Image source: Visual China
By Rao Fuying | Edited by Ye Jinyan
Presented by Shenzhen Insights · Tencent News Xiaoman Studio
On March 24, OpenAI announced on the overseas social platform X the shutdown of the Sora application. OpenAI CEO Sam Altman stated that, in addition to discontinuing the consumer version, OpenAI will also halt development of the developer version of Sora and will not support video features in ChatGPT.
Sora was first released in September 2025 and quickly gained widespread attention, hailed by the industry as the "GPT moment for video generation." However, just six months later, the product was discontinued. The stark contrast between its rise and rapid decline raises a critical question for the industry: Why did OpenAI shut down a product that was once so highly anticipated?
Why is OpenAI shutting down Sora?
Technically, Sora has yet to cross the threshold from "impressive" to "practical." Song Chunyu, Senior Partner at Lenovo Venture Capital, remarked: "In my view, the technical framework for text-to-video has not been fully realized yet; OpenAI’s Sora is essentially a demo, and the model still requires significant breakthroughs."
AI video entrepreneur Xiao Shi also confirmed this assessment: "Its generation results are impressive enough, but it has never reached the standard for stable commercial use and reliable delivery. It has consistently failed to address core commercial requirements such as high controllability, reproducibility, and batch production."
Market data more directly reflects the product’s challenges. After Sora’s standalone app launched in September 2025, downloads surpassed one million within the first ten days, briefly surpassing ChatGPT in popularity. But the spotlight faded quickly—downloads dropped 32% month-over-month in December and fell another 45% in January 2026, with user spending declining steadily alongside.
More critically, user retention: Data disclosed by Olivia Moore, a partner at Silicon Valley venture capital firm a16z, shows that Sora’s 1-day, 7-day, 30-day, and 60-day user retention rates are only 10%, 2%, 1%, and 0% respectively. Although downloads once surged, users are rapidly churning.
Cost pressures became another major factor in Sora's downfall. According to media reports, its monthly computing costs reached as high as $15 million, and the massive computational resources it consumed directly undermined the computing capacity available to other core teams at OpenAI. Long-standing internal conflicts over resource allocation played a significant role in its shutdown.
Sam Altman explicitly stated in internal communications that shutting down Sora is a critical step in the company’s strategic realignment, with future core computing power, talent, and funding being fully redirected toward enterprise productivity tools.
In addition, OpenAI plans to go public in the second half of 2026, and discontinuing the loss-making Sora business will help improve financial performance and send a clear profit outlook to the capital markets.
Along with the shutdown, the three-year partnership between Sora and Disney, which involved $1 billion in investment and over 200 IP licenses, has also been terminated.
An investor commented on the event: "Products like AI video generation are more like experiments during the transitional phase of AI development. This wave of AI advancement will continue to experience bubbles and consolidation, possibly even more intense than before, because initial expectations were set too high and the impact is broader. However, this consolidation process itself is normal."
In the eyes of industry professionals, Sora’s shutdown was hardly surprising. “Just a faded influencer,” said an AI video startup founder upon hearing the news of Sora’s closure. Since Sora transitioned to a paid model in December 2025, he and many other industry practitioners had largely abandoned the product. In his view, Sora’s peak was limited to its initial launch; subsequent model updates continuously degraded in performance, with output quality steadily declining far below market expectations.
Does Sora's exit signify a turning point in the text-to-video赛道?
Just as ChatGPT spurred a wave of large language model startups in early 2023, the release of Sora in February 2024 similarly accelerated the advancement of text-to-video models.
Prior to this, PixVerse V1 by Aishi Technology, Runway Gen1, and Pika 1.0 had already been released. Since then, startups have surged forward, and major tech giants have entered the arena. In April 2024, Shengshu Technology launched Vidu 1.0, a video large model positioned as a competitor to Sora, rolling out new features multiple times. Subsequently, Kuaishou’s KeLing was released, and MiniMax’s HaiLuo Video app also went live globally…
Now, with Sora coming to a halt, the landscape of the industry is undergoing a new round of reshuffling.
The balance of competition between big tech companies and startups is tilting in one direction. Industry insiders note that big tech firms can leverage their existing business ecosystems to create closed-loop systems, using AI video as infrastructure to support core businesses such as content platforms, achieving rapid scale and ecosystem synergy. This inherent advantage is precisely what startups struggle to match, directly impacting the speed of their subsequent commercialization efforts.
The technical challenges facing all players have yet to be fundamentally resolved. From a technical standpoint, the differences among mainstream products on the market are not significant. An employee at Hai Luo AI once stated: “In terms of duration, most outputs are around five seconds, with a maximum of 20 seconds. If generation takes too long, the risk of failure increases significantly—this is a widespread issue of insufficient stability.” A clear gap remains between technological maturity and commercial reliability.
Compliance issues regarding data copyright have become another hidden barrier. Previously, MiniMax was sued by iQiyi. Currently, companies are primarily addressing this by partnering with film and television studios and video platforms.
Mei Tao, founder of Zhi Xiang Future, has stated that high-quality copyright data assets will become one of the core competencies of AI companies: “By 2028, it is reasonable to believe that large models will have consumed all existing, readily available human-generated data. In anticipation of potential data scarcity in the future, we should proactively plan and develop strategies to address this challenge.”
Opportunities at large companies are clear, but startups still have room to succeed. “Startups need to build their own unique interactions or content. No one has yet defined the video agent space, so it’s still a race of speed and innovation,” noted an investor.