Hirose AI: OpenAI stops Sora, TikTok's parent company takes the lead—a change in the lead role for video generation AI

Chinese AI can be stopped by export controls—the premise we have believed in has collapsed first in the field of video generation.

What the FT reported on May 17 is that the video generation models of ByteDance (TikTok's parent company, unlisted) and Kuaishou (Hong Kong 1024.HK listed) have surpassed OpenAI and Google in field usage rankings. In the Arena ranking, which features evaluations from practical creators who use them for advertising, e-commerce, and entertainment, the top three are now all Chinese-made.

■ Counter-evidence data: The top three in the Arena ranking are all Chinese-made

The major models mentioned in the FT article are as follows.

Kuaishou's 'Kling', ByteDance's 'Seedance 2.0', and the startup-born 'HappyHorse 1.0'—these three occupied the top spots in the Arena-based Video Arena. Video Arena is operated by the independent evaluation platform Arena.ai (formerly LMArena) from UC Berkeley, and it uses a mechanism where users vote in blind comparisons of the outputs of two models. It is widely referenced in the industry as being closer to practical evaluation than official benchmarks, which each AI lab tends to optimize for itself.

Google's 'Veo 3' maintains competitiveness. It has the advantage of being able to access YouTube's video assets, but it has many content restrictions for developers, making it inferior to the Chinese players in terms of freedom in the field. OpenAI's 'Sora' was discontinued in March of this year due to high computational costs. Video consumes orders of magnitude more tokens than text or audio, and it could not be sustained with free distribution.

When ranked not by benchmark numbers but by 'whether it can be used in the field,' the situation was reversed. This is the starting point of this article.

■ Three structures supporting the reversal

Why were they surpassed? I will present three reasons.

The first is the 'difference in training data.' Training video models requires a massive amount of high-quality video. ByteDance and Kuaishou own some of the world's largest short-video platforms. While text can be web-scraped, video cannot. The structure of platform ownership equals training data monopoly is what worked here.

The second is the 'looseness of regulation.' Director AI's Ben Chiang told the FT that US models 'stop repeatedly due to errors in terms of use violations.' Industry insiders see the Chinese players as being proactive even in accessing copyrighted material. ByteDance faced legal threats for leaving the state where it could generate Marvel and South Park characters without permission. It is at the stage of promising to strengthen its response. The order of 'release first, restrict later' has increased the speed at which it is used in the field.

The third is 'cross-subsidization from the core business.' Video generation consumes orders of magnitude more tokens than text or audio, and as long as it is distributed for free, the deficit will expand. The FT reports that one of the main reasons OpenAI ended Sora was this computational cost. In contrast, ByteDance and Kuaishou have a structure that allows them to absorb the computational costs of their video AI business using advertising revenue earned from their core short-video platforms (TikTok and Kuaishou). The high-priced billing of $2 million (approx. 310 million yen, 158 yen/$) for pre-payment to US corporate clients is also a mechanism to secure the funds to purchase computational resources from the enterprise side as well. The limit to what can be continued is different between OpenAI, which must earn money from AI alone, and the Chinese players, whose core businesses can support AI.

US export controls focused on squeezing Nvidia chips, but training data ownership, the speed of regulatory operation, and cross-subsidization from core businesses—these three were areas that semiconductor regulations could not reach.

■ Commercialization is already underway

While we were continuing the debate about 'being able to stop it with regulations,' commercialization had already begun to run.

Vincent Yang of Firework, which provides video infrastructure for e-commerce sites, was ordered by a retailer to produce '100,000 videos for product pages.' It is a scale that would be impossible to price without AI. One video per product, and multiple versions for different customer segments—we are at the entrance to a restructuring of the economics of advertising production.

Kuaishou announced this week that it is considering the option of spinning off the Kling business into a separate company and listing it separately. We have already entered the phase where video generation AI is being used as material for listing as is.

■ Investors rethink the premise of 'all-in long on AI'

While they were surpassed in video, OpenAI, Anthropic, and Google are still the main players in the LLM (Large Language Model) and coding fields. It was reported on May 15 that Anthropic is proceeding with an additional $30 billion (approx. 4.7 trillion yen) in funding. Nvidia also continues to hold the initiative in chip supply and demand.

What was surpassed is one area called 'video generation.' However, this area is also a field that is ahead in commercialization, such as advertising, e-commerce, and entertainment. If LLMs are the main players in research and operational efficiency, video generation is highly likely to become the main player in advertising budgets and consumer touchpoints.

Let's bring this down to the investor's perspective. Now that a changing of the guard has occurred in video generation, the premise that 'you can cover the entire sector by going long on US AI stocks in bulk' has collapsed. On the US side, the front-runner for publicly available video generation products is Google's (GOOGL) Veo 3; OpenAI's Sora will be discontinued in March 2026, and Meta's Movie Gen remains in a pre-public release stage despite being announced in 2024. Microsoft does not have its own video generation model. Options for buyers are also still thin. Kuaishou is listed in Hong Kong (1024.HK) but has not been converted to ADRs, and ByteDance is private, meaning general investors cannot hold it directly, just like SpaceX or Anthropic.

While LLMs and coding remain firmly in the hands of GOOGL/MSFT/Nvidia, video generation is now almost entirely consolidated into Veo 3 among publicly available US products, and the number of Chinese stocks available to buyers is limited—this asymmetry will be the composition for the time being. If you lump everything together as 'US dominance in AI' without segmenting the areas, you will overlook the differences in profit and loss by segment from here on out.

Before you base your strategy on the premise that 'regulation will lead to victory,' keep in mind one case where the rankings were overturned by actual user experience in the field. This is the one point I want you to take away today.

--

Related article: 'The man who wrote 6 years ago about what would happen if Jicchama died, invests 397 billion dollars' →

For more details on Hirose AI, click here →

※ This article is an analysis by Hirose AI and does not represent the views of Takao Hirose himself. Please make investment decisions at your own risk.