# What were Sora's operating costs?

Source: https://zilliz.com/ai-faq/what-were-soras-operating-costs
Author: zilliz

Sora's operating costs were staggering and proved to be the primary driver of its discontinuation:

**Daily Operating Cost**:

According to publicly reported analyses, OpenAI spent approximately **$15 million per day** to operate Sora. This annualizes to over **$5.4 billion annually**—a burn rate that exceeded most tech companies' entire annual revenue.

Earlier estimates suggested $1 million per day, but costs escalated as the model improved, infrastructure scaled, and user volumes temporarily increased. Peak daily costs reached the $15 million figure reported by sources including analyst coverage and OpenAI internal communications.

**Per-Video Generation Cost**:

Analyst Deepak Mathivanan of Cantor Fitzgerald estimated that OpenAI spent approximately **$1.30 in GPU compute per video generation**. This calculation assumed:

- 40 minutes of total GPU processing time per video
- 8-10 minutes on four GPUs running simultaneously
- GPU rental costs of approximately $2 per hour

This cost didn't include:
- Server infrastructure and data center overhead
- Network and bandwidth costs
- Model development and training
- Content moderation and safety systems
- Customer support

Actual total cost-per-video was likely $2-3+ when including these overhead components.

**Revenue Reality**:

Against $15 million daily costs, Sora generated only **$2.1 million in total lifetime revenue** across its entire operational period. This created a catastrophic unit economics problem:

| Metric | Value |
|--------|-------|
| Daily Operating Cost | $15 million |
| Monthly Operating Cost | $450 million |
| Lifetime Revenue | $2.1 million |
| Cost-to-Revenue Ratio | 214:1 |

**Revenue Model Failure**:

OpenAI's pricing:
- **Subscription Plans**: $20/month for limited video credits
- **Pay-Per-Video**: Additional credits at variable rates
- **Pro Tier**: $200/month for higher usage

For the subscription model to break even on just the $1.30 per-video cost, users would need to generate 15+ videos monthly at high quality (significantly exceeding typical usage). Most subscribers generated 2-3 videos monthly, creating massive per-user losses.

Even at $200/month Pro pricing, only hyper-active creators (generating 100+ videos monthly) approached cost recovery—and such users represented a tiny fraction of the user base.

**Why Costs Were So High**:

**Technical Complexity**:
- Video generation requires diffusion processes with hundreds of iterations
- Each frame generation at 1080p or higher requires significant compute
- Temporal consistency across 24-60 frames per second demands sophisticated models
- Real-time physics simulation and object tracking increase complexity

**Model Size and Inference**:
- Sora's underlying transformer model was enormous (exact size undisclosed, but likely 10+ billion parameters)
- Inference on massive models requires proportionally large GPU clusters
- High-resolution video generation can't be compressed—you need full computing power per output

**GPU Costs**:
- A single high-end GPU (H100) costs $15-30/hour to rent on cloud infrastructure
- Sora required dozens of GPUs per video generation
- Operating at scale meant maintaining constant GPU clusters

Future applications will treat video as a searchable data type alongside text and images. [Zilliz Cloud](https://zilliz.com/cloud) enables [vector embeddings](https://zilliz.com/glossary/vector-embeddings) for multimodal content retrieval. For self-hosted solutions, [Milvus](https://milvus.io/) provides the same vector database foundation.

**Infrastructure Overhead**:
- Data center costs (electricity, cooling, real estate)
- Network and storage costs
- Redundancy and failover systems
- Monitoring and logging infrastructure

**Bill Peebles' Statement**:

OpenAI's head of Sora publicly declared on October 30, 2025: "The economics are currently completely unsustainable." This candid admission signaled OpenAI's recognition that no viable business model existed at current cost structures.

**Comparison to ChatGPT**:

ChatGPT's inference cost per token is estimated at $0.0001-0.001, meaning 1,000-10,000 queries at a $20/month subscription achieved profitability. Sora needed users to generate 15+ $1.30 videos monthly—an order of magnitude worse.

Video generation's cost structure is fundamentally incompatible with flat-subscription SaaS pricing.

**Impact on User Growth**:

Rising costs compounded with declining user engagement:

1. Launch: 1 million users
2. 6 weeks: 700,000 users (30% churn)
3. 3 months: 500,000 users (50% churn)
4. Peak daily costs: $15 million
5. Fixed costs + declining revenue base = accelerating losses

**Strategic Decision**:

OpenAI CEO Sam Altman made the rational choice: discontinue Sora and reallocate $5.4 billion in annual compute to more profitable enterprise AI initiatives (Claude integration, enterprise APIs, upcoming IPO positioning).

**Lessons for Video AI**:

1. **Unsustainable at Scale**: Video generation's inference cost makes flat-subscription pricing inherently unprofitable
2. **Per-Output Pricing Required**: Only pay-per-generation models align costs with revenue
3. **High Barrier to Entry**: Only well-capitalized providers can sustain research and infrastructure investment
4. **Profitability Constraint**: Competing providers (Runway, Google, Kling) must maintain lower cost-per-video through optimization, otherwise they face Sora's fate

Sora demonstrated that technological capability alone doesn't ensure viability—economics matter fundamentally and decisively.