Character.AI filed for Chapter 11 bankruptcy on Tuesday, October 13, the day after Columbus Day. The filing matters less than the sequence that led to it. Two deals pulled out the people and intellectual property that strategic buyers valued. What remained was 12 million users, six lawsuits from families alleging harm to minors, investigations from 46 states, an FTC inquiry, and a compute bill that revenue never covered. I'm calling that a remainder company: the entity left holding everything no buyer would take.
The sworn declaration that accompanies the filing also contains the most transparent consumer AI operating data ever made public. Treat those numbers as benchmarks. They apply to any consumer AI company claiming growth or engagement, including the ones you may be talking to right now.
The timeline, weighted by what changed trajectory
2021–2022: Founding and early traction. Noam Shazeer and Daniel De Freitas, both former Google researchers, incorporated Character Technologies in November 2021 and opened a public beta in September 2022. Users could create AI characters and hold open-ended conversations with them. By December 2022 they were generating over a billion words a day across 350,000 characters. The company built its own models and its own consumer app, with no dependency on third-party APIs.
March 2023: The $150M raise on zero revenue. Andreessen Horowitz led a $150M round at a $1B valuation, on top of roughly $40 million previously raised from individual investors including Nat Friedman, Elad Gil, SV Angel, and A Capital. a16z's stated thesis rested on proprietary models, full-stack control, and a data flywheel in which user interactions improved the AI.
May 2023–early 2024: Growth without a business model. The mobile app launched and hit 1.7M downloads in its first week. Character.AI added c.ai+, a paid tier offering faster responses and priority access. Engagement kept climbing, yet the bankruptcy declaration shows subscription revenue was still only about $600,000 a month as late as January 2024. In February 2024, a 14-year-old named Sewell Setzer III died by suicide after what his family would later allege was an emotionally dependent and sexualized relationship with a Character.AI chatbot. His mother filed suit in October 2024. By the time she did, the Google deal had already reshaped the company.
August 2024: Google's license-and-hire, the first extraction. Google did not acquire Character.AI. It structured a deal that delivered most of what an acquisition would have while leaving the liabilities in place. Google hired both co-founders and key researchers, took a non-exclusive license to Character.AI's model technology, paid $2.7B in cash plus canceled Google-held convertible instruments, and bought out venture investors at roughly $2.5 billion.
Alphabet's own books treat it more like an acquisition than the announcement did. The company applied acquisition accounting (ASC 805) and recorded $2.7 billion in goodwill plus $413 million in identifiable intangible assets, approximately $3.1 billion in recognized value. The goodwill figure happens to match the cash payment, but the two are different things. Goodwill is the residual left after identifiable assets are valued, so Alphabet first assigned $413 million to specific intangibles (technology rights, contractual rights, or similar categories) and booked everything above that as goodwill. Its annual report never disclosed what the $413 million consisted of or how it would be amortized. The amount was material enough to require recognition. Its composition is still opaque, and I can't fill that in from public filings.
What Google took: the founders, the research team, the technology rights, and the investor obligations. What stayed behind: the app, the users, every employee Google didn't hire, and every problem the company had accumulated or would go on to accumulate.
October 2024–September 2026: The obligations compounded. Families of minors filed six lawsuits. Two families sued jointly in December 2024, and three more cases followed in September 2025. Each involved a minor, and each alleged some combination of emotional dependency, sexualized interactions, self-harm, and suicide or attempted suicide. Preliminary settlements for five cases were discussed in January 2026 but never finalized, and they remain unpaid. A sixth case was filed in August 2026.
Over the same period came investigations by 46 U.S. states and territories, a civil investigative demand from the FTC, state lawsuits in Kentucky and Pennsylvania (the latter alleging bots practiced medicine without a license), a fine in Italy, and a UK data-privacy investigation. The bankruptcy declaration lists all of these. It never estimates total potential liability.
The company added advertising in mid-2025 and in-app purchases in early 2026, which took total monthly revenue from about $600,000 in January 2024 to $5.3 million by August 2026. Monthly compute alone ran over $7 million.
September 2026: Disney's license-and-hire, the second extraction. Starting in 2025, Character.AI's advisers approached 28 potential buyers or investors. They liked the user base and walked away from the litigation exposure. Disney then structured an arrangement on Google's template: a perpetual, worldwide, non-exclusive license to specified intellectual property and technology, plus hiring the remaining technical leadership, including the CTO and members of the technical team. The $20 million stalking-horse bid filed with the bankruptcy covers the IP license. The personnel moves are separate.
The template ran twice. Google took the founders and the model; Disney gets the technical leadership that remained, along with IP rights.
October 2026: Bankruptcy. Character.AI entered Chapter 11 with the service still running and no confirmed bidder proposing to continue operating it. If none emerges, the app shuts down. Google had already bought out every institutional investor, so all remaining equity belongs to current and former employees. What they own is a bankruptcy estate.
The unit economics, as diagnostic benchmarks
These figures come from the sworn bankruptcy declaration, filed under penalty of perjury. They are the first public look, under oath, at what it costs to run a high-engagement consumer AI companion at scale. Pull them out whenever a consumer AI company points to strong engagement or a path to profitability.
August 2026 operating snapshot:
- 12 million monthly active users
- ~120 minutes average daily engagement per user
- $5.3 million monthly revenue ($2.8M subscriptions, $2.5M advertising)
- $7 million+ monthly compute
- $8–9 million monthly total operating loss
| Metric | Per user, per month |
|---|---|
| Revenue | $0.44 |
| Compute cost | ~$0.58 |
| Revenue as % of compute | ~76% |
| Total operating cost (excl. litigation) | ~$1.11 |
| Revenue as % of total operating cost | ~40% |
| Net loss | $0.67–$0.75 |
Compute alone exceeded revenue by roughly $1.7 million a month, before any salary, legal bill, or office lease. The company lost money on every user interaction. Across roughly 720 million user-hours a month, each hour of engagement brought in about $0.0074 and cost about $0.0097 in compute, so every additional hour widened the loss.
When a consumer AI company cites daily engagement as proof of product-market fit, run this ratio. Character.AI's 120 minutes per user per day is what made its economics unsustainable. It was never evidence that they worked.
The filing doesn't disclose subscriber count, paid conversion, churn, or advertising yield, so the per-cohort picture is likely worse than these blended averages suggest. The blended numbers are diagnostic enough on their own.
What two extractions left behind
Two license-and-hire transactions, twenty-six months apart, left a company composed of what no buyer wanted.
Google's August 2024 deal took the founders, the research talent, and the model technology. Disney's September 2026 arrangement takes the remaining technical leadership and IP rights. In between, ownership passed entirely to current and former employees, revenue grew without ever covering compute, and the legal and regulatory burden went from one lawsuit to six, plus 46-state investigations, federal inquiries, and international enforcement.
The operating product — the thing 12 million users used for two hours a day — is the part no buyer has offered to keep running. (I covered the Manus forced unwind in issue 15, where a completed acquisition had to be reversed at the product layer. Character.AI runs the other direction: the buyers never acquired the operating product, so there's nothing to reverse. They licensed and hired what they valued, and the obligations stayed where they were.)
Where this structure goes next
High confidence that license-and-hire becomes the default structure for large companies absorbing consumer AI teams in the near term.
A consumer AI company carries two kinds of assets. Technical assets — people, models, IP — transfer cleanly through licensing and hiring agreements. User-facing assets — the installed base, its data, its behavioral expectations, and the safety and regulatory obligations that come with serving it — arrive bundled with litigation risk, content-moderation duties, and compute costs that scale with engagement. A license-and-hire lets the buyer take the first kind and leave the second.
Google built the template and Disney reused it. Expect it again, because it works for the buyer.
Moderate confidence that it persists beyond the near term. Two conditions could break it:
-
Consumer AI unit economics improve enough that owning the user relationship is worth the liability. Character.AI's numbers put that a long way off. At $0.44 in revenue per user per month against $0.58 in compute alone, the gap is structural. Inference costs would have to fall dramatically, or monetization would have to reach well beyond subscriptions and ads.
-
Regulators start treating license-and-hire deals as de facto acquisitions. Google recording $3.1 billion in acquisition-accounting assets from a "non-exclusive license" is the kind of discrepancy that eventually draws scrutiny, and the evidence above already makes this condition plausible. Alphabet's books say business combination; the deal terms say license. Both can't be precisely true, and at some point a regulator or court will ask which one governs.
What this means when you evaluate a consumer AI company
Whether you're looking at a consumer AI company as a potential employer, a competitor, or a market signal, ask one question first: can the technical team and model IP be separated from the user base and its obligations?
If yes, the most likely outcome is disaggregation. Talent and technology go to a platform company through a license-and-hire. Users, safety obligations, regulatory exposure, and the compute bill stay in an entity that can't carry them.
If you're weighing a role at a company with this profile, find out which part you'd be joining — the part that gets extracted or the part that remains. This week, Character.AI's 12 million users were still on the product two hours a day, and nobody has proposed to keep it running. The people who built it are at Google. The people who will maintain its remaining technology are going to Disney. Keep tracking which consumer AI companies can make a user-hour profitable; the ones that can't are where the next remainder company comes from.
- Apple's Huxe arrangement: Apple disclosed to European regulators a license-and-hiring structure with personalized-audio startup Huxe that follows the same template — hire the team, license the IP, let the consumer product shut down and delete user data.
- Manus's post-separation revenue: The Information reported Manus's annualized revenue run rate jumped from ~$100M to $400M–$500M during the Meta ownership period, but those figures are unaudited — a useful contrast to Character.AI's sworn bankruptcy numbers.
- Bankruptcy data restrictions: Courts can appoint a consumer-privacy ombudsman and impose conditions on transferring user data in a §363 sale, which means Character.AI's 12 million user accounts may not be freely transferable even if a bidder emerges.
- Competing stalking-horse bids: No public competing bid had appeared on the bankruptcy docket by October 16 — worth monitoring weekly, since the absence of a platform bidder is what converts Character.AI from a distressed company into a remainder company.
