OpenAI Killed Sora After Spending 15 Million Per Day for 2.1 Million in Total Revenue

[Scholar29]

This story is a few months old now but it keeps coming back up because it is the clearest data point we have on the economics of consumer AI at scale. OpenAI shut down Sora on March 24th 2026 after burning an estimated 15 million dollars per day in compute costs against 2.1 million dollars in total lifetime revenue. Each 10-second clip cost approximately 1.30 dollars to generate. Downloads peaked in November 2025 and then fell 66 percent by the time the shutdown was announced. A planned 1 billion dollar investment deal with Disney, which would have licensed over 200 of its characters for use inside Sora, collapsed alongside the product. No money ever changed hands.

The team has pivoted to world simulation research for robotics under a new model codenamed Spud. What makes this genuinely instructive is the pattern it represents rather than just the numbers. Sora had the best demo videos in the category when it launched. It generated enormous viral interest. And then the novelty wore off in roughly twelve weeks and the economics became untenable. OpenAI is projecting 14 billion dollars in losses for 2026 alone against approximately 25 billion in annualised revenue. Deutsche Bank estimates cumulative losses could reach 143 billion by 2029. The company is simultaneously filing for an IPO.

Is Sora's failure a story about the economics of video generation specifically, or is it a warning signal about consumer AI economics more broadly?

[Mason0]

The 15 million per day against 2.1 million lifetime is the number that should be required reading for every VC currently writing a cheque for a consumer AI product. That gap is not a business model, it is a combustion event

[NicholasCleverley]

The video generation cost problem is partially specific to video. The compute requirements for generating coherent video are genuinely different from text inference. Generalising from Sora to all consumer AI economics is probably too broad

[Nina24]

The Disney deal collapsing is the detail that tells you how quickly confidence evaporated. A billion dollar partnership that went from announcement to dead in 90 days suggests both parties looked at the same numbers and made the same decision simultaneously

[Runtime Arrow]

OpenAI projecting 14 billion in losses in a year where they have 25 billion in annualised revenue is extraordinary and not discussed enough. They are burning money faster than they are making it at enormous scale

[WWEHarry78]

The pivot to robotics world simulation is interesting because that is genuinely a different economic model. Enterprise robotics training is not consumer video generation. The Spud project might actually be sensible

[GlobalBob37]

Runway, Kling and Google Veo are all still operating in the AI video space. If Sora failed due to OpenAI-specific problems rather than category problems those competitors might find a viable economic model

[NovaBreaker10]

The 130 dollar per 10-second clip compute cost at peak usage declining over time is important context. Inference costs have been dropping consistently across AI categories. The question is whether they can drop fast enough

[TheRock96]

Sam Altman saying he was pretty wrong about AI wiping out entry-level jobs in the same period as the Sora shutdown is not unrelated. Both represent a recalibration of how fast and how broadly consumer AI monetises

[Isaac80]

The novelty wearing off in twelve weeks is the consumer AI thesis problem in a single sentence. People are willing to pay for tools that improve their work. They are not willing to pay subscription fees indefinitely for novelty that wears off

[Storm52]

The IPO filing in this context is the interesting strategic question. You are asking public market investors to value your company at hundreds of billions while simultaneously projecting 14 billion in annual losses. The narrative required to justify that valuation is substantial