I Did the Math on Sora. AI Video Is a Money Furnace.
OpenAI killed its most hyped product since ChatGPT. The reason has nothing to do with product quality. It’s structural economics.
If you run a restaurant, and each plate costs you $65 to make but you charge $20, you don’t have a marketing problem. You don’t have a chef problem. You have a physics problem. No amount of buzz, no celebrity endorsement, no billion-dollar partnership changes the arithmetic. Every new customer accelerates your bankruptcy.
That’s Sora.
On March 24, 2026, OpenAI announced it was shutting down Sora, its AI video generation app, six months after launching the standalone app and three months after signing a $1 billion deal with Disney. The app will go dark on April 26. The API shuts down on September 24.
I’ve been tracking the inference economics of generative AI for a while now. Sora is the first case where the numbers are so clean, so unambiguous, that there’s almost nothing to debate. This is a story about economic physics. And every other AI video startup is subject to the same physics.
Here’s what I found.
TL;DR
$15M/day in peak compute costs (analyst estimate) vs. $2.1M total lifetime revenue. Even the conservative estimate (~$1M/day) makes the math catastrophic. (Forbes, Appfigures via VentureBeat)
1% day-30 retention vs. TikTok’s 32%. Users generated videos, watched them once, and never came back. (Olivia Moore, a16z)
160x cost ratio: generating a 10-second AI video costs roughly 160 times more than generating an equivalent amount of text. This is structural. (Derived from Cantor Fitzgerald and OpenAI API pricing)
Disney learned about the shutdown less than one hour before you did. A $1 billion deal, never finalized, killed by a phone call. (WSJ via Slate)
No AI video company has proven net profitability. Not Runway (-$155M EBITDA). Not Pika ($7.6M revenue on $80M raised). Kling’s $240M ARR is the closest, and even they haven’t published profit data.
The real question isn’t why Sora failed. It’s whether AI video at consumer prices is structurally impossible in 2026.
I. $15M/Day Is Not a Typo
In November 2025, Bill Peebles, head of Sora at OpenAI, posted on X what might be the most honest sentence ever written by a tech executive about his own product:
“The economics are currently completely unsustainable.”
He was right. But the actual numbers are worse than “unsustainable” suggests.
Deepak Mathivanan, an analyst at Cantor Fitzgerald, estimated Sora’s peak inference cost at roughly $15 million per day. The calculation: each 10-second video costs about $1.30 in GPU compute (approximately 40 minutes of total GPU time across 4 H100s in parallel). Assume 25% of Sora’s 4.5 million cumulative users publish 10 videos daily, and you get 11.3 million videos per day. At $1.30 each, that’s $14.7 million. AJ Kourabi of SemiAnalysis validated the methodology.
I need to flag a critical nuance here. The $15M/day figure is the theoretical compute ceiling. The WSJ reported a more conservative ~$1M/day net burn rate, which accounts for throttling, caching, and quality degradation that OpenAI applied to control costs. Both numbers are probably correct at their respective levels of analysis: $15M is what full utilization would cost; $1M is what OpenAI actually spent after aggressively limiting the product.
Growth was the enemy, not the solution.
Either way, the revenue tells the same story. Appfigures data shows $2.1 million in total lifetime mobile app revenue. That’s not monthly. That’s total. Six months. All platforms. The app revenue number doesn’t include ChatGPT Plus/Pro subscriptions attributable to Sora or API revenue, so it’s a floor, not a ceiling. But the gap between even $1M/day in costs and $2.1M in total revenue is so large that no reasonable adjustment closes it.
A $20/month subscriber generating 50 videos cost OpenAI $65 in compute. The more people used the product, the faster it bled money. Growth was the enemy, not the solution.
II. Why AI Video Costs 160x More Than Text
This is the part that matters beyond Sora. The cost ratio is not an OpenAI problem. It’s a modality problem.
When the model looks at a 10-second video at 720p, it doesn’t see “a video.” It sees roughly 80,000 small patches of pixels, spread across space and time. For reference, a typical ChatGPT prompt is a few hundred tokens. The model has to figure out how all 80,000 patches relate to each other, and the compute needed grows with the square of the patch count. Double the resolution, quadruple the cost.
On top of that, the model doesn’t generate the video in one pass. It starts from noise (like TV static) and refines it into a coherent video over 20-50 passes, processing all frames simultaneously to keep motion and lighting consistent. Attention operations alone eat over 85% of inference time.
The comparison to images is instructive. Midjourney has been profitable since its first month, reached $500M in revenue in 2025 with 40 employees and zero venture capital. An image costs $0.03-0.05 to generate. Users happily pay $10-30/month for hundreds of images. The unit economics work because the cost per output is low enough to absorb into a subscription.
Video can’t do this. The cost per output is 20-40x higher than images. Users don’t perceive 20-40x more value from a 10-second clip versus a high-quality image. And nobody has figured out how to bridge that gap with current hardware.
III. 1% Retention: The Product That People Used Once
Beyond the economics, there’s a product problem that the economics made unsolvable.
Olivia Moore at a16z published the retention data that tells the story in a single number. Sora’s day-30 retention was approximately 1%. For context, TikTok’s day-30 retention sits around 32%.
OpenAI tried to build a social video feed (vertical scrolling, TikTok-style) to drive retention. The idea was that if people consumed AI-generated videos the way they consume TikTok content, engagement would stick. It didn’t work. TechCrunch’s Amanda Silberling described the feed as “more AI slop than AI magic.” The WSJ used the same phrase.
The app hit #1 on the US App Store (Photo & Video category) at launch in October 2025. Four months later, it was dead. The novelty wore off. People generated a few clips, shared them, and moved on. There was no recurring use case compelling enough to justify coming back.
And with each returning user costing OpenAI $1.30 per clip, the retention problem and the cost problem fed into each other. A sticky product would have bankrupted OpenAI faster.
IV. Disney Found Out One Hour Before You Did
This is the part of the story that I keep coming back to.
In December 2025, OpenAI and Disney signed a $1 billion deal. Disney would invest $1 billion in equity. OpenAI would get a three-year license to over 200 Disney, Marvel, Pixar, and Star Wars characters. It was the first time Disney had licensed IP to an AI platform. The deal was announced with fanfare.
Three months later, Disney’s technical team learned about the shutdown less than one hour before the public announcement. According to the WSJ via Slate, the money was never transferred. The deal was signed but never finalized.
Disney’s statement was diplomatic in the way that only Disney can be diplomatic when furious:
“We respect OpenAI’s decision to exit the video generation business... We will continue to engage with AI platforms while responsibly embracing new technologies that respect IP and the rights of creators.” (NBC News)
“We respect OpenAI’s decision” is corporate for “we found out from Twitter.”
V. The Creepiest App on Your Phone
There’s a part of the Sora story that the economics angle doesn’t fully capture, and it’s worth a section because it contributed to the product’s collapse.
The Anti-Defamation League tested AI video apps in October 2025 and found they responded to over 40% of racist and antisemitic prompts. Sora actually performed best among them, refusing 60% of such prompts. That “best” score is still catastrophic.
Deepfakes of Sam Altman, Robin Williams, Martin Luther King Jr., Michael Jackson, and Mr. Rogers proliferated on the platform. OpenAI restricted generation of public figures only after complaints from families and SAG-AFTRA.
Zelda Williams, Robin Williams’s daughter, posted on Instagram in October 2025:
“Please, just stop sending me AI videos of Dad.”
“You’re making disgusting, over-processed hotdogs out of the lives of human beings.”
Bernice King, Martin Luther King Jr.’s daughter:
“For me, many of the AI depictions never rose to the level of free speech. They were foolishness.”
Sora’s watermarking was bypassed within days by third-party tools. The moderation was reactive instead of proactive. The product became, in TechCrunch’s words, “the creepiest app on your phone.”
The deepfake problem didn’t kill Sora. The economics killed Sora. But the deepfakes ensured that nobody mourned it.
VI. Is Any AI Video Company Actually Profitable?
This is the question I wanted to answer before writing this piece. If Sora’s failure is structural, then competitors should show the same symptoms. They do.
Kling AI is the only player approaching possible profitability, with $240M ARR and 60M+ creators according to Kuaishou’s investor relations. But revenue is not profit. They haven’t published cost data or net margins.
Runway is the most transparent about its problems: -$155M EBITDA in 2024 on approximately $44M in recognized revenue (depending on the estimate). They’re burning faster than Sora relative to revenue.
Elon Musk posted on March 25: “The next @Grok Imagine release will be epic. We are doubling down.” He also claimed positive gross margins for Imagine. Nobody outside xAI can verify this.
No AI video company has published data proving net profitability. Not one.
VII. The Path to $0.01 Per Clip
If the problem is structural, is there a structural solution? Maybe. But not today.
Two technologies could compress video generation costs by the 100-300x needed to make consumer pricing viable:
At $0.01 per clip, a $20/month subscription with 100 videos gives you a 50% gross margin. That works. That’s Midjourney territory.
But there’s a catch (there’s always a catch). Inference costs for AI have fallen roughly 92% since 2023. GPT-4-class models went from ~$20/M tokens to $0.40/M tokens, a 280x decline in three years. And total inference spending still rose 320% in 2025, because cheaper inference means more usage (Jevons paradox). Inference is now 55% of all AI infrastructure spending, up from 33% in 2023.
For video, the Jevons paradox will only kick in once costs fall below ~$0.01/clip. At $1.30/clip, even a 10x reduction leaves you at $0.13, still too expensive for casual consumer use. The threshold is 100-300x away. That’s 18 months of hardware and algorithmic progress, and progress doesn’t arrive on schedule.
VIII. “Side Quests”: The Pre-IPO Cleanup
The shutdown timing is not accidental. Not in the conspiracy sense. In the corporate strategy sense.
On the same day OpenAI announced Sora’s death, it also announced $10 billion in additional funding. The total round, which closed on March 31 at $122 billion, values OpenAI at $852 billion post-money. Amazon contributed $50B (of which $35B is conditioned on IPO or AGI). NVIDIA put in $30B. SoftBank another $30B. The IPO filing is expected H2 2026, with a listing in Q4 2026 or Q1 2027.
Fidji Simo, CEO of OpenAI’s Apps division, told staff in a mid-March all-hands reported by the WSJ:
“We cannot miss this moment because we are distracted by side quests.”
The Graveyard
She wasn’t only talking about Sora. OpenAI killed or is killing a string of products:
The Instant Checkout is a revealing comparison. OpenAI had promised 1M+ Shopify merchants. The actual number was approximately 30 (per Forrester via CNBC). Conversions ran at one-third the rate of Walmart’s own website.
Sam Altman’s internal memo framed the recentering:
“We want to focus on capital raising, supply chain management, and building data centers at an unprecedented scale.”
Wall Street analysts praised the move as “disciplined capital allocation”, exactly what public market investors want to see ahead of an IPO. Wall Street rewards focus. Sora was noise.
The Anthropic Factor
And there’s an angle that doesn’t get enough attention. Fidji Simo explicitly cited Anthropic as the competitive threat. Anthropic’s annualized revenue reached $19 billion, narrowing the gap with OpenAI’s $25 billion. Claude Code and Claude Cowork are making Anthropic the go-to provider for enterprise. OpenAI’s recentering on coding, agents, and the “superapp” desktop experience is a direct response to Anthropic, not just to Sora’s bleeding.
The Robotics Pivot (I’ll Believe It When I See It)
OpenAI’s official narrative for the Sora team is that they’re pivoting to “world simulation for robotics.” The spokesperson told NBC: “The Sora research team continues to focus on world simulation research to advance robotics.” That’s plausible on paper. Sora’s DiT (Diffusion Transformer) architecture does learn genuine physical properties. But no robotics product has been announced, NVIDIA is three years ahead with Cosmos/Isaac in simulation, and OpenAI’s next big model, “Spud,” is focused on coding and agents, not robotics. The GPUs freed up by Sora’s shutdown are going straight to Spud.
The GPUs are more valuable running Spud than rendering 10-second clips of a capybara riding a skateboard.
I’ll believe the robotics pivot when I see a robotics product.
CFO Sarah Friar was more honest than the spokesperson:
“We just are facing a lack of compute. We’re having to make those really difficult decisions. Often we hold back models, we don’t release features. And this was an example of having to prioritize.”
That’s the real answer. Sora was eating compute that other products needed. The GPUs are more valuable running Spud than rendering 10-second clips of a capybara riding a skateboard.
My Take
Sora didn’t fail because the product was bad. The videos were impressive. Sora failed because AI video generation in 2026 is economically impossible at consumer prices.
A 10-second clip costs $1.30 to render. A ChatGPT Plus subscription is $20/month. A user who generates 16 clips has already consumed more compute than they’ve paid for. There is no pricing structure, no retention strategy, no Disney partnership that fixes this. The problem is at the level of GPU physics, not product management.
The broader context makes it worse. Goldman Sachs noted that the AI industry spent $410 billion in 2025 with “zero measurable impact on US economic growth.” OpenAI itself is projecting $14 billion in losses for 2026 and doesn’t expect to turn a profit until 2029. Sora was the most visible symptom of a systemic problem: AI companies are spending at industrial scale on products that don’t have industrial-scale economics yet.
My position: killing Sora was the right decision, made too late. The unit economics were known by October 2025 (Peebles said as much publicly). The Disney deal was signed in December, two months after the head of the product called the economics “completely unsustainable.” Either the left hand didn’t know what the right hand was doing, or the deal was signed as a PR move with no intention of seeing it through. I’m not sure which is worse.
The Limits of This Analysis
The $15M/day figure is an analyst estimate, not an OpenAI disclosure. The WSJ’s ~$1M/day is more conservative but also an estimate. Neither can be independently verified.
The $2.1M revenue figure covers mobile app revenue only (Appfigures). Total Sora-attributable revenue (including ChatGPT Plus/Pro upgrades and API usage) is probably higher, though unlikely to change the order-of-magnitude analysis.
Kling’s $240M ARR comes from Kuaishou’s investor relations, which has incentives to present the best case. No independent audit of Kling’s profitability exists.
The 18-month cost reduction timeline (distillation + Rubin) is optimistic and depends on hardware shipping on schedule and research results translating to production. Neither is guaranteed.
I cannot independently verify whether the Disney deal money was “never transferred” or merely delayed. Multiple sources (Yahoo Finance, Slate) report it as never finalized.
What This Means for You
If you’re building on AI video APIs: plan for Sora’s API shutdown on September 24. Migrate to Kling, Runway, or Veo. But understand that you’re moving from one unprofitable provider to another. Pricing is going to be volatile.
If you’re investing in AI video startups: the thesis isn’t dead, but the timeline is longer than the pitch decks suggest. Consumer-grade economics require a 100-300x cost reduction. That’s 18+ months out, assuming everything goes right. The companies that survive will be those with the lowest burn rate, not the most impressive demos.
If you’re at an AI company deciding whether to launch a video product: Sora is the cautionary tale. The technology is impressive. The economics are not. Wait for Rubin-class hardware (late 2026/2027) before committing serious compute to video generation.
If you’re a user with content on Sora: download everything before April 26. After that date, all data is permanently deleted. And don’t delete your Sora account directly; it’s tied to your ChatGPT and API access.
FAQ
Q: Couldn’t OpenAI have just raised prices?
At $1.30 per 10-second clip, you’d need to charge roughly $50-100 per video to break even. Consumer willingness to pay for a 10-second AI-generated clip is approximately zero when free alternatives exist and when the output is, in most cases, a novelty. The market won’t support pricing that matches the costs.
Q: Isn’t this the same as YouTube being unprofitable for years?
No. YouTube’s cost structure has near-zero marginal cost per additional video (storage + bandwidth). Economies of scale work in YouTube’s favor: each new user adds minimal cost. Sora’s cost structure is the opposite. Each new video requires a full GPU inference pass. There are no economies of scale. More users means proportionally more cost.
Q: Won’t cheaper chips fix this?
Eventually. NVIDIA’s Rubin GPUs (late 2026) should reduce cost per token by roughly 10x. Combined with algorithmic improvements like the DOLLAR distillation technique (278x speedup for diffusion steps), $0.01 per clip is plausible by late 2027. That’s 18 months away. Any AI video startup that can’t survive 18 months of negative margins will not be around to benefit from cheaper hardware.
Q: What about Kling? They seem to be doing fine.
Kling has the best revenue numbers in the space ($240M ARR) and benefits from lower compute costs in China. But revenue is not profit. Kuaishou has not published Kling’s cost structure, gross margins, or net profitability. The $240M figure is encouraging but insufficient to declare the problem solved.
Q: Isn’t OpenAI just refocusing for the IPO?
Yes, and that’s the right reading. Wall Street praised it as “disciplined capital allocation.” Killing Sora, the Checkout, Atlas, and other “side quests” is textbook pre-IPO cleanup. The $852B valuation depends on showing a path to profitability, and every money-losing side product undermines that narrative. This doesn’t contradict the economic physics argument; it reinforces it. Sora was killed because it was unprofitable AND because its unprofitability made the IPO story harder to tell.
Conclusion: The Restaurant That Couldn’t Charge Enough
Remember the restaurant. Every plate costs $65 to make. You charge $20. The food is excellent. The reviews are glowing. People line up at the door. And every single customer is a net loss.
That’s what happened to Sora. The technology worked. The product was genuinely impressive. And the economics made it impossible. Not hard. Not challenging. Impossible, given current hardware and algorithmic costs.
The 160x cost ratio between video and text is not a product design problem. It’s not a go-to-market problem. It’s physics. The amount of compute required to maintain temporal coherence across frames, to run dozens of denoising passes on high-resolution 3D tensors, is orders of magnitude beyond what text generation requires. Until hardware and algorithms close that gap by 100-300x, AI video generation at consumer prices will remain a money furnace.
What is certain: OpenAI is better off redirecting those GPUs to Spud, to coding agents, to the products that actually generate revenue at positive margins. Killing Sora was the right call.
What remains open: is this a temporary problem (18 months of hardware progress) or a structural one (video generation will always be too expensive for casual consumer use)? I lean toward temporary, but “temporary” means 18 months of survival on negative margins, and most startups don’t have 18 months of runway.
Justin Patterson, analyst at KeyBanc Capital Markets, summarized it cleanly:
“We are not overly surprised... Even with all of OpenAI’s resources, Sora could not attract and retain an engaged audience.”
In five years, we’ll look back at March 2026 as the month the AI industry learned that impressive technology and viable business are two different things. The question is how many more restaurants will open before the rest of the industry reads the menu.
If this analysis was useful, share it with one person who works in AI and still thinks video generation “just needs scale.”
If you’re not subscribed yet, this is the kind of thing I publish every week. No hype, no predictions without receipts.
I also published a narrative version of this story on Medium.
✦ DELANOE PIRARD ✦
Artificial Intelligence Researcher & Engineer
🌐 delanoe-pirard.com
💻 github.com/Aedelon
💼 linkedin.com/in/delanoe-pirard
𝕏 x.com/0xAedelon
👉 This article did help you ? Clap + Follow for the next one.
Sources
Tier 1: Original Reporting
Forbes (11/10/2025). OpenAI Spending $15M/Day on Sora.
WSJ (03/2026). The Sudden Fall of OpenAI’s Most Hyped Product Since ChatGPT.
CNBC (03/24/2026). OpenAI Shutters Short-Form Video App Sora.
CNBC (03/24/2026). OpenAI Secures Extra $10B.
NBC News (03/24/2026). OpenAI Shuttering Sora.
Bloomberg (03/25/2026). Musk’s xAI Doubling Down on AI Videos.
Bloomberg (03/31/2026). OpenAI Valued at $852B.
Tier 2: Analysis & Data
TechCrunch (03/24/2026). Sora Was the Creepiest App on Your Phone.
TechCrunch (03/29/2026). Why OpenAI Really Shut Down Sora.
TechCrunch (03/29/2026). Sora’s Shutdown Could Be a Reality Check Moment.
Slate (03/2026). AI Bubble Popping.
Variety (03/2026). Disney Deal.
Deadline (03/2026). Disney Investment.
Yahoo Finance. 2 Charts Explain Why OpenAI Pulled the Plug.
HPCwire (03/26/2026). OpenAI Shifts Strategy Ahead of IPO.
Tier 3: Industry & Community
Kuaishou IR. Kling AI $240M ARR.
Sacra. Pika Revenue.
Appfigures (via Forbes, VentureBeat). App analytics.
Bill Peebles. X post on unsustainable economics.
Olivia Moore (a16z). Retention data.
NPR (03/25/2026). Deepfake Concerns.
Euronews. Zelda Williams Quote.
Fortune. Bernice King Quote.
The Information. OpenAI Revenue, Losses.
WinBuzzer. Anthropic Closes Gap.
CoinDesk (04/01/2026). OpenAI $122B Round.
Introl Blog. Video Generation Infrastructure.
AI Unfiltered. Jevons Paradox and Inference Costs.
OpenAI Help Center. Sora Discontinuation.