China’s tech titans tussle in AI video gold rush

(By Caixin journalists Guan Cong and Wang Xintong)

A costly chatbot subsidy war dominated China’s artificial intelligence (AI) industry in February. AI agents were all the rage in March. By April, the spotlight had shifted again — this time to AI-generated video, where competition has intensified across nearly every layer of the tech and content ecosystem.

What sets this phase apart is not just technological progress, but commercialisation. In China, AI video is emerging as one of the few generative applications with a viable way to generate revenue, supported by strong user demand and a vast content market.

That combination is drawing companies across the value chain, from model developers and content platforms to game studios and advertisers. Yet the rapid expansion is also bringing mounting challenges, including copyright disputes, content safety risks and tightening regulatory scrutiny.

Still, this phase has global implications, serving as a test case for whether generative video can become a sustainable industry. Stronger resistance in Western markets may limit adoption there, potentially strengthening China’s position in the broader AI race.

The competition in China’s AI video space has become ferocious in just a few months. Since January, nearly ten video generation models have been released worldwide, most of them in China.

Long seen as a leader in China’s AI video field, Kuaishou Technology released an updated version of its video generation model Kling on 5 February. ByteDance Ltd. followed days later with its own update, Seedance 2.0, which quickly eclipsed the Kling model in popularity, winning attention for more dynamic camera movement and more coherent visual language.

The pace of launches has only accelerated. Kunlun Tech Co. Ltd. released its new video generation model, SkyReels V4, in late February, topping rankings by benchmarking firm Artificial Analysis for text-to-video with audio in mid-March. Alibaba Group Holding Ltd. followed in early April with its own text-to-video model, HappyHorse 1.0.

Furthermore, the battle is no longer limited to AI model developers. Video platforms and gaming companies are expanding into applications, aiming to capture demand further down the value chain. Bilibili Inc. began beta testing its AI video tool Updream in March, and iQiyi Inc. launched an AI agent platform targeting professional film and television producers. Tencent Games unveiled an AI tool that can convert video into game-ready 3D animation and later introduced a model capable of generating 3D game worlds from text prompts.

Cai Haoyu, co-founder of Mihoyo Co. Ltd., developer of the hit game Genshin Impact, has also thrown his hat into the ring via a US-based startup that released a model capable of generating real-time virtual characters with lifelike expressions that can speak and sing.

The rapid expansion reflects growing real-world use cases. AI video is now being deployed across advertising, e-commerce and entertainment. On social platforms, it powers targeted ads and viral clips. On e-commerce sites, it enables virtual product demonstrations. In entertainment, it is fuelling a surge in AI-generated animation and short dramas.

Demand is growing quickly. As of March, daily token usage on ByteDance’s Doubao large language model had exceeded 120 trillion, doubling within three months. Tan Dai, president of the company’s cloud arm Volcano Engine, attributed the surge largely to AI video and agent workloads.

According to a report released 14 April by the China Netcasting Services Association, more than 2 billion AI-generated video and audio clips were created in 2025, a more than 14-fold increase from the previous year. Douyin, TikTok’s Chinese sister app, estimates that the global market for enterprise AI video applications could reach US$36 billion by 2030.

While demand is surging, business strategies are diverging.

Some companies are already generating revenue. Kuaishou’s Kling model produced more than 1 billion RMB (US$147 million) in revenue in 2025, supported by tens of millions of users and thousands of enterprise clients.

Although the business accounted for only a small share of Kuaishou’s total revenue last year, the company continues to invest heavily and expects revenue from AI-created video to double in 2026. “Video generation models are highly complex… At this stage, we believe video generation technologies and products are still far from maturity,” CEO Cheng Yixiao said during a March earnings call. He also said the launch of Seedance 2.0 and updates from other models are expanding the overall market.

ByteDance’s Seedance 2.0 has shown that strong demand can support premium pricing. After the model shot up in popularity, users faced long waits, with some reporting delays of several hours to generate short clips. The surge in demand allowed ByteDance to raise prices. According to a 4 March announcement, the model costs 13.8 RMB for a roughly 15-second clip — close to 1 RMB per second — higher than market expectations and above competitors such as Kling.

A distributor for ByteDance’s cloud services described the model as operating in a seller’s market, with enterprise customers often required to commit to minimum spending levels.

The broader monetisation logic is becoming clear. Model developers sell access, allowing third-party applications to integrate AI capabilities via interfaces or plugins. Users purchase credits to generate content, while cloud providers monetise the computing power behind it.

Yet not all players are prioritising revenue in the short term. Alibaba’s HappyHorse model is open-source and free, at least for now. “Free access is more of a marketing strategy, but it’s unlikely to be sustainable in the AI video space,” an Alibaba Cloud executive told Caixin.

Alibaba has set an ambitious target of US$100 billion in annual revenue for its cloud and AI businesses over the next five years. To that end, the company is reorganising around token-based monetisation and model-as-a-service offerings, aiming to better serve hardware-makers, video companies and other clients that traditional cloud-sales structures struggle to reach.

Despite rapid progress, significant challenges are emerging.

Legal and ethical issues are among the most immediate. As AI-generated video becomes more realistic, disputes over likeness rights and intellectual property (IP) are increasing.

Chinese courts have begun addressing cases where AI-generated characters closely resemble real actors. In one prominent lawsuit disclosed by the Beijing Internet Court in March, actress Dilraba Dilmurat successfully challenged a production that used a synthetic face resembling her own.

Platforms are also dealing with unauthorised use of images, including those of ordinary individuals. ByteDance’s Hongguo Short Drama removed one disputed title earlier this year after the producer failed to provide enough evidence that its visual material had been used lawfully. The platform later said it had stepped up enforcement against improper use of AI drama materials and had taken action against hundreds of works after a broader review.

The issue is also affecting global expansion. ByteDance’s Seedance 2.0 drew backlash in Hollywood after users generated videos featuring copyrighted IP, including recognisable characters. The company subsequently suspended features allowing real-person facial references and temporarily halted its overseas rollout before gradually resuming the launch in late March.

On April 2, Volcano Engine’s Tan said the company had established a comprehensive copyright protection system covering prompts, intermediate materials and final outputs.

Regulatory scrutiny is intensifying as well. In March, China’s top internet watchdog required platforms to standardise labels for AI-generated and staged short videos. Six major platforms, including Douyin and Kuaishou, removed more than 37,000 problematic clips and took action against over 3,400 accounts in a single month.

A separate campaign targeting altered content involving classic TV dramas, historical and revolutionary themes, and heroic figures removed nearly 29,000 videos between February and March.

In April, Chinese state media reported that AI creation platform LiblibAI had allowed users to generate pornographic videos in some cases due to gaps in its content filters. The company later acknowledged the issue and said it had fixed it.

This article was first published by Caixin Global as “In Depth: China’s Tech Titans Tussle in AI Video Gold Rush”. Caixin Global is one of the most respected sources for macroeconomic, financial and business news and information about China.

Related: Seedance 2.0: China’s AI weapon for cultural and commercial clout | How AI made animation a mass industry in China