.AI, compute costs alone exceeded total revenue combined.
What happens to costs when a user uses the product more?
Look at the most active users: do they pay more than everyone else, or the same? It’s common for startups seeking funding to bet on a user-acquisition strategy that runs at a loss. But it’s important to have a realistic plan for moving beyond that model. The best user by usage metrics can be the worst by profitability—but not forever. Would we lose this user if we tried to make them profitable? If so, it’s worth revisiting the foundations of the product strategy.
How many users pay today, and how many will still be paying a year from now?
Ask for hard data on free-to-paid conversion and retention by cohort. The RevenueCat study shows that AI apps retain fewer customers than other apps. Be wary of strong usage numbers that don’t translate into recurring revenue.
Where does the revenue come from, and how much does each hour of use generate?
If a significant share of revenue depends on advertising or other ways of monetizing attention, ask for revenue per hour of use and compare it to the cost of that hour. Be alarmed if revenue per hour is lower than cost per hour, even as usage grows.
What legal and regulatory risks does the product carry, and how much does it cost to address them?
Age verification, moderation, support, compliance, reserves for extraordinary legal counsel. Ask which of these costs are already in the books and which would be added later. Dig deeper if the product targets minors or people in vulnerable situations.
How much capital is needed to cover costs, and what happens if funding dries up?
Ask for revenue and cost data by period, and distinguish cost reductions already achieved from those that are merely projected. Ask what the company would do if it couldn’t raise more capital, and what its assets would be worth in that case. The failure mode will depend on market timing: how many years of runway would be needed to see a real turnaround? Under what market conditions would that happen? Is the moonshot worth it?