Since the start of the year, the most talked-about stories of getting rich through artificial intelligence have generally fallen into two categories: agentic AI, as represented by Manus, and AI hardware, exemplified by Plaud.ai.
But outside these headline-grabbing cases, video generation models are quietly fueling a new wave of AI startups in China.
In June, Kuaishou’s Kling AI reached an annual recurring revenue (ARR) of USD 100 million across its app and web platforms. Among startups, MiniMax’s Hailuo AI and Shengshu Technology’s Vidu have each surpassed USD 10 million in ARR through their web-only offerings.
Several insiders told 36Kr that the actual subscription revenues for these products may be even higher.
Notably, video generation has begun generating profit even before large language models (LLMs) have achieved similar results at scale. AIsphere revealed that subscription revenue from its video model, PixVerse, now covers the majority of its operating costs, bringing the company close to breakeven.
At this year’s Beijing Academy of Artificial Intelligence (BAAI) Conference, Huang Weilin, head of image and video generation at ByteDance’s Seed division, projected that top video generation products could generate USD 100 million in ARR this year, with potential to grow that figure to between USD 500 million and USD 1 billion by 2026.
Only a year ago, however, models resembling Sora were widely viewed in China as too costly to run and too uncertain in commercial payoff to be viable.
Tencent previously reported that Wang Changhu, former head of visual tech at ByteDance and founder of AIsphere, had been advised against entering the space by Zhu Xiaohu of GSR Ventures. When MiniMax launched Hailuo in September 2024 amid intense competition from Kuaishou and ByteDance, it, too, was met with skepticism.
An investor who passed on Wang recounted the mood at the time: video models were seen as unlikely to deliver timely returns, and most startups were expected to be squeezed out by dominant incumbents, as had occurred in the LLM space.
Indeed, in 2024, many Chinese video AI startups struggled with funding and failed to gain product-market fit. Luying Technology, backed by Redpoint China Ventures and Lanchi Ventures, was acquired in December that year.
Yet within a year, AIsphere’s revenue reversed investor sentiment. “I regret it to my core. Saying video generation models wouldn’t make money was a collective misjudgment by the investment community,” that same investor told 36Kr.
What has propelled these startups forward comes down to three interrelated factors.
1. Visual appeal over technical precision
Even though video generation models are still maturing, users appear more forgiving of flaws, likely because the category is guided more by aesthetic taste than strict accuracy.
“Each company’s data strategy, even slightly immature technology or biased training data, can lead to distinctive video generation styles,” one investor said. “But that’s exactly why video creation is a diverse, taste-driven market. Every model has its own audience.”
For example, many users have observed that Kling performs especially well at generating food-related and mukbang-style content, likely due to Kuaishou’s vast trove of short-form food videos.
2. Going global with cost efficiency
Western users tend to be more willing to pay for AI-driven tools and are quicker to adopt them. That dynamic has benefited Chinese video AI startups expanding abroad.
MiniMax’s Hailuo faced resistance to its subscription pricing in China but found a much larger overseas audience, eventually surpassing USD 10 million in ARR.
Cost has also worked in their favor. Constraints in capital and compute have pushed Chinese startups toward efficiency, giving them a pricing edge globally.
Hailuo and Vidu reportedly operate at just one-tenth to one-sixth the cost of generating videos of similar duration and resolution compared to Sora.
RELATED ARTICLE
3. Virality as a growth catalyst
Platforms like TikTok and YouTube have been critical to user growth.
One AIsphere employee told 36Kr that PixVerse saw a sharp uptick in late 2024 after its “Venom” special effects went viral, racking up over a billion views across TikTok, Douyin, and other platforms. Another investor noted that special effects and filters also played a key role in user acquisition for Pika and Hailuo.
“For model companies today, climbing tech benchmarks isn’t enough anymore,” the investor said. “You need to proactively find or even invent demand that can scale.”
In video creation, users want more than just productivity tools. They want reach, visibility, and incentives. AI-generated viral effects tap directly into that demand.
There’s some validation in the numbers. According to a January ranking by Andreessen Horowitz, Hailuo AI ranked 12th globally in user traffic, surpassing OpenAI’s Sora and Kuaishou’s Kling, which ranked 23rd and 20th, respectively.
This suggests the video generation space is still fluid, offering room for startups to carve out a presence, unlike the more consolidated LLM field.
Wang once told 36Kr that video model development is somewhere between the GPT-2 and GPT-3 stage. That leaves both technical challenges and time for new contenders to emerge.
A narrowing window and rising stakes
Still, as more players adopt similar self-sustaining strategies, the first-mover advantage is waning. Existing leaders are under pressure to keep momentum.
Back in March 2024, Wang warned that those who didn’t move early would struggle: “If you didn’t raise enough money, attract users, or build a strong team and tech base during the early phase, you won’t have the resources to stay in the game going forward.”
An AI-focused investor shared a similar view. She told 36Kr that while the market isn’t yet saturated, few investors are interested in newcomers “unless someone emerges as a DeepSeek-style dark horse.”
She added that video AI startups typically attract significantly less funding than LLM-focused companies. “Over time, the resource gap will become more glaring,” she said. “Kling and Dreamina both hold a technical edge thanks to their pace of iteration.”
For companies already in the space, the message is clear: speed up, or risk being left behind.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Zhou Xinyu for 36Kr.