Spending millions of dollars daily but still failing to retain users, OpenAI cuts Sora: Disney's $1 billion partnership collapses as well

Author: Shen Chao TechFlow

Last week, OpenAI officially shut down its AI video generation app Sora. It was only about half a year since an independent app launched. According to an investigation by The Wall Street Journal, Sora’s daily operating cost was around $1 million. Its global active users fell from a peak of about 1 million to fewer than 500,000. Over the entire lifecycle, in-app purchase revenue totaled only $2.10 million. Disney previously bet on Sora with a $1 billion investment and a collaboration for character licensing, but less than an hour before the shutdown announcement, it only then found out—after which the deal immediately fell apart. OpenAI is now redeploying its computing resources toward enterprise tools and programming products to go into a possible IPO later this year with lighter overhead.

On March 24, OpenAI announced it would shut down Sora. There was no long explanation; it only posted a brief farewell on the X platform.

This AI video generation tool that once dominated tech circles has gone from dazzling launch to a quiet exit in just six months. According to the latest investigation by The Wall Street Journal, the real reason was not the data privacy disputes speculated by the outside world earlier—it was a simple arithmetic problem: Sora burns through money too quickly, and it has too few users. The cost of continuing to operate is falling behind competitors in the AI arms race.

Burning $1 million a day, with total revenue of only $2.10 million: a financial dead end for AI video

Sora’s cost structure was unsustainable from the start. According to The Wall Street Journal, Sora’s daily operating cost is about $1 million. Video generation consumes far more computing power than text. Every short video generated by a user is eating into OpenAI’s limited GPU resources.

Cantor Fitzgerald analyst Deepak Mathivanan broke down the costs in more detail: generating a 10-second video segment requires about 4 GPUs running in parallel for roughly 40 minutes, with a per-piece cost of about $1.30. That figure looks manageable when user volume is still small. But once it scales to millions of users generating multiple videos at the same time, the daily bill balloons quickly. Based on estimates by Forbes and Cantor Fitzgerald, at peak usage, Sora’s inference cost could reach about $15 million per day, which translates to roughly $5.4 billion annually.

In stark contrast is the revenue side. According to Appfigures, a mobile data analytics company, total in-app purchase revenue across Sora’s entire lifecycle was about $2.10 million. Not $2.10 million per month, not per quarter—just the total, added up from launch to shutdown.

Sora head Bill Peebles admitted as early as October 2025 on social media that Sora’s economic model was “completely unsustainable.”

Downloads plunge 66% in three months; user enthusiasm fades faster than expected

After Sora 2 launched as a standalone iOS app at the end of September 2025, early data was very promising. According to Appfigures, downloads exceeded 1 million within five days after more than 100,000 on the first day, and the pace even surpassed ChatGPT’s record from its first year. In November 2025, downloads reached a peak of about 3.33 million.

But the decline was just as swift. In December, downloads dropped 32% month over month. In January, they fell another 45% to about 1.2 million. By February 2026, downloads had decreased to about 1.13 million, a plunge of about 66% from the peak. Consumer spending fell in parallel: in January revenue dropped to about $367,000, down 32% from December’s peak of $540,000.

At the active user level, citing Similarweb data, The Wall Street Journal reported that Sora’s global user count peaked at about 1 million and then continued to fall to fewer than 500,000. Early users generated large volumes of controversial videos featuring well-known IP characters (like Mario, Pikachu, etc.), driving a wave of viral spread—but that buzz failed to translate