钛媒体APP

wrote a column ·

Sora stumbles, Jimo raises prices, Alibaba enters the scene: How will the second half of the AI video battle unfold?

(The author of this article is Singularity Research Society, published by Titanium Media with authorization)

By QiDian Research Society, Author: Xing Zhi, Editor: Meng Wen

The cost of producing a 2-minute Jimo AI video rose from 5 yuan to 40 yuan, while Alibaba’s 'Happy Horse' reached the top of the global video generation list—two events that have ignited discussions about the commercialization of AI videos.

In this clash, Jimo AI’s 'slowness' represents a defensive strategy within copyright red lines, while Alibaba’s 'speed' reflects an aggressive push for e-commerce monetization.

However, the real battleground is not in the noisy public discourse at the consumer end but rather in corporate ledgers where millions of tokens are consumed daily at the business end.

Recently, the main characters in public discourse have been Jimo and a certain 'Happy Horse.'

An old member account of Jimo was traded on Xianyu for nearly 20,000 yuan. This was an older account registered before the launch of Seedance 2.0, which enjoyed a 40% discount with 15,000 points per month.

After ByteDance’s successive price hikes, these accounts turned into scarce assets, being hoarded, sold, and even traded like a form of 'quasi-futures.'

Two months ago, when OpenClaw went viral, a RMB 499 on-site installation service also appeared on Xianyu. Both instances were the market’s spontaneous pricing of 'scarce resources,' reflecting users' assessment of the tool's value.

Amid widespread discussion sparked by Dreamlike AI’s price hike, a model called 'Happy Horse' suddenly topped Artificial Analysis’ global video generation rankings, surpassing Seedance 2.0 in both text-to-video and image-to-video categories.

Alibaba soon claimed responsibility, stating that this was their own HappyHorse-1.0, still in beta testing, with an API opening soon.

In theory, with AI developing at lightning speed, people have long become desensitized to the release of various new models. So why is the discussion around Dreamlike and 'Happy Horse' still so intense? It has even spawned arbitrage businesses involving resale.

The reason lies in the fact that video models are one of the few battlegrounds in AI capabilities that can still give ordinary people an 'Aha moment.' Compared to the 'back-end upgrades' of text models, the evolution of video models is visually apparent.

A comparison image or a five-second video is enough to trigger widespread sharing.

Netizens created a live-action version of 'Slam Dunk.' Source: Bilibili @mixzoo

It is social currency. Just like how MBTI tests have remained popular for years—not because they're truly accurate, but because after taking them, you can share results, discuss with friends, and define yourself.

The reason 'Happy Horse' made such a splash is similar: those comparison videos naturally lend themselves to dissemination on social platforms. Beta test codes are hard to come by, and the sense of scarcity itself becomes another mechanism driving spread.

More crucially, it’s the overall maturity of the AI infrastructure.

In 2023, when everyone was talking about AI, they were still discussing whether ChatGPT could chat or help you write emails.

Starting in 2024, some people in e-commerce began using AI to create product main images, and the short drama industry started experimenting with AI-generated videos. Companies gradually began integrating AI into their workflows.

User acceptance is growing, and commercial applications are slowly becoming viable.

The surge in comic dramas at the beginning of the year was a turning point. No actors or filming required—one person can produce an episode in a day. Media reports revealed that ByteDance's comic drama daily ad spending surpassed 70 million yuan for the first time, exceeding live-action short dramas. Kuaishou's comic drama daily spending also broke through 15 million yuan.

This means that AI video is no longer just a 'tool,' but a 'productivity driver.'

Once integrated into production systems, demand is no longer linear but grows exponentially.

Traditional live-action short dramas often cost millions and take months to produce, while AI comic dramas have reduced production cycles by over 50% and costs to 10%-30% of traditional methods, bringing them under 100,000 yuan.

To provide a more concrete figure, the pure AI computing power cost for a 5-second high-definition video has dropped from about 3,000 yuan in traditional production to 3 yuan or even lower.

This disruptive industry change directly ignited investor enthusiasm. Since 2025, financing in the AI video sector has surged by 370%.

Goldman Sachs estimates that the global market size of the AI video generation industry will grow from approximately $3 billion in 2025 to around $29 billion by 2030, a nearly tenfold increase over five years.

Giants like ByteDance, Alibaba, and Tencent are accelerating their layouts. On the Artificial Analysis top 10 list for text-to-video models, domestic models already occupy seven spots.

List cutoff date: April 16, 2026

The shutdown of Sora has drawn significant attention from domestic players.

Keep in mind that Sora is widely recognized as the 'pioneer' in the industry. In February 2024, it amazed the world with a 60-second demo video full of astonishing details, which Sam Altman described as 'the foundation for understanding and simulating the real world,' marking an important milestone on the path to AGI.

However, constrained by various practical issues such as computational power, cost, copyright, and commercial progress, it ended up shutting down.

During this time window, domestic models like Seedance 2.0 and Keling AI have captured a large amount of global demand originally waiting for Sora.

From Singularity's perspective, Sora’s issue lies not in its capabilities but in its approach. It aimed to be a 'world simulator': water splashes had to follow fluid dynamics, and motion had to comply with real-world physics.

This was highly groundbreaking in terms of technology but commercially meant extremely high computational costs and very low conversion efficiency.

It is rumored that at its peak, OpenAI's daily operational cost for Sora reached 15 million USD, with annualized expenditures nearing 5.4 billion USD, while cumulative revenue only amounted to 2.1 million USD, creating a massive gap between income and expenses.

In contrast, domestic video AI models did not pursue 'the most realistic world' from the beginning but instead chose highly applicable scenarios like e-commerce, advertising, and short dramas.

Keling AI was initially positioned to serve B-end professional users such as e-commerce merchants, advertising companies, and short drama studios. Generating a product promotion video can save them tens of thousands in real-person shooting costs, while generating short drama content allows for rapid trial and error iterations. Each use yields quantifiable commercial returns.

Currently, Keling AI’s cost for generating video per second is approximately $0.07, and it has already achieved positive gross profit margins at the inference level.

ByteDance's strategy is 'open externally, closed-loop internally.' Seedance 2.0 provides APIs through Volcano Engine, but integrates more deeply into its own product matrix such as Jimo, CapCut, and Douyin. Computing power costs are offset by platform revenue.

Last year, Keling AI generated an annual revenue of 140 million yuan. In January this year, Keling AI had 12 million monthly active users (on par with ByteDance’s Jimo AI during the same period), and its ARR exceeded $300 million—a scale that took Runway six years to achieve, whereas Keling accomplished it in just 18 months.

Although ByteDance has not gone public and disclosed specific revenue figures, we can refer to Keling AI for context. As of March this year, Jimo AI had approximately 79.94 million monthly active users, about ten times that of Keling AI (7.8 million) during the same period.

By comparison, ByteDance’s commercial performance will only be higher. Since Jimo leads both in user scale and commercialization, why has it raised prices multiple times consecutively?

To understand Jimo's price increase, one must first return to its 'starting point.'

During the initial launch of Seedance 2.0, Jimo offered substantial subsidies to existing users: a 40% discount and 15,000 points per month. This is a classic 'low-price traffic acquisition and user base expansion' strategy: bring users in first, establish habitual usage, and defer monetization for later stages.

This approach is no stranger to the internet industry. The subsidized pricing during this phase was never intended to be sustainable but rather a form of market investment.

The computational power consumption of video generation models is a real challenge faced by the entire industry. Sora's fate serves as a reference, and ByteDance clearly does not want to repeat the same mistakes.

On February 7, the day Seedance 2.0 was launched, the official announcement clearly stated that the API service was expected to go live on Volcano Ark in the second half of February, indicating that ByteDance had been planning for a commercial closed-loop from the start, rather than indefinitely subsidizing operations.

The reason the enterprise-level API took two months to open (actual opening date was April 2) lies in copyright compliance and content security.

Previously, during the limited internal testing of Seedance 2.0, Tim from Filmstorm released an evaluation video, pointing out two discoveries that he found 'terrifying': 'audio recognition from images' and 'scene reconstruction.' The former refers to generating highly similar sounds without audio input, while the latter involves providing a front-facing photo but being able to complete the backside of a building. These capabilities are breakthroughs at the product level, but pose enormous risks when applied to the real world.

Feng Ji, the producer of the game 'Black Myth: Wukong,' also posted during this time, referring to Seedance 2.0 as the 'strongest video generation model currently on Earth' and reminding everyone to be wary of 'the proliferation of fake videos and the trust crisis.'

Later, AI-generated 'Stephen Chow' and other celebrity lowbrow videos spread widely online, pushing AI copyright issues to the forefront of public attention.

As a result, ByteDance decided to 'hit the brakes': suspending human-like capabilities, delaying the launch, and strengthening compliance.

However, before compliance was completed, the consumer end had already taken off with full force.

Comic companies were using it, AI short drama practitioners were using it, content creators were using it, daily consumption kept rising, and the pressure on computing power and servers continued to grow. Reports indicated that during peak times, there were queues of 80,000 people waiting for over 12 hours.

In this situation, price hikes actually served two purposes: regulating supply and demand through market mechanisms, allowing users willing to pay higher prices to receive better services; and transitioning from previous subsidy investments to sustainable commercial pricing—a necessary step for internet products moving from 0 to 1 towards scaling up.

Currently, even after multiple rounds of price adjustments, Seedance 2.0 remains one of the most cost-effective options in the market. In a comparative review by Bilibili blogger Cappuccino Ultra, the video production cost for Dream was approximately $0.048 per second, still placing it among the top tier of similar products.

The key to determining whether ByteDance's strategy is effective does not lie in whether it is 'expensive,' but rather in how much money it can help users save.

In practical applications, this calculation is not difficult. Visual effects supervisor Yao Qi previously used Seedance 2.0 to produce a two-minute sci-fi short film titled 'Return Journey,' which included special effects, multi-character interactions, and complex camera movements, with a total cost of 330.6 yuan.

Juri Lu, a leading AIGC creative platform, has reduced the production cost of premium AI live-action series to about 1,000 yuan per minute; Kunlun Wanwei’s overseas AI comic drama, produced at a cost of less than $20,000, managed to generate revenue on the order of millions of dollars.

After the launch of Seedance 2.0, the production cost per minute for comic dramas further dropped from tens of thousands of yuan to just thousands, representing a 90% reduction. The production cycle for a comic drama also shortened from one month to about a week.

When AI video capabilities can reduce costs by an order of magnitude, users no longer dwell on whether 'each invocation is cheap,' but instead calculate whether the overall production cost is still lower.

This is why, even though Seedance 2.0 is priced two to three times higher than its competitors, demand from the enterprise side remains strong. For comic drama companies that consume millions of tokens daily, as long as there is no alternative solution with equivalent capabilities, the cost of migration far exceeds the rise in price itself.

On the other hand, individual users and small-to-medium teams operating on tight budgets are more sensitive. They may churn, adopt a wait-and-see attitude, or shift to alternative products, signaling the beginning of market segmentation.

As AI video becomes part of the production infrastructure, pricing is no longer merely a product strategy but a mechanism for resource allocation. This raises a more pragmatic question: Who will consume tokens, and who will pay for them?

Two days ago, ByteDance officially opened the API services for its Seedance 2.0 series, allowing both enterprise and individual users to access it. The service was also simultaneously launched in overseas markets under BytePlus.

This time, ByteDance has established a more comprehensive set of standards for portrait and copyright safety, covering all modalities involved in video generation and the entire process from pre-creation to post-production. Volcano Engine also provides a series of supporting capabilities for safe creation, catering to both enterprises and individual users.

Users can complete facial recognition and portrait authorization via the Volcano Ark console. The platform comes with over 10,000 high-quality virtual avatars, covering different ages and professional characteristics, balancing efficiency and compliance.

ByteDance's move is easily reminiscent of a counterattack against Alibaba’s 'Happy Horse.' However, this is actually just ByteDance finally delivering what was expected, albeit two months late.

Moreover, long before Alibaba's 'Happy Horse' gained popularity, ByteDance had already opened applications for enterprise-level APIs, removing the previous daunting multimillion-dollar minimum requirement that discouraged small and medium-sized businesses, replacing it with pay-per-use based on actual consumption. Now, it’s just an official opening to a wider audience.

The real variable isn’t whether Seedance 2.0 has opened its API, but rather that new changes have emerged in the B-end market landscape.

The breakout success of Happy Horse in the C-end market, along with the discussions sparked by Dream Studio's pricing adjustments, has created opportunities for other companies to enter the market.

An Alibaba Cloud sales representative mentioned in an interview with The Interface: “Many customers are currently frustrated with Volcano Engine’s monopolistic dominance—one, they have to queue for generation, and two, there’s been a price hike. If there were another company with similar capabilities and not exorbitantly priced, many customers would be very interested.”

In fact, looking purely at the product side, 'Happy Horse' and ByteDance’s Dream Studio aren't entirely competing in the same space.

ByteDance’s core users are manga-drama companies, AI short film production teams, and professional content creators who require multi-camera storytelling, consistent character portrayal, and stable industrial-grade output capabilities.

Happy Horse originated from Taotian Group's Future Life Lab, supported by Alibaba Mama. Its most natural application is in e-commerce, helping merchants turn product images into 15-second videos with contextual scenes, and generating marketing materials in different languages, with different characters, and for various scenarios across different markets.

One is the infrastructure for content production, the other is a tool for generating product content, but at their core, both compete for the same thing: Token consumption.

According to research reports by CITIC Securities, AI comic-dramas consume between 500,000 to millions of Tokens per minute, with single productions requiring over 100 million Tokens. Conservative estimates suggest that a comic-drama company consuming millions daily could incur annual API costs of tens of millions of yuan, representing an extremely high-quality and stable revenue source for cloud providers.

What ByteDance truly needs to focus on is that once 'Happy Horse' opens its API and offers competitive pricing, small and medium-sized studios entrenched in ByteDance’s ecosystem may gradually migrate.

The rise of Happy Horse is not just about rankings; it also breaks the temporary monopoly held by Volcano Engine. This benefits enterprise users as heightened competition means more choices.

However, 'Happy Horse' cannot bypass the same hurdle.

Once it scales up, it will face similar challenges such as computational power shortages and content safety issues like copyright and portrait rights. ByteDance has already spent significant time and resources addressing these, and new entrants will similarly need to pay this tuition.

In this light, the buzz on the consumer side is merely a smokescreen. The real competition lies in the efficiency of computational power consumption and the ownership of Tokens, a battle for resources among cloud vendors.

And this battle has only just begun.

48K Views