Sora Lost $1M Per Day: Disney Pulled $1B AI Video Deal

The Numbers Behind the Sora Shutdown

The headline figure, reported by multiple outlets following the shutdown announcement, is that Sora was losing approximately $1 million per day in compute costs alone. But the full picture is worse than that number suggests.

According to analysis published after the shutdown, Sora's total lifetime in-app purchase revenue reached approximately $2.1 million across its entire six-month existence. Monthly consumer spending peaked near $540,000 in December 2025 before declining to roughly $367,000 by February 2026. Against daily infrastructure costs that some reports placed as high as $15 million at peak usage, the revenue-to-cost ratio was catastrophic.

The financial trajectory was clear well before the shutdown announcement. Sora's lead researcher, Bill Peebles, described the project's financial trajectory as “unsustainable” in the weeks preceding the decision. For a company targeting an IPO, carrying a product that burns cash at this rate with declining user engagement was a liability that needed resolution.

Compute Economics: The Per-Clip Breakdown

To understand why Sora's economics could not work, you need to look at what each individual video generation actually cost. According to analyst Deepak Mathivanan of Cantor Fitzgerald, a standard 10-second Sora clip required approximately 40 minutes of total GPU time, using 8 to 10 minutes across four H100 GPUs running in parallel. At GPU rental rates near $2 per hour, each clip cost roughly $1.30 to produce.

That $1.30 figure is for a single 10-second standard-resolution clip. Higher resolutions and longer durations scaled costs dramatically. The per-generation cost spectrum reportedly ranged from $0.02 for a 480p/5-second clip to $3.00 or more for 4K content at 20 seconds. Professional users gravitating toward higher-quality output were the most expensive to serve.

Compare this to text generation. A ChatGPT response costs fractions of a cent in compute. Video diffusion models are orders of magnitude more expensive per interaction. A single 10-second Sora generation consumed GPU resources equivalent to thousands of ChatGPT queries. This is the structural problem: you can amortize text model inference costs across billions of low-cost interactions. Video inference has no equivalent path to cheap unit economics.

The pricing model compounded the problem. Sora offered a credit system within ChatGPT Plus and Pro subscriptions rather than per-second output pricing. Users paying $20 or $200 per month could generate video content whose compute cost far exceeded their subscription revenue. Competitors like Runway and Kling moved to per-second pricing models specifically to avoid this kind of subsidy trap.

User Growth and Collapse: 3.3M to Under 500K

Sora's launch generated exactly the kind of viral attention OpenAI excels at creating. Monthly downloads peaked at approximately 3.33 million in November 2025 across iOS and Google Play combined. The novelty factor drove enormous initial interest — generating video from text prompts was still a relatively new consumer experience.

The decline was steep. December installs fell 32% month-over-month. January suffered a further 45% decline, landing at about 1.2 million installs. By February 2026, downloads had fallen to just over 1.1 million — a 66% decline in three months from the November peak. The global active user count followed a similar trajectory, peaking at roughly 1 million before falling to under 500,000 by the time of the shutdown.

Several factors drove the collapse. Generation times of 3 to 8 minutes per 10-second clip were already uncompetitive by late 2025, when Kling was producing equivalent-quality clips in under 90 seconds. The credit-based system limited how many videos users could generate, restricting the iterative experimentation that professional content creators need. And the competition was not standing still — Runway, Kling, and Pika all shipped significant quality and speed improvements during Sora's six-month window.

The total download figure across Sora's lifetime reached 9.6 million across iOS and Android. That is not a small number in absolute terms, but the conversion to paying, retained users was minimal. Total consumer spending across all platforms reached just $1.4 million — suggesting that the vast majority of users tried Sora once or twice and never returned. Understanding these retention dynamics is critical for any business evaluating analytics and user engagement patterns.

The Disney $1B Deal Collapse

The Disney partnership was supposed to be the validation that Sora had crossed from a novelty toy into a legitimate creative platform. As reported by Variety, Deadline, and The Hollywood Reporter, the three-year licensing agreement would have allowed Sora users to generate video content featuring more than 200 masked, animated, or creature characters from Disney, Marvel, Pixar, and Star Wars properties. Disney had committed to a $1 billion stake in OpenAI as part of the arrangement.

The deal collapsed the moment the shutdown was announced. According to multiple reports, Disney learned about Sora's discontinuation less than one hour before the public announcement on March 24, 2026. No formal agreement had been finalized. No money had changed hands. The $1 billion investment was dead before it was born.

The communication failure is striking. A $1 billion deal partner learning about a product shutdown less than an hour before the public suggests the decision was made rapidly at the highest levels of OpenAI, without the partner engagement that enterprise relationships typically require. It also raises questions about due diligence on both sides. Disney committed to a billion-dollar deal with a product that had been live for only three months and had no proven retention metrics.

For enterprises evaluating AI partnerships, the Disney-OpenAI collapse is a cautionary data point. The speed at which AI products can go from flagship to discontinued means that partnership terms need exit protections that traditional technology deals may not include. Our analysis of AI product failures in 2026 covers the broader pattern of premature enterprise commitments to unproven AI products.

Why Consumer AI Video Economics Failed

Sora's failure was not a failure of the underlying technology. The model was, by most assessments, the most capable general-purpose video generation system available at launch. The failure was structural: the economics of serving diffusion-based video generation to consumer subscribers at fixed monthly prices cannot work at the current state of inference costs.

The comparison to text models is instructive. ChatGPT reached 100 million users and sustained engagement because text generation has immediate, repeatable utility — drafting emails, writing code, answering questions. Video generation is inherently more episodic. Most users do not need AI video daily. The use cases that do require frequent generation, like social media content production, gravitated toward cheaper, faster competitors with per-clip pricing.

This is not to say AI video generation has no future. It means that the viable business model is almost certainly enterprise-focused and usage-based — charging per second of output, per API call, or per seat in a professional workflow tool. The consumer subscription model was the wrong commercial frame for a high-cost, low-frequency product. Businesses developing content marketing strategies should evaluate AI video tools on per-unit cost rather than subscription bundling.

The Robotics Pivot: From Video to World Simulation

OpenAI's official statement framed the shutdown not as a failure but as a reallocation. The Sora research team has been integrated into a new “AGI Deployment” division focused on what OpenAI calls “world simulation research to advance robotics that will help people solve real-world, physical tasks.”

The technical rationale is sound. Sora was never primarily a video generation model in the way most consumers understood it. The model was trained to understand and simulate physical reality — how objects move through space, how light behaves, how physics governs motion. That understanding of spatial relationships and physical dynamics is precisely what robotics systems need to navigate and interact with the real world.

The pivot also preserves the research team and their accumulated expertise. Rather than disbanding a group that spent years building one of the most sophisticated video understanding systems in existence, OpenAI redirected their work toward applications with better unit economics. The Sora 2 model will reportedly remain available within ChatGPT's paid tiers, suggesting that video generation is being demoted from a standalone product to a feature — which may be the correct product positioning all along.

Lessons for AI Product Strategy

The Sora shutdown offers several concrete lessons for companies building, investing in, or deploying AI products. These apply whether you are a startup building a new AI tool or an enterprise evaluating which platforms to integrate into your workflows.

What Comes Next for AI Video Generation

Sora's shutdown did not kill the AI video generation market. It restructured the competitive landscape and clarified which business models work. Runway, Kling, and Veo all reportedly saw increased API signups in the week following the announcement. The market demand for AI video is real — the problem was Sora's commercial model, not the technology category.

For marketers and content teams evaluating AI video, the practical takeaway is to choose platforms with usage-based pricing and proven retention. Our comparison of post-Sora AI video generators covers the specific capabilities and pricing of each platform that emerged from Sora's collapse.

The shutdown timeline provides a concrete window for affected users and developers. The consumer app closes April 26, 2026. The API shuts down September 24, 2026. OpenAI has urged all users to download their content before these deadlines, as data will be permanently deleted. Teams with active Sora API integrations have six months to migrate to alternative providers — a reasonable timeline, but one that requires starting the evaluation process now.