Why AI Generation Products Should Start Charging Where Compute Costs Most

Jiang Wu · 2024-08-15

*Originally published in Chinese on August 15, 2024.*

Kling's pricing and what it says about paying users

Kuaishou, the Chinese short-video platform, has started charging for Kling, its video generation service. I can't get Kling's payment data. What I can do is monitor how much video Kling users generate across platforms, see how much new content that adds, and work backward to a rough paying ratio.

Video users are still holding back

My impression is that early payment is not high. The cause is not how hard generation is or how good the results look. For most ordinary users, video generation has few uses, and it has not yet become something they reach for often.

Where generated video is used, and where it breaks

Some small and mid-sized studios have started using generated video for client work, such as promotion and packaging for specific scenes. Because of continuity problems today, the output stays in short clips. Beyond one minute, keeping a video coherent is still a technical obstacle.

Image generation is worth more than video generation

Among consumer products, images currently earn better than video. The big traffic-driving ads and traffic-driving sites also run on an image logic. Models for small and mid-sized fixed tasks are now cheap enough to run, so image products have higher gross margins.

Image tools built for a fixed task

For tools that only generate images, the added value to users is still in doubt. Tools built for a fixed task are worth more to the people who pay for them, for example generating a résumé photo or a cover image, resizing resolution, or making cartoon-style portraits.

Competition and technical investment

In video, the product updating fastest right now is CapCut, known in China as Jianying and owned by ByteDance. AI features are already a standard module there, and I expect the pace to pick up. In images, the product users respond to best and rate highest is Meitu, the Chinese photo-editing company. Meitu does its cost math for models over a 3 to 5 year cycle, so how far the average cost gets spread out is still to be seen, and it plans to keep building its own models over the long run.

Where video generation tools fit in the market

Unlike earlier categories of tools, video generation today has too little support from large enterprise customers and too little pull from individual users. The workable step for now is to push toward small businesses and find scenarios where they will pay in stages. The consumer base is large enough that long-run investment should at least break even, and at worst a company can treat it as a technical moat for its own business ecosystem, as Douyin or Kuaishou can. Companies with no ecosystem of their own may run into trouble on the numbers.

Technical capability and user scenarios

Products will later be split by technical capability and user scenario, with a professional tier. Feeding video data in has been validated in this phase. The remaining problem is that generated video cannot be used to train tools and models, and the industry is still working on it.

Home · All articles · Topics · 简体中文