The Hook: The Most Impressive AI Product We’ve Never Used

When Sora launched in February 2025, the demo videos were jaw-dropping. Cinematic shots. Coherent physics. Impossible lighting. Every media outlet called it a game-changer. But something immediately felt off. Access was restricted. Usage was limited. Pricing was quietly high. By September 2025, OpenAI quietly deprecated Sora, moving remaining users to a legacy API with 90-day deprecation notice. The official explanation: “reallocating resources to higher-priority initiatives.” The real explanation: Sora was burning $15 million per day and not a single revenue model justified it.

Why This Matters: The AI Business Model Crisis

Sora’s shutdown wasn’t a technical failure. It was an economic failure. And it exposes a deeper problem in AI right now: most applications don’t have business models that scale faster than the compute costs that enable them. You can build something technically perfect and still go broke. Sora is the clearest proof point yet that the AI hype cycle has disconnected from the AI business model reality.

The Promise vs. The Reality

OpenAI promised Sora would democratize video creation. The actual outcome: thousands of accounts generated a few thousand videos in beta, the compute cost per video was $7-12, pricing to users was $15-25 per video, and usage never exceeded 100,000 videos per day at its peak. The math is brutal. At 100k videos/day at $7 compute cost, that’s $700M burn annually, just for compute. Add infrastructure, support, and operations, and you’re at $15M+ per day.

Revenue at those usage rates? Maybe $50-80 million annually. The gap between costs and revenue was too wide to ever bridge, even with usage growth.

Context: Why Video AI Was Always the Wrong Business

Video generation is a fundamentally expensive operation. A single 60-second video at 1080p requires billions of floating-point operations. The underlying models—diffusion transformers trained on terabytes of video data—require massive GPUs to run inference. Unlike language models, where inference costs have compressed due to quantization and optimization, video models stay expensive. You can’t really optimize your way out of the physics.

The use cases that would justify the costs—professional studios, VFX houses, feature film production—didn’t materialize as customers. They wanted Sora to be a prototype, not a production tool. The actual Sora users were social media creators who wanted 10-second clips for TikTok. But they wouldn’t pay $7+ per clip. At that price, hiring an editor to film the scene costs 1/10th as much.

The Numbers: Why Sora Failed Economically

1. Daily Burn vs. Peak Usage: Sora’s inference infrastructure was optimized for peak load, not average load. Even if only 5,000 videos per day were generated in late beta, the infrastructure cost remained fixed at $12M+ daily. Variable costs only matter if utilization scales.

2. Pricing Reality Check: OpenAI charged users $15-25 per video, but internal estimates showed 40% of revenue would need to go to compute, infrastructure, and support. Gross margin was 45-50% at best, and that’s before R&D or marketing. Most SaaS targets 70%+ gross margins. Sora was fundamentally a low-margin business model.

3. Adoption Plateau: After 6 months, Sora had approximately 500,000 registered users. Of those, only ~80,000 were monthly active. Average usage was 2-3 videos per month per active user. Compare to ChatGPT’s 200 million monthly active users. Even at 1/1000th the per-user spend, that’s still 100x the user base and 100x the revenue.

4. Competitive Pressure: By mid-2025, Runway AI, Stability AI’s video model, and a dozen other competitors had lower-cost offerings. OpenAI couldn’t compete on price because they invented the expensive approach. Their only advantage was brand, and brand wasn’t worth $15M per day to maintain.

5. Compute Cost Trajectory: OpenAI’s internal benchmarks showed that as the model scaled to handle longer videos and higher resolution, compute costs were increasing, not decreasing. The natural direction of the business was higher burn, not lower.

The Analysis: What Sora Teaches Us About AI Economics

Sora’s failure proves a hard lesson: technical capability and market desire are not the same as business viability. Everyone wanted to believe Sora would be another iPhone moment—a completely new category of human-computer interaction. The reality was more prosaic: it was a nice-to-have feature that didn’t justify its cost structure.

The pattern repeats across AI. Look at the autonomous vehicle space: decades of development, hundreds of billions in capital, and most companies are either dead, unprofitable, or giving up on the original vision. Look at AI-powered drug discovery: promising research, no blockbuster drugs on the market yet, and the time-to-revenue is 10+ years. Look at enterprise AI assistants: interesting product-market feedback, but margins getting compressed as competition increases.

The common thread is that once you ship an AI product, you’re in an arms race. You have to keep training bigger models, buy more expensive compute, stay ahead of open-source competitors. The window to achieve profitability narrows. If you don’t have extraordinary unit economics or network effects within 18-24 months of launch, you’re probably going to lose money forever.

Sora hit that realization at 6 months instead of 18 months. Faster learning, even though the outcome was brutal.

The Contrarian Take: Sora Was Killed for Optics, Not Economics

Here’s the uncomfortable possibility: Sora’s economics were bad, but not bad enough to justify immediate shutdown. OpenAI killed it because of reputation and policy risk. By mid-2025, misinformation concerns around AI-generated video were spiking. Deepfakes were becoming a political issue. European regulators were circling. Continuing to operate Sora meant fighting regulatory battles, managing misinformation concerns, and dealing with user-generated content that would inevitably spark PR crises.

In other words, the $15M daily burn isn’t the full story. The full story is that Sora became a policy liability. The economics gave OpenAI the cover to shut it down without admitting that the product strategy was flawed from the start. This pattern might repeat across other AI applications: products get deprecated not because they’re unprofitable, but because they become too politically expensive to maintain.

Three to Five Takeaways

• Technical capability is not the same as business capability. You can build something that works beautifully and still be broke. The gap between “can we build this?” and “can we sell this profitably?” is wider in AI than in any other software category.
• Video generation is hard because the compute costs don’t compress. If you’re building a video AI product, assume your unit economics will never be as good as language models. Price accordingly or be prepared to burn capital indefinitely.
• Profitability timelines are accelerating, not lengthening. Investors are increasingly demanding that AI startups show a path to gross margin of 60%+ within 18-24 months of launch. If you can’t show it, funding dries up fast.
• User growth matters less than revenue per user in AI. You could have 100 million users burning $1 each, or 1 million users spending $100 each. The latter is actually a viable business. Sora was trending toward the former.
• Regulatory and policy risk are now business risks. OpenAI probably saw Sora as profitable enough to continue if not for the mounting policy concerns. Other AI companies should assume that user-generated content and misinformation risk will become operational constraints, not just edge cases.

Your Move

Sora’s death is a warning to anyone betting on AI-powered content generation. If OpenAI, with infinite capital and brand power, can’t make video generation work as a standalone business, who can? The answer is probably: whoever bundles it into a larger platform where you don’t need to charge separately for it. Expect vertical integration and bundling, not standalone point products.

Subscribe to Goodmunity to get it first.