Sora burned $15M/day in compute against $2.1M lifetime revenue. The most expensive lesson in AI product economics.
Get new posts delivered to your inbox. No spam, unsubscribe anytime.
On March 24, 2026, OpenAI posted a brief message on X announcing that Sora --- its AI video generator that once hit #1 on the App Store --- was done. Six months of life. Roughly $15 million per day in inference costs. A total lifetime revenue of $2.1 million. That's not a typo. The product burned more money every single day than it earned in its entire existence.
This isn't just a product shutdown. It's the most expensive lesson in AI economics we've seen so far --- and every founder, engineer, and investor building AI products should be paying attention.
Let me lay out the financials, because they're almost absurd.
| Metric | Value |
|---|---|
| Launch date | September 30, 2025 |
| Shutdown announced | March 24, 2026 |
| Peak daily inference cost | ~$15 million |
| Cost per 10-second clip | ~$1.30 |
| Estimated annual compute cost | ~$5.4 billion |
| Total lifetime in-app revenue | $2.1 million |
| Peak monthly active users | ~6.1 million (Dec 2025) |
| Users at shutdown | fewer than 500,000 |
| 30-day user retention | 1% |
| 60-day user retention | ~0% |
Sources: Forbes and Cantor Fitzgerald analysis, Business of Apps, NerdLevelTech
Read that 30-day retention number again. One percent. Out of every hundred people who downloaded Sora, one was still using it a month later. By day 60, that number approached zero.
OpenAI had built something that was technically impressive and economically impossible.
Sora's trajectory looked like a rocket that ran out of fuel at 10,000 feet.
The launch was enormous. When Sora hit iOS on September 30, 2025, it racked up 164,000 downloads in 48 hours and reached 1 million downloads within five days. By November, when the Android app launched, it was pulling nearly 500,000 installs on day one. Peak monthly downloads hit somewhere between 3.3 and 6 million in November 2025.
Then the cliff. By February 2026, downloads had dropped 66% from their peak --- down to roughly 1.1 million per month. Active users cratered from 6.1 million in December to 4.7 million by February. By the time OpenAI pulled the plug, fewer than 500,000 people were using it monthly.
January 10, 2026 was the inflection point. OpenAI removed free-tier video generation, restricting access to Plus ($20/month) and Pro ($200/month) subscribers only. The thinking was obvious: stop the bleeding by cutting off non-paying users. But the users who left weren't coming back regardless.
The Wall Street Journal described the output as "more AI slop than AI magic." Users flooded social media with generated content that blatantly ripped off copyrighted characters --- Dragon Ball Z, SpongeBob, you name it. Professional video creators reported that generation speed, control granularity, and consistency weren't close to what real workflows required.
Sora was a toy. A very expensive toy.
Here's the part that stings the most.
Disney had been negotiating a $1 billion investment in OpenAI alongside a three-year licensing agreement. The deal would have given Sora access to over 200 masked, animated, and creature characters from Disney, Marvel, Pixar, and Star Wars. Imagine typing "Buzz Lightyear surfing on Mars" and getting a polished 30-second clip. That was the vision.
No money ever changed hands. The deal was never finalized.
And here's the kicker: Disney learned about the shutdown less than an hour before the public announcement. Less than an hour. To a company you're asking for a billion dollars.
Sam Altman later admitted he "felt terrible" telling Disney CEO Josh D'Amaro. D'Amaro reportedly said, "I get it." But Disney shelved the investment entirely.
Altman says OpenAI is "working hard with Disney to find a world where they can still do something amazing," which is corporate-speak for "please don't be mad, we'll figure something out."
The Disney collapse isn't just embarrassing. It reveals the fundamental problem: Sora couldn't sustain itself long enough to reach the partnerships that might have given it a viable business model.
Here's what most coverage of the Sora shutdown misses: this wasn't just a Sora problem. It's a video AI problem. The economics of generating video with AI are structurally different from text or images, and the gap is enormous.
Think of it this way:
That's not a metaphor. That's roughly the compute ratio.
| Modality | Approximate Cost | Gross Margin |
|---|---|---|
| Text (budget LLM) | $0.004 per million tokens | Up to 98.5% |
| Text (frontier model) | $3-$75 per million tokens | 50-65% |
| Image generation | $0.005-$0.17 per image | Varies widely |
| Video (low-res, 720p) | $0.50-$1.00 per minute | Negative for most |
| Video (high-end, 4K) | $12-$30+ per minute | Deeply negative |
Sources: Inference unit economics guide, AI video generation costs, Epoch AI profitability analysis
Text AI can offer generous usage within a $20 subscription. Anthropic's Claude, for example, operates at 50-55% gross margins. That's tight compared to traditional SaaS (which averages 77%), but it's survivable.
Video AI? A single generation burns through credits in seconds, consuming orders of magnitude more compute than an entire text conversation. Runway, arguably the most successful AI video company, booked $44 million in revenue against a $155 million EBITDA loss in 2024. That's a -352% EBITDA margin. Runway is surviving because it raised $860 million across 7 rounds and has a $5.3 billion valuation. But surviving isn't the same as being sustainable.
Sora's pricing tells the story. A 10-second standard clip through the API cost $1.00. A Pro HD clip cost $5.00. At $15 million per day in inference costs, you'd need to sell 3 million Pro clips daily just to break even on compute --- ignoring every other cost of running a company.
Sora's death didn't kill AI video. It just proved that the brute-force approach doesn't work. The survivors are finding different angles.
Runway (Gen-4.5) went premium. Best temporal consistency and motion control in the market. They're the filmmaker's tool --- expensive, but professionals will pay for precision. 300,000 customers and roughly $90 million annualized revenue.
Kling AI 3.0 (by Kuaishou) went cheap. Generates equivalent quality at roughly 40% of Sora's cost per second. Can produce up to 2 minutes of video versus Sora's 25-second cap. Pricing sits at $0.07 per second. Over 10 million videos generated since launch.
Google Veo 3.1 went ecosystem. Free tier with 10 generations per month for any Google account holder. Deep integration with Google Drive, YouTube Studio, and Google Ads. Google can subsidize video generation because it drives engagement across a $300 billion ad business.
Pika went creative. $8/month starting price, $470 million valuation, not competing on photorealism but on fun tooling features like Pikaswaps and Pikatwists.
The pattern is clear:
| Company | Strategy | Why It Works |
|---|---|---|
| Runway | Premium / professional | Professionals pay for control |
| Kling | Cost efficiency | 40% cheaper, 5x longer clips |
| Google Veo | Ecosystem subsidy | Video generation drives ad revenue |
| Pika | Creative / consumer | Low price, fun features, low compute |
Sora tried to be everything to everyone at a price point that couldn't sustain any of it.
Sora didn't fail in isolation. It's a symptom of a systemic problem across AI.
OpenAI projects a $14 billion loss in 2026. That's 3x worse than 2025. Cumulative projected losses through 2029: $44 billion. The company doesn't expect profitability until 2029, when revenue hits an estimated $100 billion.
Let that sink in. OpenAI's own forecast says they'll burn $44 billion before making money.
Anthropic isn't profitable either. They hit $19 billion in annualized revenue by March 2026, with staggering 1,167% year-over-year growth. But planned 2026 spending is $12 billion on training plus $7 billion on inference infrastructure --- $19 billion total, roughly equal to their revenue. Cash-flow break-even isn't expected until 2028.
90% of AI-native startups fail within their first year. Of the 14,000+ AI startups launched in 2024, roughly 3,800 shut down in 2025 and another 1,800 in early 2026 --- a 40% failure rate in under two years.
70% of companies offering AI-driven capabilities are struggling with delivery costs that undermine profitability.
The pattern keeps repeating: impressive demo, viral launch, unsustainable unit economics, death.
Remember Humane AI Pin? Burned through $230 million, shipped fewer than 10,000 units, sold to HP for $116 million. Rabbit R1? Sold 100,000 units but faced mass returns. Builder.ai? Burned $445 million "pretending AI was humans" before entering insolvency.
There's a phrase bouncing around VC circles that captures this perfectly: "You can get to 80% with 20% of the effort. But production demands 99%, and that last stretch can take 100x more work."
Sora nailed the demo. The February 2024 preview videos were stunning. They generated millions of views. They made the front page of every tech publication on the planet. Waitlists were flooded. Investors salivated.
But a demo is not a product. A product needs:
Sora had none of these at scale. The gap between "wow, look at this AI video" and "I will pay $20/month to use this daily" turned out to be a chasm.
This is the same trap that killed AI wrappers in the 2023-2025 boom. Companies layered a chatbot interface on GPT-4's API, got early traction from the novelty factor, then watched users churn once the magic wore off. ChatGPT-wrapped products demo well but rarely become mission-critical, are easy to replicate, and impossible to differentiate.
Here's a detail most articles gloss over: OpenAI is preparing for an IPO.
Target valuation: between $830 billion and $1 trillion. Expected timeline: late 2026 or early 2027.
Analysts actually viewed the Sora shutdown positively for IPO prospects. Wall Street read it as "disciplined capital allocation" --- corporate-speak for "they stopped lighting money on fire."
Sam Altman framed the pivot as reallocating resources toward "automated researchers and companies". Translation: the GPUs that were rendering AI slop are now powering coding assistants and enterprise tools that actually generate revenue.
And here's what's telling: while OpenAI was pouring compute into video generation, Anthropic's Claude Code was capturing the enterprise developer market --- the segment where customers pay real money for measurable productivity gains. Anthropic hit $19 billion annualized revenue with 80% from enterprise clients.
The Sora shutdown was a business decision dressed up as a strategic pivot. OpenAI needed to show IPO investors that it could prioritize revenue-generating products over technically impressive demos.
After watching Sora, Humane, Rabbit, and dozens of AI startups burn through billions, I think the rules are pretty clear.
Rule 1: Compute cost per interaction must be below your revenue per interaction. This sounds obvious. It isn't. Sora's cost per clip ($1.30) was higher than what most users were willing to pay. Scale makes bad unit economics worse, not better. If you lose money on every user, getting more users just accelerates your death.
Rule 2: Retention beats virality. Sora had 1 million downloads in five days and 1% 30-day retention. A product with 10,000 downloads and 40% retention is a better business. Every time. The App Store chart position is vanity. Monthly active users paying monthly is revenue.
Rule 3: Pick a lane. Runway serves professionals. Kling serves cost-conscious creators. Google subsidizes through its ecosystem. Pika serves casual creators. Sora tried to serve everyone and served no one. The "platform" approach only works if you can afford to sustain it for years. Sora couldn't sustain it for months.
Rule 4: Don't confuse demo excitement with product-market fit. Millions of people watched Sora's demo videos. That doesn't mean millions of people have a recurring need to generate AI video. There's a huge gap between "that's cool" and "I'll pay for this every month." Sora confused the first with the second.
Rule 5: The modality tax is real. Text is cheap. Images are expensive. Video is ruinous. Audio is somewhere in between. When choosing what AI product to build, the modality you pick determines your cost floor. If that floor is above what users will pay, no amount of optimization saves you.
If you're building an AI product right now, here's how I'd think about viability:
Step 1: Calculate your cost per interaction. Not your average cost --- your marginal cost at the usage level you're targeting. For Sora, the marginal cost of a 10-second clip was $1.30. At scale, that added up to $15 million per day.
Step 2: Determine your realistic revenue per interaction. Not what you wish you could charge. What users will actually pay. Sora was bundled into a $20/month subscription. If a user generates 50 clips per month, that's $0.40 per clip in revenue against $1.30 in cost. Game over.
Step 3: Check your retention curve. If 30-day retention is below 20%, you don't have a product. You have a novelty. Novelty doesn't compound. Spending money to acquire novelty-seekers is burning cash.
Step 4: Identify your subsidy model (or don't). Google can subsidize Veo because it drives ad revenue. Meta can subsidize Llama because it drives platform engagement. If you don't have a secondary revenue stream to absorb AI compute costs, your product has to be profitable on its own. Most aren't.
Step 5: Stress-test at 10x usage. If your current usage is affordable, ask what happens at 10x. Sora's costs were manageable during the preview period with limited users. At scale, they became catastrophic. The math that works for 100,000 users might destroy you at 1 million.
Sora's failure was predictable, and I think a lot of people in AI knew it but didn't say it.
The moment OpenAI announced consumer-grade AI video generation at $20/month, the math didn't work. Video generation consumes 10-15x the compute of a ChatGPT conversation. You cannot bundle that into a text-focused subscription and survive. The margins aren't there. They were never there.
But here's what bothers me more: the industry learned nothing from the 2023-2024 AI wrapper bubble. The pattern is identical --- build something flashy, ride the hype cycle, ignore unit economics, pray that scale or cost reduction saves you before the money runs out. Sora was just the biggest, most expensive version of this pattern.
I think the AI industry is heading for a serious correction. Not a crash --- AI is genuinely transformative. But a correction in expectations about what can be monetized and how. The companies that survive will be the ones that solve the unglamorous problem: making AI products that cost less to run than they charge.
Right now, text-based AI tools are the only modality with proven, sustainable economics. Image generation is borderline. Video is a money pit. Audio is unproven. Multimodal is aspirational.
If I were building an AI product today, I'd start with text, stay with text, and only expand to other modalities when the cost curves drop by another order of magnitude. The $15 million per day lesson is simple: impressive technology that burns cash isn't a business. It's a demo with a timer on it.
OpenAI killed Sora because they had to. The real question is how many other AI products are running on the same timer right now --- and whether their founders are honest enough to admit it.