OpenAI has a graveyard problem.
Six months after launching Sora, the AI video generator that let users drop their faces into fantastical scenes, OpenAI killed it. The official reason: compute had to be redirected to "the coding race." The actual reason: the math never worked. At its peak, Sora was burning up to $15 million per day in compute costs, according to Forbes estimates, against lifetime in-app revenue of roughly $2 million. By the time of the shutdown announcement, the burn rate had fallen to roughly $1 million per day, per the Wall Street Journal and TechCrunch, as user numbers had collapsed. The two estimates do not reconcile cleanly: Forbes put Sora's peak daily burn at $15 million while TechCrunch's reporting on the same shutdown cited roughly $1 million per day — a fifteenfold difference that credible analysts have not resolved. What is not in dispute is that either figure is disqualifying for a consumer product generating essentially no meaningful revenue.
The Sora shutdown is the most visible entry in what is becoming a recurring feature at OpenAI: the product that ships, the revenue that does not follow, and the deal that falls apart. As Forbes reported April 1, the company has now accumulated a list of discontinued products and abandoned partnerships that reads less like a startup and more like a corporate autopsy. The pattern matters, because OpenAI is simultaneously completing the largest private funding round in history: $122 billion at an $852 billion post-money valuation, per Bloomberg. Both facts are true at the same time.
The Disney problem
Disney found out Sora was dead less than an hour before the public announcement. This is notable because Disney had committed $1 billion to a three-year deal with OpenAI, licensing 200 characters from Marvel, Pixar, and Star Wars for use in Sora. The agreement, announced in December 2025, was framed as a watershed moment for tech-industry relations with Hollywood. Reuters reported that no money had actually changed hands before the deal was cancelled. Disney's spokesperson said the company respects OpenAI's decision. No money. No farewell. Just an hour's notice and a shelving of plans to let Sora users render Mickey Mouse or Yoda.
The Sensor Tower numbers tell the rest of the story. Over the same period that ChatGPT generated $2.48 billion in net in-app revenue per TechCrunch, Sora generated roughly $2 million, according to Appfigures. Analyst estimates vary: Sensor Tower put the figure closer to $1.4 million, while Appfigures estimated $2.14 million from 11.7 million downloads. The spread itself is the data point — when the high and low estimates of your lifetime revenue are separated by less than $1 million, the product was never real. Text generation is cheap to run and easy to charge $20 per month for. Video generation is expensive to run and consumers expected it to be free. The unit economics of AI video have never converged, and OpenAI's willingness to put a dollar figure on that failure is unusual. Most companies do not publish their money pit.
Forrester analyst Thomas Husson told the BBC that Sora was "a resource black hole" with limited monetization. Henry Ajder, an AI and deepfakes researcher, said the burn was cash OpenAI "could not afford to continue burning as initial interest has faded." The question those quotes raise, which OpenAI has not answered publicly, is whether Sora was a genuine strategic bet that failed or a demo that management knew would never pay off but needed to exist to keep the narrative going.
The capital structure behind the valuation
The $122 billion funding round announced in late March provides the context for that question. The headline number is real, but it is layered with contingencies that matter. Amazon committed $50 billion to the round, of which $35 billion is contingent on OpenAI completing a public offering or reaching artificial general intelligence. Those are not minor conditions. An IPO by the end of 2026 is the market's current best guess for the trigger; AGI is defined differently by almost everyone who uses the term. As MLQ.ai reported, Nvidia and SoftBank each put in $30 billion. The three companies account for $110 billion of the $122 billion total.
The Nvidia number carries an additional complication. Nvidia's latest annual report, filed at the end of February, included language that would not appear in a confirmed deal announcement: "There is no assurance that we will enter into an investment and partnership agreement with OpenAI or that a transaction will be completed." Nvidia has committed up to $100 billion, per earlier reporting, but the 10-K walks it back to something less firm. That gap — between what was reported and what was actually committed — is the kind of detail that makes the ceiling number harder to trust.
This is the pattern with large AI funding announcements. The headline aggregates a mix of closed cash, contingent tranches, GPU credits, and future drawdowns. Amazon's structure is the most explicit example: $35 billion of its $50 billion only arrives if OpenAI hits a milestone that OpenAI controls or the market grants. The contingent capital is a governance tool as much as a financing mechanism. It keeps OpenAI accountable to a timeline without forcing the investor to fund a company that walks away from its commitments.
The graveyard, in order
The Sora chapter is the most recent. But the list runs longer. The Oracle partnership, which would have expanded the Stargate data center footprint, ended in early March. OpenAI's Pentagon work has been a source of ongoing friction, with the government demanding use cases the company was unwilling to guarantee. Each of these is a thread; OpenAI has not published a full accounting of what it started and did not finish.
What is notable is that the company is raising capital at a scale that presupposes it will not need to raise again for a long time, while simultaneously writing off products that burn through cash at rates that would be disqualifying for any company without a $122 billion cushion. The implication is that compute allocation is now a governance decision at OpenAI: what gets funded and what gets killed is a function of which product is closest to generating revenue, not which one the company promised partners it would build.
Disney learned this in the most public way possible. Less than an hour before the announcement, a $1 billion commitment evaporated. No money changed hands, which means the practical damage is limited. But the message sent to every other partner with a term sheet on OpenAI's desk is unambiguous: the compute is finite, the timeline is short, and the things that look like commitments today are subject to revision tomorrow.
