OpenAI Shut Down Sora Because the Math Did Not Work

OpenAI announced on Tuesday that the Sora public API will be discontinued in 30 days. The reason given in the notice was the unsustainable economics of video generation at scale. The numbers that leaked to reporters over the next 48 hours put flesh on that language. Sora was reportedly burning about 15 million dollars per day in compute, against total lifetime revenue of roughly 2.1 million dollars.

Those numbers are extreme enough to be worth staring at. A product generating 2.1 million dollars in revenue cannot sustain 15 million dollars per day in cost. Not for a week. Certainly not for the months that Sora was available through the public API. OpenAI was subsidizing the product from other revenue lines the whole time.

The Sora shutdown is the clearest public example yet of a specific truth about AI products. Text generation has roughly logarithmic compute scaling per query. Image generation is more expensive but still bounded. Video generation is wildly more expensive, and the difference is not subtle. A minute of generated video can consume hundreds of times the compute of a comparable text response, depending on quality and length.

For operators, the Sora decision is a lesson in which AI workloads make sense as products and which do not, at least today. The gap between what customers are willing to pay for generated video and what it costs to generate that video is still too large to close. Every company selling AI video generation is either subsidizing the product from other revenue, accepting losses, or limiting the use case to short clips and simple scenes.

Why aren't we talking about this as a strategic story? Because it is slightly embarrassing for the field. AI pricing is supposed to drop exponentially. Most pricing has. Video is the exception. The compute-per-useful-output for video generation has improved much more slowly than for text, and pricing that works for a heavy-tail user population has not materialized.

The Sora shutdown is also a useful data point for anyone considering whether to build an AI product in a compute-heavy category. The rough rule is this. If your product's compute cost per unit of output exceeds what a typical customer will pay for that output by more than 20 percent, you do not have a product. You have a research demo with a billing system bolted on. OpenAI realized this about Sora and pulled the plug. Other labs have quietly done the same for experimental capabilities you never heard about.

There is a more practical question for operators. If you have been depending on Sora for any part of your own product, you have 30 days to migrate. Alternatives include Google's Veo, Runway's Gen-4, Luma's Dream Machine, and a handful of specialty video labs. None are cheap. Most are better-suited for specific styles of content. Evaluate carefully.

The broader lesson is about AI-product thinking. For most of the past three years, the field has trained itself to assume that if the capability exists, the product follows. Sora shows that the product does not always follow. Capability is necessary but not sufficient. The unit economics, the customer willingness to pay, and the reliability of the output all matter, and video generation today fails on at least two of those tests.

For operators building on top of AI capabilities, the move is to do the unit economics work before you commit to a product. What does each unit of output cost you? What can you charge for it? What happens to those numbers if compute prices drop 50 percent in a year? What happens if they drop only 10 percent? The answers tell you whether you have a business or a research project with a logo.

The Sora shutdown will also drive some consolidation in the video-AI market. The companies that were quietly relying on OpenAI's video capability have 30 days to make a call. Some will move to alternatives. Others will discover that their product was not actually a video product but a narrative-AI product that used video to express itself, and the move is to a different output format. The noise in the category over the next quarter will be informative to watch.