OpenAI killed Sora less than six months after launching it. The official explanation was a reset. The real story is that nobody was using it and it cost $1 million per day to run.
According to the Wall Street Journal investigation published March 29, 2026, Sora's worldwide user count peaked at roughly one million after launch — then collapsed to fewer than 500,000 within weeks. Breitbart confirmed the same trajectory using Similarweb data. The download numbers tell the same story in a different language: Appfigures Intelligence data cited by Ars Technica shows Sora peaked at 3.3 million downloads in November 2025 across iOS and Google Play, falling to 1.1 million by February 2026. Across that entire download span — 11.7 million installs — the product grossed approximately $2.14 million in revenue. That is the math that made the decision easy.
Disney, which had committed $1 billion to a partnership structured around Sora, learned the product was being terminated less than one hour before the public announcement, the Journal reported. Reuters confirmed that the $1 billion transaction never closed and no money changed hands. Bob Iger had bet on Sora as a content creation engine for Disney's platforms — a bet that OpenAI cancelled before it delivered anything. Sam Altman made the call himself, sources told the Journal: kill Sora, free up the chips, refocus.
The Vanity Fair Oscar party is where the decision crystallized. Altman pulled the plug at that event, according to people familiar with the matter, ending Hollywood ambitions that had included Marvel and Pixar cleared for Sora-generated content. The deal was dead. The compute was reallocated.
But the Sora shutdown was not an isolated misfire. It was the most visible fracture in a broader unraveling of OpenAI's commercial partnerships. According to Slate, Nvidia — which had agreed in September to provide OpenAI with a significant allocation of its AI chips — stated this month it would likely not go forward with those plans. OpenAI was also pulling back on a planned 600-megawatt data center expansion with Oracle and SoftBank in Abilene, Texas. Walmart, which had agreed to integrate ChatGPT into its online shopping pilot the previous October, ditched the experiment last week after the model consistently failed to improve store sales. Figure AI, which had been collaborating with OpenAI on humanoid robotics, cut off those efforts last month, preferring to use its own models instead. The same week OpenAI was shutting down Sora, Nvidia, Oracle, SoftBank, Walmart, and Figure AI were all walking away from the same company — each for their own reasons, all in the same direction.
While OpenAI was running Sora into the ground, the competitive landscape was moving without them. Anthropic was winning the engineers and enterprises that drive revenue. Claude Code — Anthropic's coding agent product — was, as one source put it to the Journal, "eating OpenAI's lunch" among software engineers and enterprise customers. The opportunity cost of keeping Sora running was not just the compute bill. It was the attention of a leadership team that should have been focused on the product race that matters.
The episode illustrates a specific problem in the AI infrastructure buildout that goes beyond OpenAI. The compute required to run consumer AI products at scale is not theoretical — it is a real budget line that competes with other real budget lines. TechCrunch reported the app was burning through roughly $1 million every day. When a product burns $1 million a day and has fewer than 500,000 users generating $2.14 million across 11.7 million downloads, the math is simple even if the optics are bad. The companies that can run frontier AI products at costs users will actually pay are the ones that will own the infrastructure layer. Everyone else is renting from them.
OpenAI is now reorganizing around products that have actual retention. Fidji Simo, OpenAI's CEO of applications, hosted an all-hands meeting this month to inform workers they would be doing away with "side quests" and optimizing OpenAI products for productivity on the business front. The question the WSJ reporting raises is not whether the Sora shutdown was right — it almost certainly was — but why it was launched at all with that cost structure and that user trajectory. A product that peaks at one million users and bleeds to half in months, generating $2.14 million across 11.7 million downloads, is not a consumer product. It is a cost center with a press release.
The broader signal for the industry: video generation at frontier quality is still not a viable consumer product at the compute costs required. The winners in this space will be the ones who solve the economics, not the ones who win benchmark races. OpenAI found that out the hard way. Nobody else should have been surprised.
