Some back-of-napkin math suggests OpenAI is spending more than a quarter of what it’s making to power its AI slop factory.
For a company that’s burning more than twice what it’s earning, OpenAI is giddily rolling out some ingeniously reckless new ways of racking up losses. Valued at about $500 billion, the AI behemoth on Thursday projected an annual recurring revenue run rate of $20 billion. Which is all well and good until you remember it lost more than $12 billion last quarter.
On September 30, OpenAI debuted its Sora video creation app for Apple’s iOS platform racking up a stunning 1 million downloads in a week despite an invitation-only rollout, inspiring a swoon of fawning coverage and vast slop of fantastical Ring security videos, gratuitously farting celebrities (deceased only), and some truly disturbing home shopping network ads. By Halloween the app had been downloaded 4 million times, per AppFigures, and was churning out millions of 10-second AI-generated videos per day.
So just how much money is OpenAI dumping on this firehose of imbecilic video? More than $5 billion annualized, or around $15 million per day, according to Forbes estimates and conversations with experts. When Bill Peebles, OpenAI’s head of Sora, observed on October 30 that “The economics are currently completely unsustainable,” he was right on the money.
These numbers bear some explanation and come with a handful of caveats. OpenAI declined to share specific usage data on Sora and to comment on Forbes’ estimates. That means Forbes’ calculations rely on estimates and several moving targets—including GPU prices, inference efficiency, user counts and the number of videos being posted per day.
Have a story tip? Contact Phoebe Liu at pliu@forbes.com or phoebe.789 on Signal.
Still, it’s possible to get an idea of what this all costs. Video models (like Sora 2) are much more expensive than their text counterparts (like GPT-5) because the data they ingest and spit out is more complex. For users trying to access OpenAI models directly in bulk (via its API, or application programming interface) GPT-5 costs approximately $10 for some 750,000 words. Sora 2 is more complex, since it needs to process four-dimensional data (three spatial dimensions plus time) and make sure actions make sense over a couple dozen frames per second. Generating a 10-second video—the standard length of a Sora clip that costs one “video generation” credit—costs OpenAI approximately $1.3, according to analyst Deepak Mathivanan of Cantor Fitzgerald; AJ Kourabi of SemiAnalysis said the figure “seems reasonable” but also depends on how often the different Sora models are used (some are more complex). Mathivanan’s analysis assumes each video generation takes around 40 minutes of total GPU time, or 8-10 minutes on four GPUs running at the same time, and that renting a GPU costs just under $2 per hour. Assuming OpenAI isn’t building profit margins into its API pricing yet, this estimate checks out: the company is currently charging $1 for a 10-second video generated by Sora 2 (and $3 for the more advanced model Sora 2 Pro).
Then there’s the question of how many videos users are creating on Sora. The number is wildly fluctuating, and we don’t yet know to what extent users will keep coming back for 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 per Kourabi that 25% of them post on average 10 videos a day. That comes out to 11.3 million videos per day. Multiply by $1.3 per video, which means nearly $15 million per day, or $5.4 billion per year. (This number does not account for videos Sora scraps before releasing because they are deemed infringing or violative, or drafts that use Sora credits but never get posted.)
Allowing just anyone to generate Sora videos for free is a gutsy move, though not uncommon in the tech world. OpenAI is making a play for market share and visibility at a loss, hoping that enough people use Sora for costs to come down and—once OpenAI starts charging for it—revenues to go up.
“It’s a classic internet playbook to not focus on the costs initially so much as building an audience and building an engagement because we’ve seen time and again, these companies can figure out ways to monetize this engagement,” says Lloyd Walmsley, an analyst at Mizuho who covers Meta and Google. Walmsley and Mathivanan stress that the number of GPU-minutes required to generate a second of video will decrease exponentially over time. Mathivanan estimates that inference for video models could get five times cheaper by next year and three times cheaper than that in 2027.
For now, this is a gluttonous market share grab that presumably sets the product up for aggressive future monetization. Though OpenAI CEO Sam Altman has said an ad model couldn’t possibly pay for Sora’s compute costs at the moment, perhaps a combination of advertising and power users (filmmakers or TV ad creators?) who pay handsomely for the product would. Free Sora generations also allow OpenAI to use the data from users who don’t opt out to further improve all of its models—which are hungry for video data for which humans have already written descriptions (the text prompts). That may help OpenAI’s bottom line in the future or give it a training advantage over competing models. (Kourabi thinks OpenAI’s margins for AI video, once monetized, would fall somewhere in between Meta’s and Google’s.) Finally, like all operating expenses, OpenAI can use the compute costs for running Sora to save on taxes, by reducing the newly-for-profit company’s taxable income from future profits.
Despite potential benefits, the costs are adding up so fast that OpenAI has said it plans to stop offering so much free AI video generation soon. As Altman said in a Stratechery interview in October: “There’s so much usage where people are just making funny memes to send to their three friends and that there is no ad model that can support the cost of that kind of a world.”
Read the full article here
