The viral video large model suddenly shuts down! The rise and fall—what is the reason behind it?

Recently, OpenAI announced the discontinuation of the Sora standalone app, API interface, and the embedded video feature in ChatGPT, fully exiting the consumer-level AI video generation market. (News link: Sora will be shut down) Analysts pointed out that insufficient rigid demand and revenue falling short of expectations were the main reasons for Sora’s abrupt end just six months after its launch. How the AI video sector can maintain long-term vitality has become a focal point of interest.

By analyzing Sora’s download numbers, we found that within five days of its official launch in September 2025, downloads exceeded one million, quickly reaching the top of the application store’s free chart, even surpassing ChatGPT’s popularity at the same time. However, this popularity did not last long; within less than three months, downloads decreased by 32% month-over-month, and in January of this year, downloads plummeted by another 45%. By February, Sora’s downloads had slipped out of the top 100 in the free app rankings.

On social media, we saw a post stating: “In the first two weeks after its release, my family and I made over 100 videos with Sora, it sparked our true creativity and joy. But two weeks later, we never opened it again. We feel this is where Sora truly failed.”

This post resonated with many netizens, with numerous users commenting that Sora felt more like a “novel toy” and struggled to become a tool for solving real problems.

Li Piji, a professor at the School of Artificial Intelligence at Nanjing University of Aeronautics and Astronautics, stated: When a new large model is released, it attracts a large number of users eager to try it out. Although these features are rich and interesting, if they are not integrated into users’ daily workflows and do not address high-frequency, essential problems in life, software usage will see a sharp decline, leading to a fleeting phenomenon.

Experts told reporters that as software activity continues to decline, user retention rates and the rigid demand for software will also decrease, while high costs like GPU leasing, electricity expenses, and inference fees remain unchanged. According to analysis firms, Sora’s daily operating cost reaches as high as $15 million, amounting to approximately $5.4 billion annually. Each second of video generated requires rendering around 30 images, and due to poor generation quality, many videos produced can only be discarded, resulting in a usable rate of only 5%-10%. This means that for every qualified video produced, tens of times the computing power is wasted, leading to a serious imbalance between revenue and costs, making the operational model increasingly difficult to sustain.

Domestic text-to-video large models iterate quickly

Low cost and wide application

The end of Sora has raised questions for many: do similar domestic large models that focus on video generation also face the same survival dilemmas? Industry insiders pointed out that for large models, “practicality,” meaning a rich array of application scenarios, is particularly important, as deep market integration is necessary for sustained vitality.

Similar to Sora, Seedance 2.0 also exploded in popularity overnight, dominating the online landscape. Unlike Sora, it benefits from two years of industry groundwork, allowing consumers to directly experience one-stop services for video generation, editing, and social media publishing. Additionally, another model, Kegling, has maintained stable “vitality” for nearly two years, continuously updated based on user profiles, with over 30 iterations.

Shen Linlin, a professor at the School of Artificial Intelligence at Shenzhen University, mentioned: The competition among large models is fierce, requiring companies to continuously iterate and upgrade to better meet market and user needs, improving production efficiency in various niche areas and reducing corresponding costs.

From the perspective of user costs, it can be seen that generating an ideal 15-second shot with Sora requires 5-10 attempts, costing approximately 800 yuan. In contrast, using mainstream domestic text-to-video large model software, even attempting to generate 10 shots of approximately 15 seconds only costs about 150 yuan.

Li Piji, a professor at the School of Artificial Intelligence at Nanjing University of Aeronautics and Astronautics, stated: Domestic computing power, electricity, and other resources have unique advantages, being low-cost and fast. Moreover, domestic companies focus more on lightweight and practical upgrades when iterating large models, giving them a significant competitive edge in the market.

According to incomplete statistics, over 3,000 small studios are currently using text-to-video large models as core creative tools, with some studios achieving video output increases of over three times compared to before with the assistance of these models. Furthermore, an increasing number of enterprise users are emerging, showcasing the significant market advantages in the domestic short drama and short video sectors. Recently, several short dramas and corporate promotional videos completed by text-to-video large models have gone viral online.

In the first two months of this year, the National Internet Information Office added a total of 48 new registered large models, covering various application scenarios such as industrial manufacturing, cultural tourism, and healthcare.

Shen Linlin, a professor at the School of Artificial Intelligence at Shenzhen University, remarked: Large models fundamentally serve the market and applications, and more large models are beginning to root themselves in specific application scenarios, aligning with user needs, shifting from “showing off technology” to practical application. As industries and large models form a closed loop, they are also accelerating the synchronization of the innovation chain and industrial chain, creating a stable user base and business model.

(Source: CCTV Finance)