The Model That Couldn't Pay Rent

Sora is gone as of March 24, 2026. The numbers, now that they've leaked: $15 million per day in inference costs. $2.1 million in total lifetime revenue. That's not a ratio. That's a confession.

The math was never going to work. Video generation is computationally brutal in ways that text generation isn't — you're not generating tokens, you're synthesizing frames at scale across GPU clusters that cost real money per second. The demo was beautiful. The demo environment is always beautiful. The production environment had to run 15 times the cost of every dollar it brought in, indefinitely, in hopes that adoption would somehow outrun the physics.

It didn't. Downloads fell 66% from their November 2025 peak. The Disney partnership — reportedly worth $1 billion and the kind of deal that would have made the numbers slightly less catastrophic — collapsed. Bill Peebles, who ran Sora, was apparently watching this happen in real time and saying so publicly. They kept the lights on for another five months.

Why? The IPO.

OpenAI is targeting a valuation somewhere between $830 billion and $1 trillion in a late-2026 or early-2027 public offering. You don't show up to that roadshow with a product your own team called unsustainable. You show up with a product portfolio that demonstrates multimodal ambition — text, images, voice, and video — because that's the story institutional investors need to see. The moment the S-1 prep started in earnest, the calculus flipped: Sora the live product became a liability; Sora the capability demonstration had already done its job.

The team is now "pivoting to robotics" under a model codenamed Spud. I'll start the timer on the pivot announcement.

Here's what Sora's shutdown actually reveals, and it has nothing to do with OpenAI specifically: inference costs are the new wall. The scaling era assumed that throwing more compute at a problem eventually crosses a threshold where the economics make sense. For language models, that's roughly true — the cost curve bent, the products monetized. For video, nobody has proven the curve bends in time. Runway, Pika, Google Veo — they're all running the same experiment with different budgets. The question isn't whether AI video can generate impressive outputs. It clearly can. The question is whether anyone will ever pay enough, at scale, to cover what it costs to generate them.

OpenAI's answer, for now, is no. Or more precisely: not yet, not us, not at these margins, not before the IPO.

The honest version of this story isn't "OpenAI failed at video." It's that they shipped a product they knew was underwater, kept it running for optics, and pulled the plug when the optics no longer required it. That's not a failure of the technology. That's a feature of managing an AI arms race while preparing for public markets.

Sora worked. It just didn't work for $15 million a day.

i · sources

• Why OpenAI Really Shut Down Sora — TechCrunch, 2026-03-29
• OpenAI Sora Shutdown: $15M/Day Costs, $2.1M Revenue — The Full Story — Medium, 2026-03
• Inference Costs Are the New AI Wall: What Sora's Shutdown Tells Us — MindStudio, 2026