The Concorde of AI Video: Why OpenAI Shut Down Sora and What Every Business Should Learn

Law Smith

Founder & President, Tocobaga

Law Smith is the Founder and President of Tocobaga, a Tampa-based ROI-focused marketing agency and SMB advisory. He has supported venture-backed and growth-stage companies with brand development, go-to-market strategy, and demand generation programs scaled to match their runway, drawing on 30+ business plans and 600+ integrated campaigns to help portfolio companies build the marketing infrastructure that compounds with each round. Tocobaga is a Google Partner and Squarespace Gold Partner headquartered in Ybor City.

In May 1976, two zoologists at the University of Oxford sent the journal Nature a short paper about animal parents that walk out on their mates. One author was Richard Dawkins, whose first book, The Selfish Gene, Oxford University Press published that same year.* The other was a colleague listed on the paper as T. R. Carlisle. Nature received it on May 12, accepted it 12 days later and published it on July 1 under a title only a biologist could love: "Parental investment, mate desertion and a fallacy."†

The question was old and a little ruthless. When two parents raise young together, which one leaves? The biologist Robert Trivers had built an influential theory of "parental investment" in a 1972 book chapter.* Dawkins and Carlisle admired how cleverly he used it, then wrote that "one of his most appealing usages of it is fallacious."†

The appealing usage sounded like common sense: the parent that has already poured more into the eggs has more to lose, so it should be the one that stays. Dawkins and Carlisle argued that the logic runs backward. What is spent is spent. A parent, like an investor, should weigh what staying will return from here, not what it has already paid.

The mistake needed a name, and the news supplied one. On January 21, 1976, British Airways and Air France had flown Concorde's first paying passengers, from London and Paris on the same day, at about 1,350 miles per hour.‡ Airlines had held 80 options from 18 carriers when the jet first flew in 1969, and by early 1973 most had canceled.§ It entered service anyway. Biologists came to call the error the Concorde fallacy.*

The jet earned the label. Twenty were built, and only 14 entered airline service, all with Air France and British Airways.‖ Each transatlantic seat burned about a ton of fuel, the average round trip cost about $12,000, and the Smithsonian's curator of air transportation writes that the jets often flew half full and that by 1989 the airlines realized Concorde "could never pay for itself."¶ It flew for 14 more years.

Nearly 50 years after that first paying flight, a different marvel lived the same story at software speed. On September 30, 2025, OpenAI released Sora 2 and a free social video app built around it, calling the model what "may be the GPT-3.5 moment for video."** On March 24, 2026, the company posted a five-word farewell: "We're saying goodbye to Sora."†† The app went dark on April 26, and the developer API was set to follow on September 24.‡‡ Concorde carried paying passengers for nearly 28 years. Sora's app lasted 208 days.

That arc matters far beyond Silicon Valley. Tocobaga, the Tampa Bay lead generation agency and SMB advisory behind SolvingHow.com, produces its own campaign imagery and video with several AI models side by side, and it advises founders, investors and operators whose growth plans now quietly assume a particular AI tool will still exist next quarter. Sora is the cleanest case study yet of three collisions at once: a dazzling product against its unit economics, a side project against the main bet, and every customer who built on top of it against the fine print of renting.

Why did OpenAI shut down Sora?

OpenAI shut down Sora because the AI video app cost far more to run than it earned, its audience was shrinking, and the company wanted the computing power for coding and enterprise products it judged more valuable. Reporting by The Wall Street Journal, as summarized by TechCrunch, put the app's burn at roughly $1 million a day while worldwide users fell from a peak near 1 million to fewer than 500,000.§§

OpenAI never published a one-line reason. Its help center lists dates, export instructions and refund details, and little else.‡‡ The motive came from people near the decision. NBC News, citing The Wall Street Journal, reported that top executives said the company was sharpening its focus and could not do everything at once.†† Earlier that month, according to Agence France-Presse, Fidji Simo, OpenAI's applications chief, had told employees they could not be distracted by "side quests."‖‖

Run the reported numbers and the decision looks less like a surprise than an overdue invoice. A million dollars a day is about $365 million a year. Spread across fewer than 500,000 users, it comes to more than $2 per user per day, or roughly $60 a month. A ChatGPT Plus subscription, which included limited Sora access, cost $20 a month.¶¶ On those figures, the average remaining Sora user consumed about three months of a Plus subscription every month.

Three forces arrived together:

• Cost. Generating a video clip takes minutes and far more processing power than text or images.¶¶

• Demand. Downloads and consumer spending had been falling for months before the announcement.***

• Opportunity. The Batch, the newsletter published by DeepLearning.AI, reported that OpenAI had already diverted Sora's processing resources to a new model code-named Spud, which powers coding and enterprise products, before it announced the shutdown.¶¶

TechCrunch's Connie Loizos boiled the Journal's findings down to a phrase no product manager wants on a tombstone: "a money pit that nobody was using."§§ That is harsh, and a little unfair to the people who did use it. It is also the Concorde in miniature. A technically dazzling product carried costs that scaled with every free user and revenue that did not, inside a company that needed the same scarce resource for a bigger bet.

Even the leftovers were rerouted. The help center says purchased credits can still be spent in Codex, OpenAI's coding agent.‡‡ Video money became code money. That is the whole strategy in one refund policy.

What was Sora, and why did it take off so fast?

Sora was OpenAI's text-to-video model, and Sora 2, released September 30, 2025, added synchronized dialogue and sound effects plus an invite-only iOS social app where people could insert themselves and their friends into generated scenes.** The app reached No. 1 on the U.S. App Store in October 2025.***

The model was a genuine leap. OpenAI's launch post noted that a missed basketball shot now rebounds off the backboard instead of teleporting into the hoop, and that the model could follow multi-shot instructions while keeping the world consistent from scene to scene.** The original Sora, OpenAI wrote, had been the GPT-1 moment for video. This was supposed to be the GPT-3.5 moment, the one where the public notices.**

The public noticed. TechCrunch reported, using Appfigures data, that the iOS app topped 100,000 installs on its first day even with invite codes and reached 1 million downloads faster than ChatGPT did.*** The "characters" feature, which let users drop themselves or willing friends into any scene, turned every group chat into a casting call.**

Then came the detail that matters most in hindsight. OpenAI launched Sora 2 free, "with generous limits to start," and said access remained "subject to compute constraints."** Its only stated monetization plan was to eventually let people pay for an extra video when demand outran available compute.** The company also said it was not optimizing the feed for time spent. Admirable. Also a product with no obvious way to make money from attention, attached to a very expensive way to lose it.

In December 2025, the prestige partner arrived. The Walt Disney Company agreed to make a $1 billion equity investment in OpenAI, receive warrants for more and license more than 200 animated, masked and creature characters from Disney, Marvel, Pixar and Star Wars for Sora under a three-year deal, with curated fan videos planned for Disney+.††† Concorde had 80 airline options. Sora had Disney. Both looked like proof of demand. Neither was the same thing as enough paying customers.

The table below lays out the verified Sora timeline, from launch to the API's removal date, and what each step meant for the business.

Verified Sora timeline, from launch to the API's removal date, and what each step meant for the business.

Date	What happened	What it meant
September 30, 2025	OpenAI releases Sora 2 and an invite-only iOS app, free "with generous limits to start"	Growth came first and pricing later, on a product where every clip burned costly GPU time
October 2025	App reaches No. 1 on the U.S. App Store after 100,000 first-day installs	Demand was real. So was the compute bill
October 30, 2025	Sora head Bill Peebles calls the economics "completely unsustainable"	The team named the problem within a month of launch
November 2025	Forbes estimates peak costs near $15 million a day, about $1.30 per 10-second clip	Outside math showed costs scaling with free usage while the API charged $1 per clip
December 2025	Disney agrees to a $1 billion equity investment and licenses 200+ characters for three years	The marquee partner arrived just as downloads started falling
December 2025 to January 2026	Downloads fall 32%, then 45%; monthly consumer spending peaks at $540,000	Lifetime consumer spending worked out to about 15 cents per download
March 24, 2026	OpenAI announces it will discontinue the Sora app and API; the Disney deal does not move forward	About six months from launch to exit; Disney had less than an hour of notice
April 23, 2026	GPT-5.5, code-named Spud, launches for coding, computer use and multi-step work	The compute went where paying enterprise customers were
April 26, 2026	Sora app and web experience shut down	Users had to export; data is deleted after the final export window
September 24, 2026	Scheduled removal of the Videos API and all Sora 2 models	No replacement listed; developers and automations had to migrate on their own

How much did Sora cost OpenAI to run?

No outside party knows the exact figure, but two credible estimates bracket it: The Wall Street Journal reported losses of roughly $1 million a day near the end, and a Forbes analysis in November 2025 estimated that peak usage could cost close to $15 million a day, or more than $5 billion a year.§§,‡‡‡

The Forbes math, by Phoebe Liu, is a template every founder should borrow. Deepak Mathivanan, an analyst at Cantor Fitzgerald, estimated that a 10-second clip took about 40 minutes of total GPU time, roughly 8 to 10 minutes on four chips running in parallel, at rental rates just under $2 an hour. That put the cost near $1.30 per clip.‡‡‡ AJ Kourabi of SemiAnalysis estimated that a quarter of roughly 4.5 million users posted about 10 videos a day. That is about 11.3 million videos daily. Multiply by $1.30 and the meter reads about $14.7 million a day, or roughly $5.4 billion a year.‡‡‡

The two estimates sit about 15 times apart, and both can be right. The Forbes figure modeled the fall 2025 frenzy. The Journal's figure described a shrunken audience months later. The spread is the lesson: Sora's costs moved with usage, while its revenue barely moved at all.

Then there is the price tag. OpenAI's API charged $1 for a 10-second Sora 2 video and $3 for Sora 2 Pro.‡‡‡ If the $1.30 estimate was close, the standard API price sat about 23 percent below the estimated cost of making the clip. Thirty cents lost on every clip, before anyone paid for research, salaries or storage. Estimates can be wrong, and OpenAI never confirmed its costs. But when the outside math and the inside behavior point in the same direction, the outside math deserves attention.

The inside behavior was not subtle. On October 30, 2025, a month after launch, Bill Peebles, OpenAI's head of Sora, wrote: "The economics are currently completely unsustainable."‡‡‡ NBC News reported that shortly after launch he had already announced limits on video generation because chip supply was tight.††

The revenue side was thin. Appfigures counted 9.6 million lifetime downloads and $1.4 million in lifetime consumer spending through January 2026, $1.1 million of it from the United States.*** That is about 15 cents per download. At the Journal's $1 million a day, the app's entire lifetime of in-app spending covered about 34 hours of losses. At the Forbes peak estimate, it covered about two and a quarter hours. One fairness note belongs here: The Batch reported that most of Sora's revenue came from paid ChatGPT plan subscribers, and OpenAI never disclosed how much of a Plus or Pro subscription should be credited to video.¶¶ Even generous credit would struggle to close a gap that size.

Concorde had its own version of the meter. Each seat burned about a ton of fuel across the Atlantic, and the jet's fuel burn ran about twice that of other airliners.¶,‡ With the cabin half full, each paying passenger effectively carried the fuel for two seats.¶ The $12,000 fare still was not enough.

Did people actually stop using Sora?

Yes. Appfigures data reported by TechCrunch showed Sora downloads falling 32 percent in December 2025 and another 45 percent in January 2026, to 1.2 million, while monthly consumer spending slid from a December peak of $540,000 to $367,000.*** By the time of the shutdown, The Wall Street Journal reported, worldwide users had fallen from about 1 million to fewer than 500,000.§§

Work the download figures backward and the slide gets steeper. If January's 1.2 million followed a 45 percent drop, December was about 2.2 million. If December followed a 32 percent drop, November was about 3.2 million. That is a decline of roughly 63 percent in two months.

TechCrunch's Sarah Perez listed the culprits in January. Google's Gemini and its Nano Banana image model were pulling attention away. Meta's Vibes feed offered a rival stream of AI video. Copyright restrictions tightened after studio backlash, which drained the novelty of famous characters. Many users did not want strangers using their likeness. And the Disney deal had not yet moved installs or spending.***

Quality competition arrived too. By early April 2026, The Batch noted, Sora 2 Pro ranked 19th on Artificial Analysis's text-to-video leaderboard, behind models from ByteDance, Kling, xAI and Google.¶¶ Six months earlier, its app had been No. 1 on the U.S. App Store.***

The pattern is familiar to anyone who has launched a consumer product: a novelty curve that looks like a hockey stick for six weeks and a ski slope after that. Trial is not retention. A product people try once to turn a friend into a medieval knight is not a product people pay for in March. Concorde's route map told the same story at a slower pace. Service to Washington, Rio de Janeiro, Caracas, Miami and Singapore was cut until only New York remained.¶

What is the Concorde fallacy, and did OpenAI fall for it?

The Concorde fallacy, better known in business as the sunk-cost fallacy, is the habit of continuing to invest because of what has already been spent rather than what future returns justify. With Sora, OpenAI did roughly the opposite: it ended the app about six months after launch and walked away from a Disney partnership announced about 15 weeks earlier.‖‖,†††

In 1980, Dawkins and a colleague, H. J. Brockmann, published a study of digger wasps fighting over shared nests titled "Do digger wasps commit the Concorde fallacy?" They credited the 1976 Carlisle paper with the first discussion of the biological version and defined the error plainly: "investing further in a project simply because one has invested in it heavily in the past."* They called it "a notorious theoretical error which tempts economists and evolutionists alike."* They could have added founders, boards and marketing departments.

Their alternative fits on an index card: "wise policy decisions will be based upon prospects of future returns, and only indirectly upon weight of past commitment."* Every company carries a version of the opposite habit. The annual software license already paid for. The custom integration already built. The trade show booth already designed. Each of those costs is gone whether the project continues or not, and each one quietly argues for continuing.

Concorde is the cautionary version. The governments of Britain and France absorbed its losses.‡ Airlines had concluded by 1989 that it could never pay for itself.¶ Then, on July 25, 2000, Air France Flight 4590 crashed after takeoff from Paris, killing 109 people on board and four on the ground.‡ When British Airways and Air France announced the retirement on April 10, 2003, British Airways said passenger revenue had been falling steadily while maintenance costs rose, FLYER reported that passenger numbers had never recovered from the crash, and the jet had been expected to stay in service until 2007.§§§ It took a tragedy and a revenue slide to end what the spreadsheet had ended years earlier.

OpenAI's version moved faster. According to The Batch, the Sora team had proposed training a new model to generate video inside ChatGPT, possibly replacing the standalone app, and OpenAI declined because the training cost was too high.¶¶ That is the anti-Concorde decision. A team that had built something beautiful asked for more money to make it bigger, and leadership said no on the strength of future returns.

Disney found out the way partners usually do when a company gets disciplined in a hurry. The Journal reported that Disney learned of the shutdown less than an hour before the public announcement.§§ Disney's spokesman stayed gracious: "We respect OpenAI's decision to exit the video generation business and to shift its priorities elsewhere."‖‖ A $1 billion handshake ended with a 60-minute heads-up. Discipline and courtesy are different skills.

What does the Sora shutdown reveal about OpenAI's finances?

It reveals a company growing revenue at historic speed while spending even faster on computing power, and rationing every chip. On March 31, 2026, OpenAI said it had closed a funding round with $122 billion in committed capital at an $852 billion post-money valuation and was generating $2 billion in revenue per month.‖‖‖

OpenAI reported more than 900 million weekly ChatGPT users, more than 50 million subscribers and enterprise revenue above 40 percent of the total, on track to match consumer revenue by the end of 2026. It said it was growing revenue four times faster than the companies that defined the internet and mobile eras.‖‖‖ More than $3 billion of the round came from individual investors through bank channels, and an expanded credit line of about $4.7 billion sat undrawn.‖‖‖

The cost side came from internal documents reported by The Information and summarized by The Decoder. OpenAI's 2025 revenue was $13.1 billion. Inference costs, the price of actually running the models for users, quadrupled in 2025. Adjusted gross margin fell to 33 percent from 40 percent the year before, below a 46 percent target. The company projected cash burn of $25 billion in 2026, $57 billion in 2027 and $85 billion in 2028 before turning cash-flow positive in 2030.¶¶¶

Translate the margin first. At 33 percent, every $100 of revenue leaves $33 of gross profit before salaries, research or marketing. For a company that sells software, that is a thin cushion, and video was the most expensive thing on the shelf.

Then translate the burn. Twenty-five billion dollars a year is about $68 million a day. Sora's reported $1 million a day was about 1.5 percent of that. Even the Forbes peak estimate would have been about 22 percent. So Sora did not threaten OpenAI's survival. It competed for the company's scarcest input, and it lost the argument to products with paying enterprise customers attached.

The pressure showed at the top. In early April, Business Today, citing The Information, reported that chief financial officer Sarah Friar had warned colleagues OpenAI might not be "IPO-ready" in 2026, flagged more than $600 billion pledged for cloud server capacity over five years as a risk, and that the company expected to spend upward of $200 billion before becoming cash-flow positive.**** The same report said operating chief Brad Lightcap was moving into a special-projects role that included a potential joint venture with private equity firms to distribute OpenAI's technology.**** Private equity, in other words, was being courted as a sales channel.

By late September, the top line had sprinted again. Reuters, following an Axios report, cited a source saying OpenAI's annualized revenue was approaching $70 billion, with enterprise sales more than doubling since July, and noted reports that the company was preparing to go public.†††† As of early October 2026, it remained private.

What was Spud, and why did it need Sora's chips?

Spud was OpenAI's internal code name for GPT-5.5, released on April 23, 2026, as its most capable model for coding, computer use and multi-step office work, and The Batch reported that OpenAI diverted Sora's processing resources to Spud before announcing the shutdown.‡‡‡‡,¶¶

Axios reported that GPT-5.5 could plan and use tools across multi-step workflows without step-by-step prompting, and that early testers reported saving up to 10 hours of work a week.‡‡‡‡ OpenAI co-founder Greg Brockman called it "a new class of intelligence" and added a line that doubles as Sora's eulogy: "We are moving to a compute-powered economy."‡‡‡‡

In a compute-powered economy, every GPU hour has a price and an alternative use. A 10-second video clip, by the Forbes estimate, consumed about 40 GPU minutes.‡‡‡ Those same minutes could serve coding agents and enterprise workflows that customers pay for on contract. OpenAI's Codex had more than 2 million weekly users by March 2026, up fivefold in three months, and an ads pilot reached more than $100 million in annualized revenue in under six weeks.‖‖‖ When the choice is between a consumer toy earning 15 cents per download and a coding agent compounding at that rate, the toy loses.

The enterprise push had a consulting layer, too. On February 23, 2026, OpenAI announced multi-year Frontier Alliances with BCG, McKinsey, Accenture and Capgemini to help companies deploy what it calls AI coworkers through its Frontier platform. McKinsey and BCG handle strategy, operating model redesign and change management. Accenture and Capgemini wire Frontier into enterprise systems and run it over time, working alongside OpenAI's own forward deployed engineers.§§§§ For management consultancies, that is a big new practice area. It is also a practice built on another company's roadmap, and Sora just showed how quickly a roadmap can change.

The Sora researchers were not shown the door. OpenAI moved them toward longer-term work on world models and robotics.¶¶ The talent was redeployed. The product was not.

My read is that Concorde's sponsors never faced a Spud, a clearly better use for the same money and machines waiting in the next hangar, and that absence is part of why the jet lingered. Opportunity cost is what makes killing a project rational. A company that knows exactly what it would do with the money and the chips finds it far easier to walk away from the nest.

Should anyone trust annualized run-rate revenue?

Trust it as a speedometer, not an odometer. Annualized run rate takes a recent period, often a single month, and multiplies it into a year, and Reuters noted in September 2026 that the figure can be misleading for exactly that reason.††††

OpenAI's own numbers show why it matters. Two billion dollars a month in March 2026 implies a $24 billion run rate.‖‖‖ Actual revenue for calendar 2025 was $13.1 billion, and an early-2026 internal forecast reported by The Information called for about $30 billion in calendar 2026.¶¶¶ By late September, annualized revenue was reportedly near $70 billion.†††† All of those numbers describe the same company within about a year. For a company growing as fast as OpenAI, a rising run rate is real information. For a company with churn, seasonality or a one-time spike, it is a mirage with a decimal point.

Sora is the small, perfect example. Its best month of consumer spending was December 2025, at $540,000.*** Annualize that and the app had a $6.5 million run rate. One month later spending was $367,000, a $4.4 million run rate. The run rate fell by about a third in 30 days. Anyone who valued the app on its December number paid for a peak.

The ads pilot deserves the same treatment. "More than $100 million in ARR in under six weeks" is impressive, and it means roughly $8 million to $9 million a month at the time of the announcement, not $100 million collected.‖‖‖ Both statements are true. Only one is cash.

For SaaS founders and the venture investors who fund them, the fix is boring and effective:

• Ask for trailing twelve-month revenue next to any run-rate figure.

• Ask which period was annualized, and why that period.

• Ask for revenue retention by cohort, so churn cannot hide inside growth.

• Ask for gross margin after compute or hosting costs, not before.

• Ask what share of revenue comes from pilots, promotions or one-time deals.

A board that gets those five answers can enjoy a big run-rate number without being hypnotized by it.

What should SaaS founders learn from Sora's unit economics?

Founders should calculate the cost to serve each user, especially the heaviest users, before celebrating growth, because when cost scales with usage and price does not, every new user digs the hole deeper. On the Forbes estimate, a Sora power user making 10 clips a day consumed about $13 of compute daily, more than half of a $20 monthly Plus subscription in a single day.‡‡‡,¶¶

Do that math out loud. Ten clips at $1.30 each is $13 a day. Thirty days is about $390 a month, against a $20 subscription. That user costs roughly 19.5 times what a Plus subscriber pays. In the Forbes model, a quarter of all users behaved that way.‡‡‡

Here is how the same trap shows up in an ordinary software company, using round, illustrative numbers rather than any real firm's. A SaaS company sells a $49-a-month plan and adds an AI video feature at no extra charge. Each generated clip costs the company $1 in compute. Ninety percent of subscribers make five clips a month, costing $5 each. Ten percent are power users who make 200 clips a month, costing $200 each.

• Blended compute cost per subscriber: (0.9 x $5) + (0.1 x $200) = $4.50 + $20.00 = $24.50 a month.

• That is half of the $49 price, gone, before payroll, hosting or sales.

• Each power user costs about four times what that user pays.

• If the feature's marketing works and the power-user share doubles to 20 percent, blended cost rises to $44 a month, about 90 percent of revenue.

The feature is a hit. The company is in trouble.

OpenAI saw the escape hatch on day one. Its launch plan was to let people pay for extra videos when demand exceeded compute.** The lesson for founders is to build the meter before the crowd arrives: usage caps on free tiers, credit systems for expensive features, pricing tied to whatever drives cost, and a weekly report showing the margin on the top 10 percent of users. Venture-backed companies can subsidize growth on purpose. They cannot subsidize it by accident for long.

Concorde had the inverse problem and the same result. It never had too many customers. It had too few at a price that covered a ton of fuel per seat.¶ Too many free users and too few full-fare passengers are two doors into the same room.

How should PE and VC boards handle "side quests"?

Boards should treat every side project as a fresh investment that must re-earn its capital on future returns, with kill criteria written before launch, a named alternative use for the money and a deadline for the decision. Simo's warning about "side quests," followed within weeks by the Sora shutdown, is what that discipline looks like at an $852 billion company.‖‖,‖‖‖

Private equity operating partners know this conversation by heart. A portfolio company launches a new product line, a new region or a new brand. The launch gets applause at the all-hands meeting. A year later, nobody wants to be the person who points out that the line consumes a third of management's attention and a sliver of gross profit. Venture investors see the same thing in startups that chase a second product before the first one has a repeatable sales motion.

The Concorde timeline puts a price on hesitation. Airlines concluded in 1989 that the jet could never pay for itself, and it flew until 2003.¶ Fourteen years of a known answer. OpenAI's Sora timeline ran from a public admission of "completely unsustainable" economics in October 2025 to a shutdown announcement in March 2026, a little under five months.‡‡‡,††

A practical decision rule for an operating partner, a venture board or a founder:

1. Write the kill criteria at launch. For example: if cost to serve exceeds 60 percent of revenue per active user after two quarters, or if paid conversion stays below a named threshold, the project ends or is restructured.

2. Name the alternative use. Every dollar, engineer and GPU hour on the side quest gets priced against the best thing it could be doing instead.

3. Ban sunk costs from the memo. Past spending is context, never a reason. Future returns come first.*

4. Separate applause from revenue. Press coverage, app store rankings and celebrity partners are signals of attention, not proof of a business.

5. Plan the landing. Redeploy the talent, give partners real notice and give customers time and tools to move their work.

OpenAI nailed the first three and stumbled on the fifth. Disney's one-hour notice is the part no board should copy.§§

What happens to brands and marketers who built on Sora?

They lose access on the vendor's schedule, not their own. The Sora app and website shut down on April 26, 2026, OpenAI said it would permanently delete Sora user data after any final export window, and the Videos API and every Sora 2 model reached their scheduled removal date on September 24, 2026, with no replacement listed.‡‡,‖‖‖‖

OpenAI notified developers on March 24, 2026. Its deprecation table covers the Videos API, sora-2, sora-2-pro and three dated model snapshots, and the replacement column for each shows only a dash.‖‖‖‖ Any e-commerce brand that had wired Sora into an automated ad pipeline, any agency that built a client deliverable around a Sora-specific look and any creator whose portfolio lived inside the app got the same message: export now and migrate alone.

Disney, a $1 billion partner, got less than an hour of warning.§§ A five-person direct-to-consumer brand gets a help article.

Tocobaga has repeated one line to clients for years: "You might think you own your website, but you might be renting it." The same sentence now applies to every AI workflow. Prompts typed into a vendor's interface, likenesses uploaded for cameos, generated clips sitting in a vendor's library and a process that only works inside one company's app are rented, no matter how creative the work felt. When the landlord changes the building's purpose, the tenant moves.

Marketers should notice what OpenAI kept and what it deleted. It kept the research team, the compute and the customers' purchased credits, which now buy coding time.‡‡ It deleted the consumer product and, after the export window, the consumer data. That is a rational choice for OpenAI. It is a warning for every brand that treated a free creative tool as a permanent piece of its marketing infrastructure.

How can a business keep its AI content workflow portable?

By owning every input and output, writing briefs that describe the shot rather than the tool, and routing each shot to whichever model handles it best, a practice some creators call model arbitrage. Tocobaga's own content production works this way, with several image and video models running side by side rather than one tool carrying the whole workflow.

The video model market will keep reshuffling. Google DeepMind now calls Veo 3.1 its leading video generation model, says Veo 3 generates sound effects, ambient noise and dialogue natively, and offers access through Gemini, its Flow tool and the Gemini API.¶¶¶¶ Leaderboards shift by the quarter: Sora went from the No. 1 app in the U.S. App Store to a model ranked 19th on one prominent leaderboard in about six months.***,¶¶ A workflow that assumes today's leader stays the leader has an expiration date.

A practical checklist for a marketing team, an agency or a founder:

• Keep a master prompt library in a company-owned document, written in plain language: subject, setting, camera, lens, light, motion, sound and mood. Prompts are intellectual property. Store them that way.

• Maintain a tool-independent brand kit: color values, type choices, logo files, voice rules and written likeness releases for anyone who appears on camera.

• Export every final cut and every usable raw clip to company-owned storage the day it is generated. Assume any vendor library could disappear.

• Write briefs by shot, not by platform. "Fifteen-second product reveal, golden hour, slow push-in" travels to any model. "Make it look like our Sora style" does not.

• Test at least two models for each recurring shot type every quarter, and track cost per usable second, not cost per generation. Rejected takes are part of the price.

• Keep humans in the edit. The model makes clips. People make campaigns.

• Document the exit for every automation. If a scheduled workflow calls a single video API, write down what replaces it before the vendor writes the deprecation notice.

The common mistakes are predictable: treating a free tier's price as permanent, storing the best work in a vendor's gallery, letting one employee's account become the archive and confusing a model's signature look with the brand's look. Each is cheap to fix before a shutdown and expensive after.

None of that requires an AI budget the size of a Disney deal. It requires the same habits Tocobaga's house rules already enforce for websites and ad accounts: the client owns everything, contracts never auto-renew and no single vendor gets to hold the keys. For a no-pressure intro call about building an AI content workflow a business actually owns, reach Tocobaga at (813) 934-6605 or through SolvingHow.com.

Was shutting down Sora a mistake?

Possibly in one narrow sense: OpenAI surrendered leadership in a category it helped create and burned goodwill with Disney and creators. On the evidence available, keeping Sora alive at the reported cost would have been the bigger mistake.

The case against the shutdown deserves a fair hearing. The Batch argued that OpenAI gave up its leadership in video generation and opened the field to strong competitors, and its editors wrote that the era in which an impressive AI demo is "sufficient to establish leadership may be drawing to a close."¶¶ The Disney agreement would have brought $1 billion of equity, licensed characters and a showcase on Disney+.††† Creators who built audiences on the app lost their home. And there is a real technical counterargument: computing gets cheaper. Nvidia says its newest chips cut the cost of running models like GPT-5.5 by up to 35 times per token, according to Axios.‡‡‡‡ Concorde's fuel never got 35 times cheaper. A patient OpenAI might have waited for the hardware curve to rescue the economics of video.

The case for the shutdown is the arithmetic. A product with about 15 cents of lifetime consumer spending per download, an API price below its estimated production cost and a user base cut in half was competing for chips with a coding agent whose weekly users grew fivefold in a quarter.***,‡‡‡,‖‖‖ Waiting for cheaper hardware means paying today's prices for tomorrow's hope, and that is the Concorde fallacy wearing a hoodie.

My take: right call, clumsy exit. OpenAI was correct to stop funding a side quest on the strength of past investment and a famous partner. It was careless to give Disney an hour and consumers an export button. The decision was disciplined. The landing was not.

The mistakes on the customer side are just as instructive. Brands that treated a free, invite-only app as dependable infrastructure made a bet they never priced. Agencies that sold clients a single-vendor look sold a depreciating asset. Investors who read app store rankings as revenue mistook a parade for a business.

The jet, the nest and the video app

Air France's first Concorde in passenger service, F-BVFA, logged 17,824 flight hours and now rests in the Smithsonian's collection at the Udvar-Hazy Center.¶,‡ The jet got a museum. Sora's videos got an export button and a deletion date.‡‡

Dawkins and Carlisle were writing about animal parents, not jets or apps. Their point was about the moment a parent decides whether to stay with the nest. The one that leaves is not heartless. It is the only one doing math about the future. Concorde's sponsors stayed with the nest for nearly 28 years, long after the airlines stopped believing it could pay for itself.¶ OpenAI left after 208 days, while its biggest partnership was barely 15 weeks old.

The lesson is not "never build the jet." Build it. Fly it. Love it out loud. Then decide, before the first passenger boards, the number at which the company walks away, and keep the footage on a drive the company owns. Nobody will ever wheel a dead app into a museum. The work only survives if somebody kept a copy.

Why did OpenAI shut down Sora?

OpenAI shut down Sora because the AI video app cost far more to run than it earned, and the company wanted its computing power for coding and enterprise products. The Wall Street Journal reported losses of roughly $1 million a day as worldwide users fell from about 1 million to fewer than 500,000. OpenAI announced the decision on March 24, 2026, about six months after launching the Sora 2 app.

When did the Sora app and API shut down?

OpenAI announced the shutdown on March 24, 2026. The Sora web and app experiences were discontinued on April 26, 2026. The Videos API and all Sora 2 models were scheduled for removal on September 24, 2026, and OpenAI's developer documentation listed no replacement. OpenAI said it would permanently delete Sora user data after any final export window ended.

How much did Sora cost OpenAI per day?

OpenAI never disclosed the figure. The Wall Street Journal reported losses of roughly $1 million a day near the end of the app's life. A November 2025 Forbes analysis estimated that peak usage could cost nearly $15 million a day, based on about $1.30 of computing per 10-second clip and roughly 11.3 million videos a day. The two estimates describe very different levels of usage.

What happened to Disney's $1 billion deal with OpenAI?

In December 2025, Disney agreed to make a $1 billion equity investment in OpenAI and license more than 200 characters for Sora under a three-year deal. When OpenAI discontinued Sora in March 2026, the deal did not move forward. The Wall Street Journal reported that Disney learned of the shutdown less than an hour before the public announcement, and Disney said it respected OpenAI's decision.

What is the Concorde fallacy?

The Concorde fallacy is another name for the sunk-cost fallacy: continuing to invest in a project because a lot has already been invested, rather than because future returns justify it. Oxford zoologists Richard Dawkins and T. R. Carlisle discussed its biological version in a 1976 Nature paper. The name refers to the supersonic Concorde, which kept flying for years after airlines concluded it could never pay for itself.

What was OpenAI's Spud model?

Spud was the internal code name for GPT-5.5, which OpenAI released on April 23, 2026. OpenAI positioned it for coding, computer use, office work and multi-step workflows that need less step-by-step prompting. According to The Batch from DeepLearning.AI, OpenAI diverted the computing resources that had been running Sora to Spud before it announced Sora's shutdown.

What are the best alternatives to Sora for AI video?

The best choice depends on the shot. Google DeepMind's Veo 3.1 generates video with native sound effects and dialogue and is available through Gemini, Flow and the Gemini API. Leaderboards in 2026 also ranked models from ByteDance, Kling and xAI above Sora 2 Pro. Many teams now test several models side by side and route each shot to the one that handles it best.

Is annualized run-rate revenue a reliable number?

Annualized run rate is useful as a speedometer, not a total. It multiplies a recent period, often one month, into a year, so a strong month, a seasonal spike or a short pilot can make a business look bigger than its trailing revenue. Investors should ask which period was annualized, compare it with trailing twelve-month revenue and check cohort retention and gross margin after compute costs.

Sources

*

Dawkins, Richard, and H. J. Brockmann. "Do Digger Wasps Commit the Concorde Fallacy?" Animal Behaviour 28 (1980): 892 to 896. Supports the definition of the Concorde fallacy, the credit to Dawkins and Carlisle (1976) for its first biological discussion, the "notorious theoretical error" and "prospects of future returns" quotes, Trivers's 1972 chapter on parental investment and the 1976 Oxford University Press publication of The Selfish Gene. https://www.gwern.net/doc/psychology/cognitive-bias/sunk-cost/1980-dawkins.pdf

†

Dawkins, Richard, and T. R. Carlisle. "Parental Investment, Mate Desertion and a Fallacy." Nature 262 (July 1976): 131 to 133. Supports the authors and their Oxford affiliation, the May 12 receipt, May 24 acceptance and July 1, 1976 publication dates, and the quote calling one of Trivers's usages "fallacious." https://www.nature.com/articles/262131a0

‡

Edwards, Owen. "When Concorde First Flew, It Was a Supersonic Sight to Behold." Smithsonian Magazine, January 20, 2016. Supports the January 21, 1976 first paying flights from London and Paris, the 1,350 mph cruise, fuel burn about twice other airliners, the governments absorbing the losses, the July 25, 2000 crash killing 109 aboard and four on the ground, and F-BVFA's placement at the Udvar-Hazy Center. https://www.smithsonianmag.com/smithsonian-institution/when-concorde-first-flew-it-was-supersonic-sight-behold-180957875/

§

Thomas, Geoffrey. "Fifty Three Years On, Concorde Lives in the Hearts of All Who Look Up and Dream." AirlineRatings, February 21, 2022. Supports the 80 options from 18 airlines held when Concorde first flew in March 1969 and the cancellation of most options by early 1973. https://www.airlineratings.com/articles/fifty-years-concorde-lives-hearts-look-dream

‖

Riva, Alberto. "Concorde at 50: Where Are They Now?" The Points Guy, March 2, 2019. Supports the 20 Concordes built and the 14 that entered airline service with Air France and British Airways. https://thepointsguy.com/news/concorde-at-50-where-are-they-now

¶

van der Linden, F. Robert. "Au Revoir Concorde." The Torch, Smithsonian Institution, March 4, 2019 (originally published in Air and Space magazine). Supports the roughly $12,000 average round-trip fare, about a ton of fuel per seat, flights that often ran half full, the 1989 conclusion that Concorde "could never pay for itself," the route cuts that left only New York, the 2003 retirement, and F-BVFA's 17,824 flight hours as Air France's first Concorde in service. https://torch.si.edu/2019/03/au-revoir-concorde

**

OpenAI. "Sora 2 Is Here." OpenAI, September 30, 2025. Supports the launch date, the "GPT-3.5 moment for video" framing, synchronized audio, the characters feature, physics examples, the free launch "with generous limits to start," compute constraints, the stated monetization plan and the decision not to optimize for time spent. https://openai.com/index/sora-2/

††

Kopack, Steve, Jared Perlo, and Daniel Arkin. "OpenAI Shutting Down Sora App." NBC News, republished by NBC Bay Area, March 24, 2026. Supports the "We're saying goodbye to Sora" post, the Journal's report that executives said the company was sharpening its focus, and Peebles's early limits on video generation due to chip supply. https://www.nbcbayarea.com/news/national-international/openai-shutting-down-sora-app/4057424/

‡‡

OpenAI Help Center. "What to Know" (Sora discontinuation, article 20001152). Accessed October 9, 2026. Supports the April 26, 2026 app and web discontinuation, the September 24, 2026 API end date, export instructions, permanent data deletion after any final export window, and purchased credits remaining usable in Codex. https://help.openai.com/en/articles/20001152-what-to-know

§§

Loizos, Connie. "Why OpenAI Really Shut Down Sora." TechCrunch, March 29, 2026. Supports The Wall Street Journal's findings of roughly $1 million a day in losses, worldwide users falling from about 1 million to fewer than 500,000, Disney learning of the shutdown less than an hour beforehand, and the "money pit" characterization. https://techcrunch.com/2026/03/29/why-openai-really-shut-down-sora/

‖‖

Agence France-Presse. "OpenAI Kills Sora Video App in Pivot Toward Business Tools." Gulf News, March 25, 2026. Supports Fidji Simo's "side quests" warning to staff, the shutdown about six months after launch and the Disney spokesman's statement. https://gulfnews.com/technology/companies/openai-kills-sora-video-app-in-pivot-toward-business-tools-1.500485367

¶¶

DeepLearning.AI. "OpenAI Exits Video Generation." The Batch, April 3, 2026. Supports video generation's compute intensity, ChatGPT Plus pricing and Sora access, most Sora revenue coming from paid plan subscribers, the diversion of Sora's processing resources to Spud, the rejected video-in-ChatGPT proposal, the team's move to world models and robotics, Sora 2 Pro's 19th-place leaderboard ranking and the editors' quote on AI demos. https://charonhub.deeplearning.ai/openai-announced-it-would-shut-down-sora-its-once-state-of-the-art-video-model/

***

Perez, Sarah. "OpenAI's Sora App Is Struggling After Its Stellar Launch." TechCrunch, January 29, 2026. Supports Appfigures data on 100,000 first-day installs, the No. 1 U.S. App Store ranking, 1 million downloads faster than ChatGPT, the 32 and 45 percent download declines, the $540,000 and $367,000 monthly spending figures, 9.6 million lifetime downloads, $1.4 million lifetime spending and the competitive reasons cited. https://www.techcrunch.com/2026/01/29/openais-sora-app-is-struggling-after-its-stellar-launch/

†††

The Walt Disney Company and OpenAI. "The Walt Disney Company and OpenAI Reach Landmark Agreement to Bring Beloved Characters from Across Disney's Brands to Sora." Press release, Business Wire, December 2025. Supports Disney's $1 billion equity investment, warrants, the license for more than 200 characters from Disney, Marvel, Pixar and Star Wars, the three-year term and Disney+ plans. https://www.businesswire.com/news/home/20251210820829/en/The-Walt-Disney-Company-and-OpenAI-Reach-Landmark-Agreement-to-Bring-Beloved-Characters-from-Across-Disneys-Brands-to-Sora

‡‡‡

Liu, Phoebe. "OpenAI Could Be Blowing as Much as $15 Million per Day on Silly Sora Videos." Forbes, republished by Forbes Australia, November 10, 2025. Supports the estimate of about $1.30 per 10-second clip, 40 GPU minutes per clip, roughly 4.5 million users and 11.3 million videos a day, nearly $15 million a day and over $5 billion a year, the $1 and $3 API prices, and Bill Peebles's "completely unsustainable" quote. https://www.forbes.com.au/news/innovation/openai-could-be-blowing-as-much-as-15-million-per-day-on-silly-sora-videos/

§§§

"Concorde to Retire." FLYER, April 10, 2003. Supports the joint British Airways and Air France retirement announcement, British Airways' statement on falling passenger revenue and rising maintenance costs, passenger numbers not recovering after the 2000 crash, and the earlier expectation of service until 2007. https://flyer.co.uk/concorde-to-retire/

‖‖‖

OpenAI. "Accelerating the Next Phase of AI." OpenAI, March 31, 2026. Supports the $122 billion round at an $852 billion post-money valuation, $2 billion in monthly revenue, more than 900 million weekly users and 50 million subscribers, enterprise revenue above 40 percent, more than $3 billion from individuals, the undrawn $4.7 billion credit facility, Codex growth and the ads pilot's ARR. https://openai.com/index/accelerating-the-next-phase-ai/

¶¶¶

"OpenAI Adds $111 Billion to Its Cash Burn Forecast as AI Costs Spiral Beyond Projections." The Decoder, summarizing reporting by The Information, accessed October 9, 2026. Supports 2025 revenue of $13.1 billion, the 2026 revenue forecast of about $30 billion, quadrupled inference costs, the 33 percent adjusted gross margin, and projected cash burn of $25 billion, $57 billion and $85 billion in 2026 to 2028 before positive cash flow in 2030. https://the-decoder.com/openai-adds-111-billion-to-its-cash-burn-forecast-as-ai-costs-spiral-beyond-projections/

****

"Why OpenAI IPO May Not Happen in 2026 as CFO Sarah Friar Flags Financial Risks." Business Today, April 6, 2026 (citing The Information). Supports Sarah Friar's reported "IPO-ready" concerns, the more than $600 billion in cloud capacity pledges, spending of upward of $200 billion before positive cash flow, and Brad Lightcap's special-projects role including a potential private equity joint venture. https://www.businesstoday.in/technology/story/why-openai-ipo-may-not-happen-in-2026-as-cfo-sarah-friar-flags-financial-risks-524178-2026-04-06

††††

Hu, Krystal, and Jaspreet Singh. "OpenAI's Annual Recurring Revenue Nears $70 Billion, Axios Reports." Reuters, republished by The Star, September 29, 2026. Supports annualized revenue approaching $70 billion, enterprise sales more than doubling since July, the caveat that run-rate figures can mislead, and reports of IPO preparation. https://www.thestar.com.my/tech/tech-news/2026/09/29/openai039s-annual-recurring-revenue-nears-70-billion-axios-reports

‡‡‡‡

Mills, Madison, and Ina Fried. "OpenAI Releases 'Spud' GPT Model." Axios, April 23, 2026. Supports GPT-5.5 as the model code-named Spud, its coding, computer-use and office-work focus, testers saving up to 10 hours a week, Greg Brockman's quotes and Nvidia's claim of up to 35 times lower cost per token. https://axios.com/2026/04/23/openai-releases-spud-gpt-model

§§§§

OpenAI. "Frontier Alliance Partners." OpenAI, February 23, 2026. Supports the multi-year Frontier Alliances with BCG, McKinsey, Accenture and Capgemini, each firm's role, and their work alongside OpenAI's forward deployed engineering team. https://openai.com/index/frontier-alliance-partners/

‖‖‖‖

OpenAI. "Deprecations." OpenAI API documentation, accessed October 9, 2026. Supports the March 24, 2026 developer notice, the September 24, 2026 removal date for the Videos API, sora-2, sora-2-pro and dated snapshots, and the absence of a listed replacement. https://developers.openai.com/api/docs/deprecations

¶¶¶¶

Google DeepMind. "Veo." Accessed October 9, 2026. Supports Veo 3.1 as Google DeepMind's leading video model, natively generated sound effects, ambient noise and dialogue, and availability through Gemini, Flow and the Gemini API. https://deepmind.google/models/veo/