OpenAI shuts down Sora: a unit economics lesson for AI product managers

OpenAI announced on March 24, 2026 that it is shutting down Sora — its standalone AI video generation app and API — just six months after public launch. The app shuts down on April 26, 2026, and the API on September 24, 2026. Video generation through ChatGPT ends with it.

The economics were the deciding factor. Estimated inference costs ran approximately $15 million per day against total lifetime revenue of roughly $2.1 million. Each 10-second generated clip cost OpenAI about $1.30 to produce. Peak usage reached around one million active users before collapsing to under 500,000 — a drop of roughly 75% from the November 2025 high. No subscription price that was viable for users could cover that cost structure at consumer scale.

OpenAI’s stated rationale is simplifying the product portfolio and redirecting compute toward coding tools, enterprise customers, and robotics research — areas with higher return on GPU time. A practical factor is also the company’s anticipated IPO timeline, where a product generating $15 million in daily costs against negligible revenue would appear directly in financial disclosures.

The shutdown had collateral damage. Disney had committed $1 billion to a licensing partnership that would have allowed Sora to generate video content from over 200 characters across Disney, Marvel, Pixar, and Star Wars properties. Disney was notified less than an hour before the public announcement, and the deal collapsed.

For product managers, Sora offers a clear case study in three patterns that surface repeatedly in AI product failures. First, technical quality is not a sufficient condition for commercial viability: Sora’s video output was genuinely impressive, but impressive output at $1.30 per clip cannot be priced into a subscription that users will pay. Second, novelty-driven adoption does not indicate product-market fit: a 75% retention drop from peak to six months later shows that users found limited recurrent utility, regardless of initial enthusiasm. Third, compute cost is a first-class product constraint in AI development — not an engineering detail but a variable that determines what business models are structurally possible before a single user is acquired.