OpenAI Killed Sora: $15M/Day Costs, $2.1M Revenue

$15 million a day in inference. $2.1 million in total revenue. Six months of life. A lesson the industry would rather not hear.

---

On March 24, 2026, OpenAI announced it’s shutting down Sora. The generative video app that was supposed to revolutionize content creation. The one with a billion-dollar Disney deal. The one that was going to do for video what ChatGPT did for text.

It lasted six months.

It didn’t die because the technology wasn’t there. It died because the math didn’t work. Not even close.

The Numbers Nobody Wanted to See

Here’s the breakdown:

• Inference cost: $15 million/day ($5.4 billion annualized)
• Total revenue over six months: $2.1 million
• Cost per 10-second video: $1.30
• Active users at shutdown: fewer than 500,000
• Download decline: from 3.3 million (November 2025) to 1.1 million (February 2026) — a 66% drop in three months

Read that again. Fifteen million dollars a day to generate revenue that wouldn’t cover a single hour of operations.

The Disney Deal That Never Was

Disney had signed a $1 billion agreement. Three-year license. Over 200 characters from Disney, Marvel, Pixar, and Star Wars available for AI video generation.

They found out about the shutdown less than an hour before the public announcement.

Not a single dollar ever changed hands.

The Deepfake Problem

Within weeks of launch, deepfakes of Martin Luther King Jr. and Robin Williams were circulating. The daughters of both figures protested publicly. The MLK estate threatened legal action.

OpenAI’s initial policy was opt-out: your likeness was available by default unless you explicitly asked to be excluded. That’s like a bank opening a credit line in your name without asking and telling you “you can cancel it whenever you want.”

Why This Matters Beyond Sora

Sora isn’t an isolated case. It’s the most visible symptom of a structural problem: generative model inference costs don’t scale with the revenue they produce.

Training a model is expensive. But running it millions of times a day for users paying $20/month (or nothing) is where the hole becomes unsustainable.

This is exactly what FinOps for AI has been warning about for months: inference costs in production are a trap for anyone who isn’t measuring them.

And it’s not just a video problem. If 35,000 LLM calls can cost €150 in three days on a personal system, imagine what it costs to serve generative video to millions of users.

The Lesson for Businesses

OpenAI is preparing for an IPO in late 2026. Wall Street doesn’t want consumer experiments. It wants recurring enterprise revenue. Sora was the exact opposite: a cash furnace with declining metrics.

If your company is considering building something with generative AI — especially video, image, or audio — ask yourself these questions before signing anything:

1. What does each inference cost? Not the pilot. Full production scale.
2. Who’s paying? If the user isn’t paying enough to cover the cost, you’ve got a Sora.
3. What’s the plan B when GPU costs rise? Because the hardware supply crunch isn’t over.
4. Are you solving a real problem or demoing technology? Sora was impressive. And completely useless as a business.

Not Everything Is AI

The dominant narrative says AI will change everything. And maybe it will. But “change everything” doesn’t mean “be profitable from day one” or “scale with no cost ceiling.”

$7 trillion looking for returns. And the flagship product of the industry’s leading company lasted six months.

That doesn’t mean AI is a bubble. It means the difference between the companies that will survive and those that won’t comes down to one word: margins.

The most impressive technology in the world is worthless if you lose money every time someone uses it.

Keep exploring

• 35,000 LLM Calls to Say ‘Nothing Found’ — The hidden costs of multi-agent systems when nobody’s watching the meter