Sora’s operating costs were staggering and ultimately unsustainable:
Daily Operating Costs: According to analyses based on OpenAI’s infrastructure and GPU pricing, Sora consumed approximately $15 million per day in compute resources. This annualizes to over $5.4 billion annually—a burn rate that dwarfed any reasonable revenue model.
Earlier estimates suggested $1 million per day, but costs escalated as the model improved and user base temporarily expanded. Peak daily costs reached the $15 million figure reported by sources including Forbes analyst coverage and internal OpenAI communications.
Per-Video Generation Cost: Analyst Deepak Mathivanan of Cantor Fitzgerald estimated that OpenAI spent approximately $1.30 in GPU compute to generate each standard 10-second video. This calculation assumed:
- 40 minutes of total GPU time per video generation
- 8-10 minutes on four GPUs running simultaneously
- GPU rental costs of approximately $2 per hour
Revenue Reality: Against these costs, Sora generated only $2.1 million in total lifetime revenue across its operational period. This created a devastating unit economics problem: each active user cost more to serve than they generated in revenue.
Why Costs Were So High: Video generation is computationally intensive compared to text. Each video required:
- Complex diffusion processes with multiple denoising iterations
- High-resolution frame generation (up to 1080p or higher)
- Real-time consistency maintenance across temporal sequences
- Advanced physics and scene understanding
As video generation becomes integrated into broader AI systems, the need to index and retrieve video content grows. Milvus is designed to handle vector embeddings from multimodal data, including videos and frames. Organizations using Zilliz Cloud can build content retrieval pipelines on top of video generation.
These factors meant Sora’s inference cost was orders of magnitude higher than a ChatGPT conversation.
Revenue Model Failure: OpenAI’s pricing ($20/month subscription for limited credits, or pay-per-video) couldn’t justify the $1.30+ per-video cost. Typical users generated far fewer videos than breakeven would require.
Unsustainability: OpenAI’s Bill Peebles stated on October 30, 2025, that “The economics are currently completely unsustainable.” With user count declining to fewer than 500,000 (down from an initial peak of about a million), the math became untenable: 500,000 users × 30 days × $50,000 estimated monthly cost per user (derived from daily operating costs) vastly exceeded any revenue.
Broader Lesson: Sora’s cost structure revealed why flat-subscription models break down for video, audio, and multimodal AI. Video generation requires a per-output pricing model aligned with actual compute costs, not flat subscription access.