OpenAI thought it could own AI videos. The reality was too expensive - CNN
Sora was supposed to change everything. Eleven months later, it has changed exactly one thing: the math on OpenAI's quarterly burn rate.
OpenAI's AI video generator, which the company announced to extraordinary fanfare in February 2024 before releasing a consumer app in October 2025, has been quietly shut down. The financial record of its brief commercial life is now available, and it is not close. Sora grossed approximately $2.14 million in total revenue across 11.7 million downloads, according to Appfigures Intelligence data reported by Ars Technica. Its peak of 3.3 million downloads in November 2025 had collapsed to 1.1 million by February 2026 — a 67 percent drop in three months. Daily active users fell 34 percent over the same period. Against this, a single quarter of ChatGPT subscriptions generated $1.9 billion in net in-app revenues, according to Sensor Tower data cited by the BBC.
The comparison is not kind to Sora. It is not even in the same category.
"Sora is a resource black hole with limited monetization," Forrester analyst Thomas Husson told the BBC. That framing — resource black hole — is the one that survived contact with the numbers.
The Disney deal is the detail that crystallizes the whole thing. Reuters reported that Disney and OpenAI had discussed a $1 billion investment and content partnership. The deal never closed. No money changed hands. Disney walked away before a single dollar transferred. The $1 billion figure represents a hypothetical future that OpenAI's own leadership apparently concluded was not coming — Bill Peebles, who led the Sora project, posted publicly that the economics were currently completely unsustainable. He is correct.
The gap between Sora's demo and its economics is not a mystery. Video generation is genuinely hard in a specific way that text generation is not: every output requires a full model inference run, the outputs are large files that require storage and delivery infrastructure, and the use cases that excite people most — replacing B-roll footage, generating b-roll for commercials, producing personalized content — are exactly the use cases that require commercial licenses, rights clearance, and consistency across frames that current models cannot reliably deliver. The demo looked like magic. The product looked like a toy. Paying $120 a year for a toy that occasionally produces something you can use is a different value proposition than paying $20 a month for a chatbot that writes your emails.
The analyst consensus on why OpenAI pulled the plug is straightforward: OpenAI is still unprofitable, investor and competitive pressure is intensifying, and $15 million a day for a product generating $1.4 million in lifetime net revenue is not a line item that survives internal scrutiny. "This is cash they likely decided they can't afford to continue burning as initial interest has faded," Henry Ajder, an AI industry analyst, told the BBC. The tense is important. They decided. This was not a science experiment that failed. It was a business decision that was made.
What is the lesson? That AI video is dead? That generative media is a dead end? That would be the wrong lesson. Stable Diffusion shipped a video product. Runway raised at a substantial valuation. Google continues to invest in Veo. The technology is real. The question was never whether the outputs were impressive. The question was whether the unit economics of impressing people at scale would ever converge with the cost of generating those outputs. Sora answered that question empirically: not yet, and not soon.
The question that actually matters for the industry is whether video generation follows the same cost curve as text generation. Text inference got cheap because hundreds of millions of people used it and the fixed costs of training amortized across a large revenue base. Video inference is still in the early phase where the people building it are paying the full cost of the research, the hardware, and the delivery infrastructure without a corresponding revenue base to absorb it. The same was true of ChatGPT in 2020. The difference is that ChatGPT found a product-market fit at $20 a month for a text box that scales cheaply. AI video has not yet found its equivalent.
OpenAI will try again. The research investment is not wasted. But the lesson of Sora is not subtle: a beautiful demo is not a business. The people who passed on a $1 billion Disney deal without changing hands understood something the viewers of the demo did not. The demo was not the product. The demo was the hypothesis. The hypothesis was wrong. Now they know.