The race in video generation large models is rapidly accelerating into a capital-intensive phase.
On May 11th, market reports indicated that Kuaishou is planning to spin off its video generation AI model business, Kling AI, for financing at a valuation of $20 billion. This funding round aims to raise $2 billion and is currently in discussions with investors, including Tencent.
The same market sources also reported that Kling's current Annual Recurring Revenue (ARR) has reached $500 million, which has doubled since before the Lunar New Year.
According to reports, sources close to the matter have revealed that Kuaishou is indeed engaged in financing communications based on a $20 billion valuation for Kling AI. However, neither Kuaishou nor Tencent has officially commented on the matter.
Kuaishou-W (1024.HK) surged over 11% at the market open, reaching HK$57.4, its highest price since March 19th. In the U.S. market, Kuaishou's stock closed up nearly 9.85% on May 11th, trading at $1.45 per share.
Beyond Kuaishou, other AI concept stocks in the Hong Kong market also strengthened. At the time of writing, Nobi Kan surged 19.72%, and Haizhi Technology Group rose 8.72%.
It is important to note that as giants like Alibaba, ByteDance, and Kuaishou ramp up their investments in video large models, the industry is heading towards a price war.
"The terminal selling price hasn't changed, but the technology keeps iterating. The objective reality is that as competition intensifies, there is less and less room for price premiums," said a senior executive from a domestic video model company.
Founders of several AI short drama companies echoed this sentiment. "Fundamentally, as model companies continuously engage in 'training,' model costs are becoming increasingly lower," stated a founder of a well-known AI short drama company.
In 2025, Kling AI's annualized revenue reached $240 million. Source: Kuaishou Annual Report.
Video Large Models Become a Capital-Intensive Business
Objectively, the cost of iterating video large models is rising.
On one hand, there are expensive computing power costs. Media reports suggest ByteDance plans to increase its 2026 capital expenditure by approximately 25% to 200 billion yuan, a significant portion of which is earmarked for AI computing infrastructure.
Kuaishou also mentioned during its earnings call that it expects total capital expenditure for 2026 to reach around 26 billion yuan. The additional 11 billion yuan investment compared to 2025 will primarily support the computing needs for the Kling large model and other foundational models, as well as cover routine server procurement for offline data storage/processing and data/computing center construction projects.
Simultaneously, labor costs in the AI market are soaring.
For instance, ByteDance has established a separate "Doubao Stock" incentive plan for its Seed team. Additionally, ByteDance previously implemented additional option grants for core technical staff in the Seed department, tiered based on performance and rank. Employees can receive ByteDance options worth approximately 90,000, 110,000, or 130,000 yuan per month, vested monthly, with the first batch distributed over 18 consecutive months, potentially reaching a million-yuan level.
Tencent, meanwhile, is upgrading its competitive response through its "Qingyun Plan." In 2025, Tencent's "Qingyun Plan" identified large models as the key area with the largest investment, expanding recruitment by over 50% and offering more market-competitive compensation packages for top technical talent. The plan also includes a special recruitment channel for AI large models, where selected candidates can participate in the development of Tencent's Hunyuan large model.
Data released by Maimai on February 5th shows that major internet companies are posting AI-related positions on the platform covering product, operations, growth, R&D, algorithms, and more, all offering high salaries to attract talent.
For example, Yuanbao user operations and event operations positions offer annual salaries exceeding 750,000 yuan, the "Doubao AI Product Manager" position offers 600,000 yuan, and the "Doubao AI Application Engineer" position offers nearly 1 million yuan. During the same period, the "Qianwen APP User Growth Algorithm Engineer" position had a maximum annual salary of up to 1.28 million yuan, and the "Qianwen App User Growth" position offered up to 1.12 million yuan.
Coupled with the previously mentioned industry difficulty in raising prices, capital is becoming the core competitive factor in the future video model wars.
Kling's Mobile Revenue Continues to Grow
It is noteworthy that domestic video models are on the rise.
According to the Artificial Analysis text-to-video ranking, domestic models Alibaba's Happyhorse-1.0, ByteDance's Seedance2.0, and Kuaishou's Kling 3.0 occupy the top three positions.
"The fundamental technological gap between domestic models isn't large. Our current core competitiveness lies in providing deeper services for vertical video markets," said the aforementioned domestic video model executive. He also noted that his company's related business has been consistently growing.
Furthermore, with the rise of AI short dramas, the potential market for video large models is vast.
Sensor Tower data shows that in March and April 2026, Kling's mobile revenue grew month-over-month by 24% and 10% respectively, with MAU data maintaining a relatively high level above 7 million.
However, it must also be recognized that video large models still have some distance to go before truly transforming the entertainment market.
In April, Zeng Yushen, Kling AI's Global Operations Head, mentioned at a related forum during the Beijing International Film Festival that AI large models have not yet solved the issue of native 4K. "AI large models indeed have some shortcomings in achieving film/TV-grade industrial-level quality standards, but solving these problems in the future is only a matter of time," she said.
Zeng Yushen believes AI will inevitably give rise to new content forms. "It will most likely be a new type of content that is personalized, interactive, and generated in real-time, perhaps similar to an electronic version of 'Sleep No More' or 'Westworld.' AI may not necessarily change the entire film/TV capital market landscape; it might happen, but perhaps not in the near term," she stated.
Overall, whether from a technological or market competition perspective, the competition in video large models is inevitably moving towards an even more capital-intensive stage.
Comments