IPO countdown pushes OpenAI to cut Sora: Who can still win in the AI video track?

By | Rongzhong Finance

The spring breeze hasn’t warmed every corner yet, but the global AI video track has already gone through an ice-and-fire kind of dramatic turn.

On March 25, OpenAI officially announced it would stop the standalone public service entry for Sora, fully internalizing its technical capabilities and integrating them into an enterprise-level API and the ecosystem of specific partner organizations. The move was widely interpreted as “OpenAI cutting Sora.” But if you cut through the fog of emotion, this is actually a profound strategic pullback and large-scale reorganization—amid the ticking countdown to the IPO, OpenAI chose to sever a limb to survive.

Image source: Screenshot of Sora’s official X (formerly Twitter) account

Almost at the same time, across the ocean in China, ByteDance’s Seedance2.0 model launched a fierce revolution in the short drama track. It’s no longer flashy 60-second movie-like trailer previews; instead, it’s hundreds or thousands of daily-updated AI short dramas with coherent plots—and even stable, recognizable character faces—streaming into major traffic platforms at very low cost, quickly running through the business closed loop from generation to monetization.

Between this withdrawal and this advance, it signals that the AI video industry has finally said goodbye to the “fun but useless” Demo era. OpenAI’s trade-off, at its core, is the painful adjustment as AI video shifts from flashy show-and-tell to a monetizable, scalable industrial inflection point. And ByteDance Seedance2.0’s success in using short dramas as its landing scenario proves that at the key moment in the IPO sprint for large models, controllable technology, controllable costs, and a closed business loop are the industry’s only main line.

Sora’s dilemma: when cool technology hits the IPO life-and-death line

In early 2024, Sora’s sudden debut was viewed as a “iPhone moment” for video generation. But two years later, it became OpenAI’s heaviest burden on its financial statements.

Sora’s failure wasn’t because the technology wasn’t advanced enough; the problem is that it always stayed at the “toy” stage and never evolved into a tool.

Even though the videos Sora generates look impressive in a single frame, they suffer from severe “uncontrollability”: character consistency is hard to maintain over long shots, complex camera-movement instructions often cause the visuals to break down, and it can’t precisely match existing post-production workflows (such as layered rendering and Alpha channel output). As a result, Sora’s video usability is extremely low, and the vast majority of generated content can only be used to show off on social media—it can’t be directly embedded in commercial advertising or film/TV production workflows.

User data doesn’t lie; capital’s patience is limited.

In September 2025, OpenAI officially released Sora2, and within five days downloads exceeded one million. However, after the celebration comes a rapid cooldown. According to Appfigures data, Sora’s downloads fell month over month by 32% in December last year, and continued to plunge another 45% in January this year.

Silicon Valley VC firm a16z partner Olivia Moore once broke the news on social media that while Sora’s downloads are still growing, early user retention data is quite weak. In the screenshot of data she shared, Sora’s 1-day, 7-day, 30-day, and 60-day user retention rates were 10%, 2%, 1%, and 0%, respectively.

It shows that users stay in a loop of “generate—be amazed—share—forget,” lacking scenarios for sustained payment and reuse.

More deadly is the high cost. An analysis firm SemiAnalysis estimates that Sora’s average daily operating cost is around $15 million. Its annualized burn rate is about $5.4 billion, including GPU rental, electricity costs, inference costs, and more—but cumulative total revenue accounts for a very small portion of overall revenue.

This kind of model—high cost, low retention, low income—is not only a financial black hole for a company preparing to target an IPO, but also a poison pill for valuation.

2026 is a key year for OpenAI to prepare for going public.

According to CNBC, OpenAI is currently in the core preparation period ahead of its IPO. This giant with a valuation as high as $730 billion is facing a transformation inflection point from scaling up to realizing profitability. OpenAI’s Chief of Applications business, Fei Qi・Ximuo, stated clearly in a recent internal all-hands meeting that the company will “aggressively pivot to high-productivity use cases,” concentrating resources on the enterprise market and directly competing head-to-head with Anthropic’s Claude.

Against this strategic backdrop, entertainment tools aimed at consumers—such as Sora—gradually become cost burdens that drag down performance. On the same day, OpenAI also announced it would shut down its instant shopping feature and start integrating a desktop super-app that brings together the browser, ChatGPT, and code tools. A series of moves with different paths all lead to the same destination: by slimming down the business and focusing resources, it’s full speed ahead toward the IPO.

At this point, Sora’s challenges are not just about retention and costs; there are also legal and compliance hurdles.

Because since 2024, copyright lawsuits related to Sora’s training data have never stopped. A collective lawsuit filed by major players such as Disney and Universal Pictures accuses Sora of using vast amounts of copyright-protected film and television materials for training without authorization. This lawsuit not only brings enormous potential liability for damages, but also traps Sora’s commercialization process in long-term uncertainty. Ahead of the IPO, any major pending legal case is a red line for regulators and institutional investors.

Sora’s shutdown isn’t a failure of technology; it should be seen as a return to commercial logic. Under IPO pressure, the company must strip away non-core, high-investment, weakly monetizable business burdens with risks that can’t be controlled, and tell the capital markets a story about profits—not dreams.

In addition, this event also marks a fundamental shift in industry consensus about the entire AI video sector. Over the past two years, the industry has been racing to answer questions like “Who generates video that looks more like movies?” and “Who can simulate more complex physical collisions?” But now the yardstick has changed, and the market has finally realized that if a perfectly made 60-second video can’t help merchants sell products and can’t help production studios save budgets, then it’s nothing more than an expensive pile of pixel waste.

The industry’s focus has shifted from simply competing on model parameters to deepening efforts in vertical scenarios, integrating workflows, and building closed business loops.

AI video isn’t done cooling off; Chinese players focus on short dramas to connect the monetization closed loop

While giants across the ocean are struggling to deal with Sora’s business model, China’s AI video track has taken a completely different path.

Although products such as Alibaba’s Tongyi Wansiang and Kuaishou’s Keling AI have made technical breakthroughs, what truly drew industry attention was the perfect coupling between ByteDance’s Seedance2.0 and the short drama—an industry-level trillion-scale market.

Seedance2.0, launched by ByteDance, didn’t choose to challenge Hollywood-grade long films. Instead, it precisely entered the short drama market—highly distinctive in China—and successfully opened the path from technology to commercialization.

You have to know that before Seedance2.0 arrived, all AI video generation (including Sora 1.0) was essentially a “probability game.” For example, you input a piece of text, and the AI generates a video for you like opening a blind box—the better-looking it is depends entirely on luck. This is called “gacha.” But Seedance2.0’s most core revolution is that it can “understand human language,” and even more, it can “understand the visuals.” It adopts a brand-new dual-branch diffusion transformer (DiT) architecture and introduces “multimodal reference capability.”

Unlike Sora’s pursuit of extreme generality, Seedance2.0’s technical route has a strong engineering-and-verticalization flavor. Short drama production doesn’t need every frame to reach an Oscar-level standard, but it does require extremely high stability, controllability, and production efficiency.

Therefore, Seedance2.0 has undergone deep, dedicated optimization for the short drama scenario, including multimodal input and script understanding, breakthroughs in character consistency, director-level shot control, audio-visual synchronization, and lip-sync driven generation, among others.

These technical improvements may not look as flashy as Sora generating a “whale swimming in the clouds,” but they truly transform AI video from “unusable art” into “industrial materials that can be produced at scale.”

Why short dramas? It’s an inevitable result of supply-demand matching.

A report from Guanyan Shijie shows that the AI short drama industry is still in an early development stage, with plenty of room for the market to be further tapped. Based on the current pace of technological development and commercial effects, it’s expected that AI short drama supply will grow by dozens of times. It’s expected that in 2026, the user base for AI manhua/literary dramas (including AI lookalike short drama) will increase from about 120 million in 2025 to 280 million, and the market size is expected to reach 24 billion yuan.

This is a super track characterized by short cycles, high turnover, strong willingness to pay, and easy distribution.

Traditional short drama filming requires renting locations, building sets, hiring actors, and assembling a production crew. A medium-sized short drama typically costs 300,000–500,000 yuan, with a filming cycle of 7–10 days. Meanwhile, AI short dramas produced through the full Seedance2.0 workflow don’t require real locations or live actors; per-episode costs can be compressed to a few thousand yuan or even lower, and the production cycle is shortened to the hour level.

According to media reports, multiple practitioners said Seedance2.0’s usability rate is over 90%. Taking the production of a 90-minute piece as an example, the theoretical cost is about 1,800 yuan. But because 80% of the generated results need to be discarded, the actual cost is close to 10,000 yuan. Seedance2.0 compresses actual costs to around a little over 2,000 yuan, saving about four-fifths.

“Currently the strongest video generation model on the ground—it marks the end of AIGC’s childhood.” After trying Seedance2.0, Feng Ji, CEO of Game Science and producer of “Black Myth: Wukong,” commented in this way.

With endless expansion of topics—ancient costume, fantasy/supernatural, sci-fi, end-of-the-world—those topics that require huge amounts of visual effects in live-action filming only need a few lines of code in AI video. Seedance2.0 enables creators of small and medium scale to handle large worlds, greatly enriching the content ecosystem of short dramas.

Short dramas have traits of high tolerance for image quality, fast requirements for pacing, and extremely sensitive cost control—making them the perfect vessel to absorb AI video production capacity. Here, AI is no longer just decorative icing on the cake; it becomes the core engine for cost reduction and efficiency gains.

Unlike Sora’s vague monetization model, Seedance2.0 has already run through a clear monetization closed loop. Its cooperation strategy shows a clear layered structure: the innermost layer is ByteDance’s suite of products—Meng, Doubao, and Xiaoying AI enjoy full capabilities and priority update rights. Tomato Novel’s IP reserves and Hongguo short drama distribution channels form an internal industrial chain, with a very short conversion path. The next outward layer is deep bundling with leading content companies. For example, after the ManMan platform under iQIYI’s subsidiary (Paoman) integrates Seedance2.0, the per-episode production cost for manhua dramas drops from 2,000 yuan to 200 yuan, producing over 10,000 episodes per month. Tencent Literature Group has also announced it will integrate the Seedance2.0 model, open up licensing for top IP such as “Joy of Life” and “Ghost Blowing Out the Lamp,” and explore an AI derivative creation plus official revenue-sharing model.

Unlike OpenAI’s reliance on future expectations to support a valuation in the hundreds of billions of dollars, ByteDance has proven with Seedance2.0 short dramas that large models can not only tell stories, but also make money. This kind of verifiable profit narrative is especially valuable in the current capital winter.

Conclusion

Sora’s shutdown is a necessary “de-bubbling” action for AI video technology. It signals the end of that rough-and-ready era where you could raise tens of millions just by showcasing an impressive video and achieve a valuation beyond 100 million.

History is always surprisingly similar. In the early days of the internet, many flashy tech-demo portals went bankrupt; only platforms that solved real needs—information retrieval, e-commerce transactions, social connections—eventually won out. Now, AI video is going through the same kind of cleansing.

From Sora to Seedance, we see not only the strategic divergence between two companies, but also a profound industry-wide shift from technology-driven to industry-driven. Sora’s retreat tells us that technology without a business closed loop is just castles in the air. Seedance2.0’s rise, in contrast, proves that only by putting technology into concrete industrial scenarios, solving real pain points, and creating tangible value can AI video evolve from a “fun toy” into an indispensable tool.

Endless information and precise interpretation—available in the Sina Finance app