"Four years of burning cash, and the valuation was wiped out."

Author | Jiachang

Produced by | Jixin

On October 12, 2026, AI companion star Character.AI filed for Chapter 11 protection in a Delaware court and entered a court-supervised asset auction. Disney served as the stalking horse bidder at $20 million, acquiring a non-exclusive perpetual license to part of the intellectual property along with transitional technical services; the platform app and user assets are not part of the deal. In its bankruptcy filing, the company self-reported assets and liabilities both in the $50 million to $100 million range—the two sides nearly offset, meaning its net assets have approached zero. This unicorn, valued at $1 billion in 2023 and rumored at its peak to be worth $5 billion, has a ledger that is plain to see: compute burns over $10 million a month, subscription revenue peaks at only $2.8 million, and the longer users chat, the more the company loses. Costs, peers, founders, and the users left behind—all four threads are tied to this ledger.

1 From the $5 billion rumor to the $20 million floor price

Character.AI began on the eve of the large-model explosion. In 2021, Noam Shazeer, a core author of the foundational Transformer paper "Attention Is All You Need," and his colleague Daniel De Freitas left Google, and the product launched the following year. Users could build their own characters and chat with historical figures, anime characters, and completely fictional personas; the platform grew through its UGC character library, and user dwell time was high enough to make contemporary AI products envious.

In March 2023, a16z led a $150 million Series A when the company had not a cent of revenue and was already valued at $1 billion; by September, market rumors said it was raising at a valuation of over $5 billion, but that round never closed.

The high valuation did not hold. In the October 2026 filing, assets and liabilities were each between $50 million and $100 million. Disney's Disney Worldwide Services served as the stalking horse bidder, offering $20 million for a non-exclusive perpetual license to part of the intellectual property and related claims, and also using newly hired former employees to provide transitional technical services to the platform. CEO Colin Adams said the board evaluated all options and that a court-supervised sale was the best path to "maximize value for all stakeholders." The auction window is expected to close in late November to early December; if no new buyer takes over, the platform will shut down as planned, and the characters and chat histories created by tens of millions of users will become inaccessible data.

2 The three fatal numbers in the ledger

The bankruptcy filing lays the costs on the table. CEO Colin Adams wrote a blunt line in his affidavit: "Character's technology is impressive, but it is too expensive to operate." How expensive, exactly? The filing contains three numbers.

Compute. In 2024 it burned an average of $13.8 million a month, falling to $10 million in 2025, and totaling $56.1 million in the first eight months of 2026, about 45% of operating expenses. On the other side, subscription revenue peaked at $2.8 million a month, sometimes less than $1 million, and total revenue including advertising peaked at $5.3 million a month. As of August 2026, it was losing $8 million to $9 million a month, which the company called a level already "substantially improved." Full-year 2025 revenue was not enough to pay three months of compute bills.

The three numbers point to the same deadlock: the more users, the higher the cost. Monthly active users have long been around 20 million; spread across each user, annual revenue is about $2.5, while each user consumes expensive inference compute every day. Every extra hour a user chats, the GPU burns another hour. In a business where the more popular a user is the more it loses, scale only accelerates the bleeding, and there is no moat to speak of.

3 Giants pick the peaches, the soul is pulled out

The company had its peaches picked by giants twice. In August 2024, Google paid about $2.7 billion for a technology license and brought Shazeer, De Freitas, and about 30 researchers back to DeepMind, with Shazeer becoming co-technical lead of Gemini. Investors were bought out at about $88 per share and a company valuation of about $2.5 billion, with remaining equity left to stay-on employees. The company kept its brand and product but lost its founders and core in-house team—Silicon Valley calls this a "reverse acquihire."

After losing its soul, Character.AI effectively became a product company. It abandoned in-house large models and switched to fine-tuning open-source models such as Meta's, focusing its efforts on product features like character consistency and memory. CEO Karandeep Anand, who took over in June 2025, redefined the company as an "AI entertainment company," aiming to build immersive content across Chat, Books, Series, and Comics.

But the story did not continue. In June 2026, Shazeer left Google to join OpenAI; in September, Anand moved to Disney as CTO, with a group of technical staff following. The vacant CEO seat was taken by Colin Adams—a veteran executive skilled in operational restructuring; less than six weeks after taking office, he signed the Chapter 11 filing on October 12.

A month earlier it was still poaching people for Disney; a month later Disney was sitting in the stalking horse seat at the bankruptcy auction. The Verge put it sharply: this was a Silicon Valley-style "acqui-hire," with the giant taking the talent and assets and leaving the high-cost operations, legal risks, and community responsibilities in place.

4 Why cost-cutting could not save the company

Switching from in-house models to fine-tuned open-source models was originally a correct cost-cutting move. Average monthly compute spending fell from $13.8 million in 2024 to about $7 million in the first eight months of 2026, and Character.AI's affidavit calling the "$8 million to $9 million monthly loss" a "substantial improvement" was not mere courtesy.

But the money saved was quickly swallowed by deteriorating revenue. As models became lighter, character consistency in long conversations and cross-turn memory became more fragile; and Character.AI's core experience was precisely the long-term relationship between a user and a persona that "remembers you."

Monthly active users fell from a peak of about 28 million in mid-2024 to around 20 million, a drop of nearly 30%. Remaining users still chat for hours every day, and the GPUs keep burning; departing users took away already thin subscription and advertising inventory. Starting in 2025, the company also inserted full-screen mobile ads and web banners for free users.

Retention was already declining, and advertising pushed marginal users further away.

This is a classic death spiral: cost-cutting via smaller models leads to a degraded experience; monthly active users churn and the revenue base shrinks; fixed compute costs are spread thinner, forcing more ads or budget cuts; the experience degrades further and users leave again. The cost curve went down, but the revenue curve fell faster. By August 2026, the company was still losing $8 million to $9 million net every month; the endpoint after "improvement" was still bleeding.

The problem was never "not saving enough." Every extra hour a user chats, the GPU burns another hour, and that hour brings in almost not a single cent of extra revenue. As long as this unit economic model does not hold, any cost-cutting only delays the bleeding rather than stopping it.

5 Peers' ledgers: who can make the math work

Character.AI is not the only AI companionship company burning cash; the difference is only in what each relies on to survive until now. Replika has done emotional companionship for nearly a decade, selling paid features such as voice and selfies through subscriptions; it monetized early but never scaled large, becoming a small and beautiful business. ByteDance's Xingye and Maoxiang in China take the free-plus-ads-plus-in-app-purchases route, backed by the traffic and compute of a major company, essentially subsidizing an AI story with profits from the main business.

Character.AI's predicament did not accumulate overnight. An independent startup wants both scale and profit, but paid conversion cannot support inference costs; the middle ground in this position is empty. Either be internalized by a major company's costs, or be small enough not to attract attention—a standalone "companion chat" business cannot support a unicorn.

Copyright and regulation added another cut. In September 2025, Disney sent a lawyer's letter accusing the platform of systematically reproducing Disney character IP and of some content posing risks to minors; within a week the platform removed its most popular characters including Mickey Mouse, Donald Duck, Captain America, and Luke Skywalker. In 2024, the suicide of 14-year-old Sewell Setzer III in Florida triggered the first U.S. lawsuit over "an AI chatbot causing the death of a minor"; in January 2026, the Garcia lawsuit and four other family lawsuits were settled, and the company also banned minors entirely.

Public opinion split into two camps on Hacker News. One camp cursed burn-cash culture, repeatedly citing the $100,000 DoorDash expense bill and the $278,000 average base salary in the bankruptcy filing; the other camp cursed the business model, saying that ranking first in user time while being unable to monetize is a structural fatal flaw of consumer AI, and that it would be the same no matter who did it.

But Hacker News represents only the Silicon Valley engineer community. The people truly left behind have almost no channel to speak: the characters created by tens of millions of users and years of chat histories may disappear along with the app after the December auction is settled, and paying users' refund priority ranks far down the creditor list. The claim that "the founders cashed out twice" does not hold legally; both the Google and Disney deals were normal commercial contracts with no fraud allegations, but the disappointment of ordinary employees and silent users is real.

The capital narrative had long since shifted. In 2022 and 2023 investors were willing to pay for "user stories"; after 2025 they want verifiable cash flow and controllable marginal costs. A product that relies only on companion chat with limited paid conversion cannot raise a new round. The lesson it leaves is also simple: when the cost of an AI product rises linearly with user time while willingness to pay does not rise with it, scale is not a moat—it is the speed of bleeding.