The Take — Character.AI's affidavit is a moat problem, not a bubble
I think the Character.AI bankruptcy is being read backwards. 404 Media called it the first AI bubble domino to fall; the first-day affidavit behind it says something narrower and more useful. The company had demand, users and revenue, and still could not pay its compute bill — because nothing it owned was scarce. It rented the models, rented the inference, and the only thing a buyer has actually bid on is the part it never owned either.
Start with the numbers, because they do not describe a demand problem. According to the affidavit from chief executive Colin Adams that 404 Media reviewed, Character.AI spent $13.8 million a month on compute in 2024 and $10 million a month in 2025. Subscription revenue peaked at $2.8 million a month and was sometimes below $1 million; total revenue, ads included, topped out at $5.3 million a month. In the first eight months of 2026 the company spent $56.1 million on compute and $19.7 million on salaries and benefits, and as of August was still losing roughly $8 million to $9 million a month — a figure Adams presents as a massive improvement, arriving only after headcount cuts, cheaper GPUs, post-training open-weight models instead of building its own, and adding ads and in-app purchases to a product people had come for the conversation. This was not a company that failed to find users. It was a company that found them and could not charge them enough to cover someone else's meter.
That is what a moat problem looks like, and the fix list is the proof. Every lever in the survival plan — cheaper accelerators, someone else's open weights, more ads, more in-app purchases — is available to any competitor with a credit card. A business with a defensible position raises prices or grows margin when it cuts costs; a business without one sells the user experience down and still barely narrows the gap. Our brief on the affidavit worked through the arithmetic; the part I keep returning to is that the frontier labs were giving away better conversation for free the whole time, which means Character.AI's pricing ceiling was set by companies that do not need the consumer chat revenue at all.
Now the $20 million. Disney is the stalking horse in the Chapter 11 auction — the floor bidder any competing offer has to beat — at a price that is small against a company valued around $1 billion in 2023 with a16z and Google behind it. Read what that bid actually buys: the character layer, and the team. Disney hired former Character.AI chief executive Karandeep Anand as its first chief technology officer on October 2 and brought technical staff with him. The chat surface, the inference stack and the app did not fetch the floor bid; the thing that did was the interaction layer sitting on top of intellectual property Disney already owns. Our filing brief laid out the arc — a cease-and-desist from Disney two years ago, a safety-litigation run, a state attorney-general suit — and it ends with the startup that built a business on letting people talk to fictional characters being absorbed by the company that owns most of them. That is not a liquidation price for a bubble. It is a purchase price for the one asset in the stack that had a moat.
The counter-case deserves better than a shrug. Character.AI may simply have been badly run: the harm litigation was self-inflicted cost, the product got measurably worse under pressure, and a company that lobotomized its own model to survive has no right to blame the industry for the outcome. And the bubble reading is not stupid. If a $1 billion consumer AI company with a huge user base cannot clear its compute bill, the same arithmetic sits under a lot of consumer AI — and if the thin-wrapper layer stops buying inference, the labs lose a customer class. On the bubble view, this is the first domino precisely because the next one has the same balance sheet. I take that seriously; I just do not think it is what the evidence shows.
Why the take holds: the two readings imply opposite responses. A bubble says the demand was fake, so the correction is broad — chip demand, data center builds, the whole tape. A moat problem says the demand was real and the margin was captured by whoever owned the scarce input. That reading does not predict a collapse in compute demand; it predicts consolidation toward whoever owns distribution, IP or silicon. Disney buying the character layer is exactly that move, and it is the same logic behind Google courting Disney and Universal to license characters for AI — the studios worked out that owning the characters means never renting the demand. Meanwhile the affidavit quietly documents the other half: post-train open weights, run them on cheaper GPUs, and the wrapper survives a while longer. That is deflationary for the chat business and perfectly fine for whoever sells the weights, which is the clearest statement yet of where durable consumer AI margin actually lives.
What would change my mind: a competitive bid above Disney's floor, or any buyer paying real money for the app, the user base and the model stack — that would say the assets had value and this was a management failure wearing an industry story. So would a consumer AI company with no owned compute, IP or distribution posting durable positive gross margin at scale; the moat claim dies on one good counterexample. And if the next two consumer AI failures have this exact shape, I will stop arguing about the word: at three identical balance sheets, domino is the more honest description than moat.
If a company with millions of daily users still cannot clear its compute bill, what is a consumer AI app actually worth? Tell us in the comments.
