9.6 million downloads. $2.1 million in total revenue. An estimated $15 million per day in compute costs at peak.
On March 24, OpenAI shut down Sora, the AI video generation app it launched six months earlier. Bill Peebles, the head of Sora, had posted on X back in October that “the economics are currently completely unsustainable.” By March, OpenAI made the call.
This was not a product that failed to find users. It was a product that succeeded at the wrong thing.
The numbers that looked like success
Sora hit number one in the App Store on day one. It reached 3.3 million downloads in its first month, faster initial adoption than ChatGPT. Disney signed a three-year licensing agreement, reported by Variety, worth $1 billion, covering more than 200 characters from Disney, Marvel, Pixar, and Star Wars.
By every metric that typically appears on a product dashboard, Sora was performing. Downloads were in the millions. Press coverage was constant. Brand partnerships were landing.
But the decline was already visible. Downloads dropped 32% in December, then another 45% in January, falling to 1.2 million. The US accounted for $1.1 million of the total $2.1 million in lifetime revenue. The gap between compute costs and revenue was not narrowing. It was widening.
The real calculation
Every GPU running Sora was a GPU not running ChatGPT, not powering enterprise API customers, not training OpenAI’s next model. Sam Altman told employees to stop pursuing “side missions.” Sora turned out to be the most expensive one.
The timing was not accidental. OpenAI is laying the groundwork for a potential IPO, having recently hired a head of investor relations and discussed internal targets of filing in the second half of 2026. Unprofitable, compute-intensive consumer products are exactly what a company trims before going public. The Sora team’s research on world simulation and physics understanding is being redirected toward robotics, where the same capabilities have a clearer path to commercial value. That pivot lands in a market where $14 billion in robotics funding poured in during 2025 alone.
Internally, compute is also being reallocated toward a new language model codenamed “Spud,” which Altman expects to release within weeks.
The Disney deal collapsed in the process. As recently as the day of the shutdown announcement, teams from both companies had been meeting about the Sora project. Thirty minutes after that meeting ended, according to the Hollywood Reporter, the Disney side was informed OpenAI was killing the app. No money ever changed hands.
What this exposes about AI resource allocation
Most companies measure AI initiatives by adoption: active users, engagement rates, internal satisfaction surveys. These are the metrics that made Sora look healthy.
What almost never gets measured is opportunity cost. When tool proliferation is already straining teams, adding another initiative without asking what it displaces is how organizations end up busy but unproductive. A pilot project with 500 active users looks reasonable on a dashboard. But if those same compute hours, engineering resources, and management attention went to the initiative that actually moves the business forward, the math changes entirely.
OpenAI, valued at $730 billion after its recent $120 billion funding round, ran this calculation and chose to walk away from 9.6 million users. The question was not whether Sora was working. The question was whether the resources it consumed could work harder elsewhere.
The pattern worth watching
This is not the first product OpenAI has pulled back. The company also scaled back its ChatGPT shopping features, as reported by Axios, in the same period, part of a broader shift from consumer experimentation toward enterprise focus and core model development.
The pattern: a company with functionally unlimited access to capital is still making hard choices about compute allocation. If OpenAI cannot afford to run everything it builds, the constraint is not money. It is physics - the availability of chips, energy, and data center capacity that the entire industry is racing to secure.
For companies running their own AI portfolios on a fraction of those resources, the Sora shutdown is a stress test worth borrowing. Not every initiative that shows adoption is worth the resources it consumes. The question that killed a product with 9.6 million users is the same one that applies at any scale: what else could these resources be doing?