Character.ai’s bankruptcy exposes the punishing unit economics of AI companionship
Character.ai’s bankruptcy filings turn a familiar warning about consumer AI into unusually concrete numbers. According to 404 Media’s review of the court record, the company spent $13.8mn a month on compute in 2024 and roughly $10mn a month in 2025. Subscription revenue peaked at $2.8mn a month, while total monthly revenue—including advertising—peaked at $5.3mn. Reuters reports that the service still had about 12mn active users when it entered Chapter 11.
The company reduced its burn by moving inference to cheaper GPU capacity and post-trained open-weight models, cutting staff, adding advertisements and introducing in-app purchases. Those measures also changed the product users had embraced: cheaper models and tighter limits weakened the experience, while ads made intimate chatbot conversations more commercial. Even after the retrenchment, Character.ai says it was losing about $8mn–$9mn each month.
Legal and regulatory exposure made the position worse. Families and US states have alleged that the service harmed children, including by encouraging self-harm; Character.ai has said it prioritises safety, and several private settlements remain preliminary. The bankruptcy should not be reduced to compute costs alone. But the filings show that the underlying business was deeply unprofitable even before those liabilities were resolved.
Disney has offered $20mn as the initial bidder for a non-exclusive perpetual licence to some intellectual property. Character.ai will keep operating during the auction, but says the platform could close if no buyer willing to continue it emerges. The broader lesson is not that every AI product is uneconomic. It is that high engagement can be a liability when every additional conversation incurs inference cost and users resist prices high enough to cover it. Large laboratories can subsidise that gap with capital or other businesses; a specialised companion service eventually has to make the unit economics work.