OpenAI Spent $15M/Day on Sora. It Made $2M. Total.

Key Takeaways

• Sora is dead after six months: OpenAI shut down its AI (Artificial Intelligence) video app on March 24, 2026, after earning just $2.1 million in total in-app revenue against estimated peak inference costs of $15 million per day.
• The inference cost wall is structural: Generating a single AI video clip costs 10-50x more compute than a ChatGPT text conversation. No pricing model can bridge that gap at consumer scale as of March 2026.
• This is the Iridium playbook: In 1999, Motorola’s Iridium satellite phone system cost $5 billion to build and worked perfectly. It went bankrupt in nine months because calls cost $5 per minute while cellular dropped to pennies. Sora died of the same disease: perfect technology, impossible economics.
• Consumer AI faces a reckoning: If OpenAI, with billions in backing, can’t make video generation pay, the entire category of compute-heavy consumer AI products is on notice.

Six Months, $2.1 Million, and a $1 Billion Corpse

On March 24, 2026, OpenAI posted a farewell on X: “We’re saying goodbye to the Sora app.”

The statement was polite, grateful, and completely devoid of the number that mattered: Sora’s total lifetime in-app purchase revenue was $2.1 million.

That’s not a typo. Two point one million dollars. Total. Over six months. From an app that hit #1 in the App Store’s Photo & Video category within 24 hours of its September 2025 launch and peaked at 3.33 million downloads in November.

Meanwhile, Cantor Fitzgerald analyst Deepak Mathivanan estimated Sora’s peak inference costs at approximately $15 million per day, pegging each clip generation at roughly $1.30 in compute. At that burn rate, Sora’s entire six months of revenue wouldn’t cover three and a half hours of peak compute.

$\text{Hours of compute covered by total revenue} = \frac{\$2.1\text{M}}{\$15\text{M/day}} \times 24 \approx 3.4 \text{ hours}$

The $1 billion Disney deal signed in December 2025, covering character licensing across Disney, Marvel, Pixar, and Star Wars, is also dead. No money ever changed hands.

Bill Peebles, OpenAI’s own head of Sora, admitted on social media that “the economics are currently completely unsustainable.”

This isn’t a story about a bad product. It’s a story about a structural wall that every compute-heavy consumer AI product will hit.

The Inference Cost Wall: Why Video Breaks the Math

To understand why Sora died, you need to understand the difference between training a model and running one.

Training is a one-time cost. You spend hundreds of millions of dollars teaching the model to understand video, and then that cost is amortized across every user forever. The hyperscaler capex cycle that this site has covered extensively is largely a training and infrastructure story.

Inference is a per-use cost. Every time a user clicks “generate,” the model runs. And video inference is a monster. A single high-quality video clip requires 10-50x more compute than a typical ChatGPT conversation. A user generating 20 videos per month consumed more compute than a thousand text sessions.

Here’s the math that killed Sora:

Metric	ChatGPT (Text)	Sora (Video)
Compute per interaction	1x (baseline)	10-50x
Subscription price	$20/month	$20/month (bundled)
Revenue per compute unit	Sustainable	~2% of cost
User behavior	Daily, habitual	Episodic, project-based

ChatGPT works because text inference is cheap enough that a $20/month subscription covers it. Sora broke because video inference is 10-50x more expensive, but users weren’t willing to pay 10-50x more. They expected it bundled or cheap, just like text.

This is the inference cost wall: the gap between what compute-heavy AI costs to run and what consumers will pay to use it. For text AI, that gap has closed. For video AI, it hasn’t.

The Retention Collapse: Nobody Came Back

Cost was only half the problem. The other half was that nobody needed Sora more than once.

By February 2026, fewer than 5% of users who signed up for Sora were returning monthly. Downloads cratered from 3.33 million in November 2025 to 1.13 million by February 2026, a 66% decline in three months.

The pattern was clear: users would generate a handful of clips for a specific project or curiosity, then never come back. AI video generation is inherently episodic rather than habitual. You don’t open Sora over breakfast the way you open ChatGPT. You use it when you need a clip for a presentation or a social post, and then you’re done for weeks.

This meant Sora was burning an estimated $15 million per day to serve a user base that was both shrinking and barely engaged. Compare that to ChatGPT’s reported 900 million weekly active users, who generate recurring, habitual demand that justifies the infrastructure.

And the competitive moat evaporated almost immediately. Runway Gen-3 launched in October 2025, Kling 2.0 in November, and Google Veo 2 in January 2026, all reaching comparable or superior quality within months. When OpenAI tried to scale Sora for broader access, quality degraded dramatically, and users noticed. There was no switching cost, no data lock-in, and no network effect.

The Iridium Parallel: Perfect Tech, Impossible Economics

This has happened before.

In 1998, Motorola launched Iridium, a satellite phone network that cost $5 billion to build. It worked exactly as advertised: you could make a phone call from anywhere on Earth, from the middle of the Pacific Ocean to the top of Mount Everest. The technology was a genuine engineering triumph.

It went bankrupt on August 13, 1999, nine months after launch.

The technology wasn’t the problem. The economics were. Iridium handsets cost $3,000 and calls ran up to $5 per minute. Meanwhile, terrestrial cellular networks were expanding fast and dropping their prices to pennies per minute. Iridium forecast 500,000 subscribers; it got 10,000.

After bankruptcy, Iridium was sold for $25 million, a 99.5% loss for original investors. The satellites kept working. A new owner restructured the business, targeted niche markets (maritime, military, remote industry), and built Iridium into a profitable company with $831 million in 2024 revenue.

The parallels to Sora are precise:

Factor	Iridium (1998)	Sora (2025-2026)
Technology	Worked perfectly	Worked perfectly
Infrastructure cost	$5 billion	Est. $500M-$2B in compute
Per-use economics	$5/minute calls	~$1.30/clip generation
Cheaper alternative	Cellular ($0.10/min)	Text AI ($0.001/query)
User retention	10K vs 500K forecast	5% monthly retention
Time to failure	9 months	6 months
What happened next	Sold for $25M, rebuilt	Compute redirected to ChatGPT

The lesson from Iridium isn’t that satellite phones were stupid. It’s that the economics of serving a mass market with expensive technology don’t work when a cheaper alternative covers the vast majority of use cases. Sora’s “cheaper alternative” isn’t another video generator. It’s text.

The Gray Area: Costs Are Falling, Just Not Fast Enough

Here’s where the Sora pessimists need to pump the brakes.

AI inference costs have been plummeting. Running a GPT-4-class model cost approximately $20 per million tokens in late 2022. By early 2026, equivalent performance costs $0.40 per million tokens, a 1,000x reduction in just over three years.

Gartner published a forecast on March 25, 2026, one day after Sora’s death, projecting that inference costs for a one-trillion-parameter model will fall an additional 90% by 2030.

Four compounding forces are driving this decline: hardware efficiency gains delivering 2-3x more throughput per GPU (Graphics Processing Unit) generation, software optimization pushing GPU utilization from 30-40% to 70-80%, model architecture improvements like MoE (Mixture-of-Experts) delivering 3-5x lower compute per token, and quantization techniques cutting memory requirements by 2-4x.

The question isn’t whether AI video will become affordable. It’s when. Even with another 10x improvement from March 2026 costs, video generation at roughly $0.13 per clip still doesn’t work at the scale where advertising or cheap subscriptions can absorb it. Reaching sub-penny generation costs, where consumer-scale economics start to breathe, likely requires the 90% additional reduction Gartner projects by 2030.

Iridium’s sequel is instructive here. Satellite communication didn’t die. It got cheaper, found niche markets, and eventually became viable. Iridium serves over one million subscribers as of 2018. But it took 20 years and a complete reimagining of the business model.

Consumer AI video will follow the same trajectory. The technology will survive. The economics will eventually work. But the companies funding the gap between now and then will bleed.

What Dies Next: The Consumer AI Reckoning

Sora isn’t the last casualty. It’s the canary in the coal mine.

Every AI product that depends on heavy inference faces the same wall. As you move beyond text, the economics get worse:

Product Category	Inference Cost vs. Text	Consumer Price Sensitivity	Viability (March 2026)
Text chat (ChatGPT)	1x	Low ($20/mo works)	Viable
Image generation	5-10x	Medium	Marginal
Audio/Music generation	10-20x	High	At risk
Video generation	10-50x	Very high	Dead (for now)
Real-time 3D/Game	50-100x	Extreme	Not viable

The further you move from text toward richer media, the worse the economics get. Each step up in compute intensity hits the same inference cost wall. And consumers have been trained by two decades of free internet services to expect AI to be cheap or free.

This leads to a structural split: AI becomes an enterprise tool, not a consumer product. Corporations can absorb higher per-user costs for AI video tools that replace a video production team. Consumers cannot and will not.

OpenAI seems to understand this. Their official statement cited a pivot toward “world simulation research to advance robotics that will help people solve real-world, physical tasks.” Translation: enterprise and industrial applications where the customer can afford the compute bill.

The IPO Math: Why Sora Had to Die Now

There’s a strategic dimension here that the “product failure” narrative misses.

OpenAI is preparing for an IPO. Every dollar of compute burned on Sora was a dollar not earning revenue through ChatGPT, their product that actually makes money. With capital expenditure budgets already stretched across the industry, compute allocation is a zero-sum game.

Killing Sora wasn’t just cutting losses. It was portfolio optimization. OpenAI chose to redirect Sora’s GPU capacity toward ChatGPT and its API (Application Programming Interface) business, where the unit economics work. For an IPO-track company, that’s rational: show investors a path to profitability by shutting down the money furnace.

The Disney deal’s collapse reinforces this. A $1 billion investment sounds enormous, but it wouldn’t have changed Sora’s unit economics. If generating clips costs $1.30 each and users won’t pay more than a few dollars a month, no amount of Disney IP changes the math.

The Iridium Rule for AI Investors

Every technology bubble follows a pattern: build the infrastructure, discover the economics don’t work, crash, then rebuild on the wreckage. Iridium’s $5 billion loss became a $25 million acquisition that became a company generating $831 million in annual revenue.

The infrastructure doesn’t disappear. The investors do.

For anyone evaluating AI companies right now, the Iridium Rule is simple: separate the technology question from the economics question. “Does this AI model work?” is a different question from “Can this AI model generate more revenue than it costs to run?” Sora answered “yes” to the first and “absolutely not” to the second.

The companies that will survive the inference cost wall are the ones serving customers who can afford the compute: enterprises, governments, and specialized industries. The companies betting on mass-market consumer AI video, music, and real-time generation at subscription prices are building Iridium phones in a cellular world.

The technology is real. The costs are falling. But the gap between where inference costs sit in March 2026 and where they need to be for consumer viability is measured in years, not months. Sora was six months early to a party that won’t start until the end of the decade.