The Most Expensive Failure in AI History

How OpenAI burned $5.4 billion annualized on a product that earned $2.1 million total and what every practitioner should learn from the wreckage

Author

Justin Wright
March 30, 2026 • Est. Reading Time: 9 minutes

Happy Monday!

OpenAI killed Sora on Tuesday. Only six months after launching to massive fanfare, the AI video generator is gone. The app shuts down April 26, and the API follows in September.

The numbers behind the shutdown are staggering. At its peak, Sora burned an estimated $15 million per day in inference costs. Each 10-second video cost roughly $130 in compute. Total lifetime revenue from in-app purchases was only $2.1 million. Not $2.1 million per month, $2.1 million total across the product's entire existence.

Sora's head of product, Bill Peebles, admitted on social media last year that "the economics are completely unsustainable." He was right, and the collateral damage extends beyond OpenAI: a $1 billion Disney licensing and investment deal died with it. No money ever changed hands.

This could be the clearest signal yet about what actually survives in AI and what doesn't. The most dangerous number in this industry isn't a benchmark score, but the inference cost per unit of revenue.

OpenAI shut down Sora after burning an estimated $15 million per day in inference costs against $2.1 million in total lifetime revenue. Downloads fell 75% from peak. A $1 billion Disney deal collapsed. The shutdown is a pre-IPO cleanup, but the lesson is broader: consumer AI products that can't close the gap between inference costs and revenue don't survive, regardless of how impressive the technology is.

TL;DR

The Numbers That Killed Sora

Sora's launch was genuinely impressive by consumer metrics. It topped the App Store's Photo and Video category within a day, hit 1 million downloads in under five days, and peaked at 3.33 million monthly downloads in November 2025.

Then gravity took over. Downloads dropped 32% in December, another 45% in January, and by shutdown, user growth had fallen nearly 75% from its peak. The viral moment faded, and the users who stayed weren't paying. The entire product generated $2.1 million in in-app purchases across its lifetime while consuming compute resources at a rate that would have cost $5.4 billion annualized.

The unit economics were never close. Unlike text inference, where costs have dropped dramatically over the past two years, video generation remains extraordinarily expensive. Each generation requires massive GPU compute for diffusion model processing, and the cost curve hasn't bent the way text and image generation costs have. OpenAI couldn't charge enough to cover inference, couldn't reduce inference costs fast enough to match what users would pay, and couldn't find an advertising or licensing model that bridged the gap.

Sora's Economics at a Glance

Metric

Number

Peak daily inference cost

~$15 million

Annualized inference cost

~$5.4 billion

Total lifetime revenue

$2.1 million

Cost per 10-second video

~$130

Peak monthly downloads

3.33 million (Nov 2025)

Downloads at shutdown

~830K/month (75% decline)

The Disney Casualty

In December 2025, Disney announced a three-year deal to license over 200 characters for use in Sora-generated videos. This included Mickey Mouse, Marvel heroes, Pixar characters, and Star Wars IP. Disney also committed a $1 billion equity investment in OpenAI. The plan was for curated Sora videos to appear on Disney+.

Unfortunately, the deal never closed. When OpenAI announced the shutdown, a Disney insider told Deadline: "The deal is not moving forward." The billion-dollar investment evaporated alongside the product it was meant to support.

For practitioners, the Disney collapse illustrates a specific risk in AI partnerships: when your partner's product economics are unsustainable, no amount of brand power or content licensing can save it. Disney bet on Sora's technology without adequately stress-testing Sora's business model. That's a mistake enterprises are making across the AI landscape right now, signing deals based on capability demonstrations without asking hard questions about unit economics.

The Pre-IPO Cleanup

OpenAI is targeting a Q4 2026 IPO. The Sora shutdown reads very differently through that lens.

No prospective public investor wants a product burning $15 million per day against $2.1 million in total revenue on the books. Killing Sora before the roadshow is a balance sheet decision. Clean up the losses, sharpen the enterprise story, go public on numbers that make sense.

OpenAI's applications CEO Fidji Simo told employees the company is "orienting aggressively" toward high-productivity use cases. The compute freed up by shutting Sora is being redirected to coding tools, enterprise productivity, and the ChatGPT super-app strategy. Sora's research team continues as a unit focused on world simulation and robotics, where the commercial applications have pricing power a consumer video app never had.

The pattern is familiar from previous tech IPOs: cut the money-losing consumer experiments, consolidate around revenue-generating enterprise products, and present a clean growth story. OpenAI is following the playbook because the playbook works. But it also means the company that launched Sora as a vision of creative AI's future quietly admitted that vision alone doesn't pay for itself.

What This Actually Means for the Industry

Sora's failure isn't unique. It's a preview of what happens when inference economics collide with consumer pricing expectations across every modality.

Text inference costs have dropped enough to support viable consumer products. Image generation costs have fallen significantly. But video, 3D generation, and real-time multimodal processing remain in a fundamentally different cost bracket. The companies still operating in AI video, like Google's Veo, Runway, and Luma AI, are navigating the same equation. None of them have solved the unit economics at consumer scale either.

For practitioners, the Sora postmortem offers three concrete lessons:

First, capability and viability are different things. Sora generated stunning video. It also generated stunning losses. Before building on any AI capability, stress-test the inference cost per transaction against what users will actually pay. If the gap is orders of magnitude, the technology isn't ready for production regardless of how impressive the demos look.

Second, consumer AI has a pricing ceiling that enterprise AI doesn't. Consumers expect free or near-free access to AI tools. Enterprise customers will pay $50 to $200 per seat for productivity gains. The inference costs of advanced AI capabilities increasingly only work at enterprise price points. That's why OpenAI is pivoting hard toward business customers.

Third, the pre-IPO cleanup tells you where the real money is. OpenAI is redirecting Sora's compute toward coding tools and enterprise agents. Anthropic built its business around enterprise from the start. The market is converging on a simple reality: the AI products that survive are the ones where inference costs are a fraction of the value delivered, not a multiple of it.

The Bottom Line

Sora was the most technically impressive consumer AI product of 2025. It was also, by the numbers, the most expensive product failure in AI history. $15 million per day in costs against only $2.1 million in total revenue. A $1 billion Disney deal that died on the table.

The technology works, it was just that the economics don't. And in a pre-IPO environment where every dollar of compute needs to justify itself, "works but doesn't pay for itself" is a death sentence. For anyone building AI products, Sora's six-month lifespan is the clearest case study yet: the benchmark that matters most isn't accuracy, creativity, or user delight. It's inference cost per dollar of revenue. Everything else is a great looking demo.

In motion,
Justin Wright

If the most well-funded AI company in history couldn't make consumer video generation economically viable in six months, what does that tell us about the timeline for AI capabilities that are even more compute-intensive, and about which AI products will actually survive the transition from impressive demo to sustainable business?

Food for Thought

If you haven’t listened to my podcast Mostly Humans: An AI and business podcast for everyone yet, new episodes drop every week!

Episodes can be found below - please like, subscribe, and comment!