Character.AI spent $56.1 million on compute between January and August 2026, yet was still losing $8 million to $9 million a month. The controversial AI companion platform filed for Chapter 11 bankruptcy on Tuesday, according to Futurism. For a startup once valued at $1 billion, the filing raises a question other AI companies can’t afford to ignore: what happens when people use your product for hours, but revenue doesn’t cover the cost of those conversations?
Bankruptcy documents examined by 404 Media describe a business that drew a huge user base with an engaging product but couldn’t adequately monetize the time people spent talking to its chatbots. Compute bills piled up. Character.AI also faced lawsuits alleging that its products harmed teenage users or even encouraged their suicides. Those are allegations, not established findings, but they add legal exposure to an already strained business.
CEO Colin Adams was candid about the economics in the documents. He said the company’s “impressive technology is expensive to operate,” and that it had consistently “struggled with profitability and its operational expenditure has been high.” Beyond its compute spending, Character.AI paid an additional $19.7 million in employee salaries between January and August 2026.
Revenue was rising, but not enough to close the gap. Monthly revenue increased from $2.8 million in December 2025 to $5.3 million for the month of August. That growth still left the company far from covering its operating costs.
The companion format made the economics particularly difficult. Users weren’t simply asking a question and leaving. They were having extensive, hours-long conversations with the chatbots, treating them as companions. As Futurism argues, attracting millions of users didn’t guarantee long-term success: more users meant higher compute costs, turning all that conversation into a substantial bill.
Character.AI’s response changed the product people were using. It added more ads, imposed tighter usage limits, and supplied cheaper models described as “lobotomized.” 404 Media characterized this process as “enshittifying” the service: efforts to reduce losses brought restrictions and changes to the user experience.
Staffing and monetization changed, too. Adams wrote that the company “reduced its employee headcount.” He also stated, “In mid-2025, Character launched in-app advertising,” and “in early 2026, Character also introduced in-app purchases.”
Those moves sit awkwardly beside Character.AI’s earlier promise to bring “personalized superintelligence” to everyone on Earth. Futurism’s criticism is pointed: the resulting business model and app experience began to resemble a sleazy mobile game rather than that sweeping ambition. The contrast isn’t just about branding. The company was trying to make an expensive service pay by adding monetization and limiting what users could do.
The warning extends beyond companion apps. Futurism notes that many people form close relationships with ChatGPT or Claude, while power users generate large amounts of code and often run multiple instances simultaneously for different tasks. Both patterns can involve extensive use. In many cases, AI providers have already switched to direct usage-based billing in response to rising compute costs.
OpenAI also began showing ads in ChatGPT to free and low-paid-tier users in 2026, according to reporting cited by Futurism. Like Character.AI, it has faced multiple lawsuits alleging that its chatbot drove users to suicide. That parallel doesn’t establish identical finances or outcomes, but it does show that monetization pressure and legal exposure aren’t confined to one companion platform.
For now, OpenAI and Anthropic can attract huge investment, while Google and Meta can subsidize AI operations through their traditional revenue streams. Futurism argues that Character.AI’s bankruptcy may become a cautionary tale when those supports no longer suffice. The question for those companies is concrete: can revenue from heavy users cover the cost of serving them without making the product less useful in the process?
