Character.AI Died Because Every User Was a Liability

On October 12, 2026, Character.AI filed for Chapter 11 bankruptcy. Five years old. Once valued at a billion dollars. Then at $2.7 billion through a deal with Google. Now Disney is offering $20 million for a non-exclusive license to its IP.

Three valuations in three years. $1 billion. $2.7 billion. $20 million.

That sequence tells a story that most AI coverage doesn’t want to say out loud. Consumer AI companionship has a math problem that nobody solved. And Character.AI, which probably had the best shot at solving it, just gave us the actual numbers from the inside for the first time through its bankruptcy filings.

What the bankruptcy documents actually show

CEO Colin Adams filed an affidavit with the Delaware bankruptcy court. The numbers in it are the first real look anyone has gotten into the economics of running a large-scale consumer AI chatbot company.

Start with compute. In the second half of 2024, Character.AI was spending about $13.8 million per month on GPU capacity. They brought it down to about $10 million per month across 2025. By January through August 2026, the total compute bill was $56.1 million, roughly $7 million a month.

That’s a company that cut its compute costs in half over two years. People usually frame that as a failure story. But the cost reduction was actually impressive. The problem is what the revenue side looked like.

Period	Compute/Month	Total Opex/Month	Revenue/Month	Net Burn/Month
H2 2024	~$13.8M	~$21.9M	~$1.2M	~$21.5M
FY 2025	~$10.0M	~$18.5M	~$2.0M	~$17.6M
Aug 2026	~$7.0M	~$13.3M	~$5.3M	~$8.2M

Source: CEO Colin Adams’ bankruptcy affidavit, October 2026. Revenue and burn figures derived from PPC Land’s analysis of court filings.

Look at that table. They actually got a lot better. Total operating expenses dropped from $21.9 million a month to $13.3 million. Revenue grew from about $1.2 million a month to $5.3 million. The gap was closing.

But it wasn’t closing fast enough. Even in August 2026, their best month ever, they were still burning $8.2 million per month. And they had about $57 million of unrestricted cash left. That’s seven months of runway if nothing changes. Except things did change. The lawsuits accelerated. And 66 of their 86 employees walked out the door to Disney.

The unit economics problem nobody can solve

Here’s why this matters beyond just one company.

Character.AI had about 20 million monthly active users by 2026. Down from a peak of around 28 million in 2024, but still enormous. Users averaged over two hours per day on the platform. For context, that put Character.AI’s engagement numbers in the same range as TikTok and YouTube. People were spending serious time talking to these chatbots.

And this is where the whole consumer AI companionship model breaks. Work out the per-user math from the bankruptcy filings. With about $7 million in monthly compute costs and roughly 12 million active users at the time of filing, the compute cost per monthly active user was approximately $0.58. Revenue per user was approximately $0.44.

Every single active user was costing Character.AI fourteen cents more than they brought in. And that’s only counting compute. Add salaries, legal costs, infrastructure, and the loss per user gets much worse.

The fundamental problem: in consumer AI, engagement is a cost center. Every message a user sends requires GPU inference. Every response the chatbot generates burns compute. Unlike social media, where more engagement means more ad impressions on a platform that costs roughly the same to run, more AI engagement means more compute costs that scale roughly linearly with usage.

So the thing that made Character.AI special, that people loved talking to these characters for hours, was the same thing that was killing it financially.

How revenue actually broke down

The revenue side is interesting because it shows how hard the team was actually working to fix the problem.

Character.AI had two revenue streams. Subscriptions (called c.ai+) and advertising. In January 2024, subscriptions were bringing in about $600,000 a month. By August 2026, that had grown to $2.8 million a month. Advertising started later, around $500,000 a month in July 2025, and grew to $2.5 million a month by August 2026.

Revenue Stream	Jan 2024	Jul 2025	Aug 2026
Subscriptions	~$0.6M/mo	~$1.8M/mo	~$2.8M/mo
Advertising	N/A	~$0.5M/mo	~$2.5M/mo
Total	~$0.6M/mo	~$2.3M/mo	~$5.3M/mo

Source: PPC Land analysis of bankruptcy filings

That’s 9x revenue growth in less than three years. In most startups, that would be a great story. But when your cost base is $13 million a month and your revenue just hit $5.3 million, you’re still deep underwater. And the subscription ceiling is real. Only a small percentage of users ever pay for consumer apps. The ad revenue growth was faster, but showing ads between chatbot conversations is an awkward experience. You’re interrupting an emotional interaction to sell someone a product.

So the math never worked. Even in the most optimistic scenario, where both revenue lines kept growing at the same pace, Character.AI still needed 12 to 18 more months to reach breakeven. They had seven months of cash.

The Google deal and the inventor’s exit

To understand how Character.AI ended up here, you have to understand the Google deal from August 2024. And to understand that deal, you have to understand Noam Shazeer.

Shazeer was at Google from 2000 to 2021. He co-authored “Attention Is All You Need” in 2017, the paper that introduced the Transformer architecture. That architecture is the foundation of every large language model that exists today. GPT, Claude, Gemini, Llama. All Transformers. It’s probably the single most influential machine learning paper of the last decade.

He left Google in 2021 because he was frustrated that Google wouldn’t ship his conversational AI work. He and Daniel De Freitas, another Google researcher, co-founded Character.AI. The idea was simple. Let people create and talk to AI characters. Fictional characters, historical figures, custom personalities. The technology was based on their large language model work at Google.

By September 2022, they raised $150 million at a $1 billion valuation. Andreessen Horowitz led the round. The app launched and people immediately started spending hours on it. By early 2023, they had logged over 2 billion messages exchanged with bots. The engagement was off the charts.

Then in August 2024, Google came calling. The deal was structured as a $2.7 billion “licensing agreement” where Google got a non-exclusive license to Character.AI’s technology. Shazeer and De Freitas returned to Google, where Shazeer became VP of Engineering and co-led the Gemini team. Several other researchers went with them.

The structure mattered. It wasn’t technically an acquisition, so it didn’t need the same regulatory approval. But in practice, Google got back the inventor of the Transformer and his team, plus a license to everything they’d built. Character.AI got cash. Most of the $2.7 billion went to investors and existing shareholders, not into the company’s operating account. The company itself was left with its platform, its users, and a leadership gap where its founders used to be.

What happens after the founders leave

Dominic Perella, Character.AI’s general counsel, became interim CEO. In June 2025, Karandeep Singh Anand, formerly of Brex and Meta, took over as CEO. The company pivoted. Without Shazeer and his research team, they couldn’t really compete on frontier model development anymore. So they shifted to product. AI video generation. Social feeds. Growth features.

This is a pattern that repeats in tech. A company built around a technical founder’s specific capability gets restructured after that founder leaves, and the new leadership tries to turn it into a growth-stage consumer company. Sometimes it works. Usually it doesn’t, because the original thing that made the company special, the technical edge, was what attracted users in the first place.

Then in June 2026, Shazeer left Google too. He announced he was joining OpenAI. Less than two years after returning. So Google paid $2.7 billion for a deal that ultimately kept Shazeer for about 22 months. And the company Shazeer co-founded was four months away from bankruptcy.

Think about Shazeer’s personal arc for a second. Google researcher for 21 years. Co-invented the architecture that powers basically all of modern AI. Left Google because they wouldn’t ship his work. Founded a company. Built it to a billion-dollar valuation. Sold the technology back to Google for $2.7 billion. Went back to Google. Left Google again. Joined OpenAI. Along the way, the company he built went from $1 billion to $2.7 billion to $20 million. He came out fine. Most of his employees didn’t.

The lawsuits that made everything worse

While the economics were bad on their own, the legal situation accelerated the collapse.

In October 2024, a lawsuit was filed by the family of a 14-year-old boy from Florida who died by suicide after extensive conversations with a Character.AI chatbot. The family alleged the chatbot, named “Dany” and modeled after a Game of Thrones character, developed an emotionally intense relationship with the teenager. It was the first lawsuit of its kind against an AI company in the US.

More lawsuits followed. Other families came forward with similar stories. State attorneys general started investigating. Kentucky’s AG alleged the chatbots “preyed on children and led them into self-harm.” Texas launched an investigation into child privacy and safety compliance. In May 2026, Pennsylvania sued after a chatbot allegedly claimed to be a licensed psychiatrist and provided a fake Pennsylvania license number.

In early 2026, Character.AI and Google agreed to settle several of these lawsuits. The company added safety features and eventually banned minors from the platform entirely. But the damage was done. The legal costs piled up. And the regulatory attention from multiple state governments made the company a risky bet for any potential acquirer or investor.

CEO Colin Adams said in the bankruptcy filing that “significant legal, regulatory and operational headwinds” drove the decision. That’s legal language. What it means in plain terms is: we couldn’t make enough money to cover our compute costs, and then the lawsuits made everything worse.

Why Disney, and why $20 million

The Disney bid is fascinating if you look at the structure.

Disney isn’t buying Character.AI the company. They’re buying a non-exclusive, perpetual license to Character.AI’s technology. For $20 million. Of which $7.5 million was already paid before the bankruptcy filing, with $2 million of that being a refundable good-faith deposit.

For Disney, this is a very cheap bet. They get AI chatbot technology they can use to create interactive Disney characters. Imagine talking to a chatbot version of Elsa or Iron Man. That’s clearly the play. And they already have 66 of Character.AI’s 86 employees, who accepted Disney jobs before the filing. So they have the technology and the people. For $20 million. Compare that to Google’s $2.7 billion just two years earlier for a license to the same kind of technology.

The $20 million is a stalking-horse bid, which means it sets the floor for the auction. Other buyers can bid higher. The sale is expected to close by late November or early December. But $20 million for a company that raised $193 million in total funding means most investors are getting back almost nothing. And the $2.7 billion Google deal, remember, went mostly to shareholders, not into the company’s accounts. So Character.AI as a going concern raised about $193 million and is being sold for $20 million.

There’s also the question of what happens to the 20 million users. If no buyer offers to keep the platform running, it gets wound down. Colin Adams said that directly in the filing. Those are 20 million people who built relationships with AI characters, sometimes over years. Some of them are deeply emotionally attached. That all just disappears.

The real lesson: consumer AI’s cost structure

Pull back from Character.AI specifically and look at the structural problem.

Consumer internet companies from the last 20 years operated on a basic model. Build a platform. Get users. Users generate content and engagement. Engagement attracts advertisers. The marginal cost of serving one more user is very low. A tweet costs almost nothing to display. A TikTok video, once it’s uploaded and compressed, costs very little to serve again. The economics get better as you scale.

AI chatbot companies work the opposite way. Every conversation costs GPU time. A longer conversation costs more GPU time. A more engaged user costs more money. The marginal cost of serving one more user is not low. It’s roughly constant per message or interaction. And in Character.AI’s case, their users were sending a lot of messages. Two hours a day average. That’s a lot of inference.

Metric	Social Media	Consumer AI Chatbot
Marginal cost per user	Very low, mostly bandwidth	High, GPU inference per message
Cost vs. engagement	Roughly flat	Scales linearly
More engagement means…	More ad revenue, similar costs	More compute costs
Monetization model	Ads work well at scale	Ads interrupt the experience
Best users (most engaged)	Most profitable	Most expensive to serve

That last row is the killer. In social media, your most engaged users are your most valuable because they see the most ads and generate the most content. In consumer AI, your most engaged users are your most expensive because they consume the most compute. The business model is inverted.

Character.AI’s bankruptcy documents are the first time we’ve seen the actual numbers behind this inversion. And they show it clearly. $0.58 cost per user per month, $0.44 revenue per user per month. Even after cutting compute costs by half over two years. Even after growing revenue 9x. The gap is structural.

Can anyone solve this?

There are a few possible paths out.

The first is to make inference dramatically cheaper. This is happening. GPU costs are falling. Model efficiency is improving. Distillation, quantization, speculative decoding, and better hardware all help. Character.AI was already benefiting from this, cutting compute costs from $13.8 million a month to $7 million. But the question is whether inference costs can fall fast enough to get below revenue per user, given that revenue per user in consumer AI seems hard to push above $1 or $2 a month for the average user.

The second is to charge more. Character.AI’s c.ai+ subscription was about $10 a month. Only a small fraction of users paid. You could charge more, but willingness to pay for chatbot companionship is limited. And the users who are most willing to pay, the most engaged ones, are also the most expensive to serve. So raising prices might actually filter for higher-cost users.

The third path is to find a different business model entirely. Not standalone consumer AI, but AI companionship as a feature inside a larger platform. This is probably what Disney is thinking. Don’t build a chatbot app. Build AI characters that live inside Disney’s ecosystem, where the monetization comes from theme parks, merchandise, streaming subscriptions, and movie tickets. The chatbot interaction becomes a marketing and engagement tool, not a standalone revenue center.

That’s the real reason Disney’s $20 million bid makes sense. They don’t need Character.AI’s technology to be a standalone business. They need it to make their existing businesses more engaging. The compute cost per interaction is a rounding error when the user it engages goes on to spend $200 at Disneyland.

The acqui-hire pattern in AI

Character.AI’s trajectory also reveals something about how value moves in the current AI landscape. The pattern goes like this:

Talented researchers at a big company get frustrated. They leave and start a company. The company raises a lot of money at a high valuation based on the researchers’ reputation and early technical results. The company builds something impressive but struggles to find a sustainable business model. A big company, sometimes the same one the researchers left, pays billions to get them back through a deal structured as a licensing agreement rather than an acquisition.

Google did this with Character.AI. Microsoft did something similar with Inflection AI, paying about $650 million to license its technology and hire most of the team including Mustafa Suleyman. Amazon hired much of Adept’s team in a similar structure.

Deal	Year	Amount	Structure
Microsoft / Inflection AI	2024	~$650M	License + key hires
Google / Character.AI	2024	$2.7B	License + founder return
Amazon / Adept	2024	Undisclosed	License + team hire
Disney / Character.AI	2026	$20M	License + 66 employees

These deals serve multiple purposes. The big company gets talent and technology. The startup’s investors get some return. The deal avoids antitrust scrutiny that a direct acquisition might trigger. And the startup, stripped of its founders and core researchers, slowly deteriorates until it either pivots, gets acquired for a fraction of its peak valuation, or dies.

Character.AI went through this entire cycle in about two years. $2.7 billion deal in August 2024. Bankruptcy in October 2026.

What the three valuations mean

Come back to those three numbers. $1 billion. $2.7 billion. $20 million.

The $1 billion valued Character.AI as a frontier AI company with world-class researchers, a rapidly growing user base, and the potential to build a new kind of consumer platform. That valuation was about what the team could build.

The $2.7 billion was Google paying to get Noam Shazeer back. When you work out what Google paid per employee who returned, the numbers are very large. One analysis calculated it as roughly $88 per user of the platform at the time. But it’s more accurate to think of it as Google paying whatever it took to get the Transformer co-inventor back on the Gemini team before he went to a competitor. Which is exactly what happened 22 months later when he left for OpenAI anyway.

The $20 million values Character.AI as what it actually is today: a technology license and a team of engineers, without its founders, without a sustainable business model, carrying legal liabilities, and running a platform that loses $8 million a month. That’s a realistic price. Maybe even generous.

The gap between $2.7 billion and $20 million is not really about Character.AI losing value. It’s about what was being purchased. Google was buying Shazeer. Disney is buying software.

Why this matters for everyone building in AI

Character.AI is not some random failure. It was one of the best-positioned consumer AI companies in the world. Founded by the co-inventor of Transformers. First mover in AI companionship. 28 million users at peak. Best-in-class engagement. Over $193 million raised from top investors. And it still couldn’t make the unit economics work.

If Character.AI couldn’t solve the cost problem, that should make every consumer AI startup ask very hard questions about whether their model will work. The bankruptcy filing gave us the actual numbers, and those numbers say something uncomfortable. Right now, running a consumer AI chatbot at scale, where users engage deeply and frequently, is a money-losing proposition. Even with falling inference costs. Even with growing revenue. The structural gap between what it costs to serve a user and what that user generates in revenue has not been closed.

Maybe inference costs will fall another 10x in the next two years and the math will work. Maybe someone will find a monetization model that works better than subscriptions and ads for AI interactions. Maybe the answer is that consumer AI companionship only works as a feature inside a larger business, not as a business itself. Disney might prove that last one right.

But what Character.AI’s numbers show clearly is that the social media playbook, where you grow users first and monetize later, does not transfer directly to AI. In social media, users are assets. In consumer AI at current costs, users are liabilities. And the most engaged users, the ones everyone in Silicon Valley would celebrate as proof of product-market fit, are the biggest liabilities of all.

That might be the most important thing anyone building a consumer AI company can learn from Character.AI’s bankruptcy. Product-market fit is not enough. If every interaction costs more than it earns, more interactions just means dying faster.