Some easy calculations suggest that OpenAI spends greater than 1 / 4 of its revenue running its AI slop factory.
For an organization that burns greater than twice its income, OpenAI is quickly introducing some ingenious and ruthless recent ways to build up losses. The AI giant is value around $500 billion as of Thursday projected an annual recurring revenue rate of $20 billion. That’s all well and good until you remember it lost greater than $12 billion last quarter.
On September 30, OpenAI introduced its Sora video creation app for Apple’s iOS platform, which saw a formidable 1 million downloads in per week, although the rollout was invite-only. This sparked rave coverage and an enormous amount of unbelievable ring security videos, gratuitous farting celebrities (deceased ones only), and a few truly disturbing home shopping network ads. By Halloween, the app had been downloaded 4 million times and produced thousands and thousands of AI-generated 10-second videos each day, in response to AppFigures.
So how much money is OpenAI spending on this flood of silly videos? That means it’s greater than $5 billion a 12 months, or about $15 million a day Forbes Estimates and discussions with experts. As Bill Peebles, OpenAI head of Sora, observed On October 30, he said, “The economy is completely unsustainable right now,” and he was exactly right.
These numbers require explanation and are available with some caveats. OpenAI declined to share specific usage data about Sora and comment Forbes’ Estimates. That means Forbes’ Calculations are based on estimates and multiple moving targets – including GPU prices, inference efficiency, user numbers, and the variety of videos posted per day.
Do you’ve a story tip? Contact Phoebe Liu at pliu@forbes.com or phoebe.789 on Signal.
However, it is feasible to get an idea of what all of it costs. Video models (like Sora 2) are rather more expensive than their text counterparts (like GPT-5) because the info they ingest and spit out is more complex. For users attempting to access OpenAI models directly in bulk (via the API or application programming interface), GPT-5 costs about $10 for about 750,000 words. Sora 2 is more complex since it has to process four-dimensional data (three spatial dimensions plus time) and be sure that actions make sense over a couple of dozen frames per second. According to Cantor Fitzgerald analyst Deepak Mathivanan, making a 10-second video — the usual length of a Sora clip, which costs one “video generation” credit — costs OpenAI about $1.3. AJ Kourabi of SemiAnalysis said the number “seems reasonable” but additionally is dependent upon how often the various Sora models are used (some are more complex). Mathivanan’s evaluation assumes that every generation of video takes up about 40 minutes of total GPU time, or 8-10 minutes with 4 GPUs running concurrently, and that renting a GPU costs slightly below $2 per hour. Assuming that OpenAI doesn’t yet factor profit margins into its API pricing, this estimate is: The company currently charges $1 for a 10-second video generated by Sora 2 (and $3 for the more advanced Sora 2 Pro model).
Then there may be the query of what number of videos users create on Sora. The number fluctuates widely and we do not yet know to what extent users will keep coming back to learn more – or when OpenAI will stop offering free access to AI video generation. But take Sora’s estimated 4.5 million app users and assume that 25% of them post a mean of 10 videos per day, in response to Kourabi. That corresponds to 11.3 million videos per day. Multiply by $1.3 per video, which implies almost $15 million per day or $5.4 billion per 12 months. (This number doesn’t bear in mind videos that Sora cuts out before publishing because they’re deemed infringing or infringing, or drafts that use Sora credits but are never published.)
Allowing anyone to make Sora videos without spending a dime is a daring move, although not unusual within the tech world. OpenAI is playing at a loss for market share and visibility, hoping that enough people will use Sora in order that costs go down and – once OpenAI starts charging for it – revenue increases.
“It’s a classic Internet playbook to start by not focusing so much on cost, but rather on building an audience and building engagement, because we’ve seen time and time again that these companies can find ways to monetize that engagement,” says Lloyd Walmsley, an analyst at Mizuho who covers Meta and Google. Walmsley and Mathivanan emphasize that the variety of GPU minutes required to generate a second of video decreases exponentially over time. Mathivanan estimates that inference for video models could turn into five times cheaper by next 12 months and 3 times cheaper in 2027.
OpenAI’s Sora app shot to the highest of the App Store’s free app charts in October. It fell to fifth place on November ninth.
Photo illustration by Cheng Xin/Getty Images
For now, it is a voracious grab for market share, presumably setting the product up for aggressive future monetization. Although Sam Altman, CEO of OpenAI, has said that an promoting model couldn’t possibly cover Sora’s computing costs for the time being, perhaps it may very well be a mix of promoting and power users (filmmakers or TV business creators?) who pay well for the product. Free Sora generations also allow OpenAI to leverage the info of opt-out users to further improve all of its models – who’re hungry for video data for which individuals have already written descriptions (the text prompts). This could improve OpenAI’s bottom line in the long run or give it a training advantage over competing models. (Kourabi assumes that OpenAI’s margins for AI videos after monetization can be somewhere between those of Meta and Google.) Finally, like all operating expenses, OpenAI can use the computational costs of running Sora to avoid wasting taxes by reducing the newly for-profit company’s taxable income from future profits.
Despite potential advantages, costs add up so quickly that OpenAI has announced it should soon stop offering as much free AI video generation. As Altman said in a technique interview in October: “There are so many people who just make funny memes to send to their three friends that there is no advertising model that can support the costs of such a world.”
