2026.09.24 | TB | China AI | Wondershare on China AI Apps & Seedance

Client-led • 2 Aug 2026

Companies: 1024 HK, 700 HK, 100 HK

Expert Senior Executive, AI Product at Wondershare Technology Group Co Ltd

Agenda

• Around 10 years' experience in AI products and large AI model applications, familiar with text-to-video and large AI model commercialisation
• Deep insights into large video AI models such as Seedance and Kling AI, plus the business model of automatic video generation for short drama
• Can discuss market growth and token invocation in the text-to-video space, the MAUs and gross margins of Kling AI and Seedance, plus 3D models' underlying evolution

Questions

[00:00:12] With the release of Seedance 2.5 and MiniMax's new models this weekend, the entire AI application sector has seen heightened market enthusiasm. What is your view on the overall growth of the market? What is the overall token invocation volume of these text-to-video companies? What is the month-on-month growth rate?

Expert [00:00:12]: It is indeed very popular this week, especially with the release of Seedance 2.5, which might be a bit more eye-catching than MiniMax. It has a few core highlights. First, its entire native video duration was probably around 15 seconds before and now it is 30 seconds. I will not go through the reference inputs one by one, because there are actually several segments in the entire market now. Regarding the AI video track, one uses AI foundation models, such as MiniMax, Volcengine and ByteDance's Seedance 2.0. They belong to one track. There are also native AI aggregation platforms or various current aggregation platforms for AI short dramas, which constitute another track.

Client [00:01:58]: The underlying layer for short dramas should also be foundation models, right?

Expert [00:01:58]: Yes, but it does not conduct model R&D. It only focuses on aggregation, which means it only earns the margin in between.

[00:02:08] The short drama business is essentially inputting novels and other content to generate short drama output. The application developers earn revenue, then pay the foundational model developers and make a profit from the margin in between, right?

Expert [00:02:08]: Yes, this is actually the path of text-to-video editing. Most of them have a limited profit margin, about 30%.

Client [00:02:34]: Given 30%, the foundation model they use is probably not the latest generation, right?

Expert [00:02:34]: They previously used Seedance 2.0. For example, the API price of Seedance 2.0 is about RMB 1 per second.

Client [00:03:11]: Hasn’t it already launched Dreamina AI? There is a price for Dreamina AI, which is equivalent to the retail price, right?

Expert [00:03:11]: Yes, the retail price, which is basically equal to RMB 1 per second.

Client [00:03:24]: Does RMB 1 per second refer to the price of 2.0?

Expert [00:03:24]: That’s right. 2.5 has just been launched and its commercial price has not been negotiated yet. The second is aggregation platforms. The third is that you already have very mature AI products and are making them AI-enhanced, such as CapCut or say Filmora. You have your own products and are making them AI-enhanced, mainly in these segments. Like the launch of Seedance 2.5 this time, previous surveys on some other aggregation platforms showed that Seedance 2.0 basically occupied 70-80% of the market and most people used Seedance 2.0.

[00:04:18] Seedance 2.0 uses ByteDance's proprietary foundation model with internal settlement, correct?

Expert [00:04:18]: Basically, it is accessed through Volcengine. Volcengine charges clients based on the number of calls or by token. It requires you to sign an annual framework agreement before it negotiates discounts with you. At present, it seems that the biggest competitive advantage is the price of the model itself and it only makes money from this part.

[00:05:27] In reality, those who can earn excess returns right now are mainly those doing secondary development on foundation models, such as Seedance, right?

Expert [00:05:27]: It is an aggregation platform where you can do video editing. The core of video editing here is that it aggregates different models, such as Seedance 2.0, Runway and MiniMax, as well as Sora previously and domestic models such as Vidu and Kling. After aggregation, it makes money from subsidies, resulting in a very low gross profit margin.

[00:06:12] What is the current ARR [annual recurring revenue] of Seedance?

Expert [00:06:12]: Seedance’s ARR is not officially disclosed. However, it should have 50-60 million MAUs every month.

Client [00:06:39]: Is there any data on the MoM growth rate of MAUs?

Expert [00:06:39]: Some theatre-oriented products have a MoM growth rate of 30 %, which is quite fast. However, the difficulty lies in that such theatre-oriented products don’t have strong product competitiveness because users may use models. In China, it depends on how each company negotiates discounts with Seedance.

[00:09:57] Could you introduce the overall situation of Kling AI?

Expert [00:09:57]: Most of Kling AI’s revenue comes from overseas markets. As far as I know, Kling AI doesn’t have a big copyright risk to individual customers because it started early. It took off in H2 2024 and Seedance was launched during this Spring Festival.

Client [00:10:24]: What is the approximate MAU of Kling AI?

Expert [00:10:24]: Kuaishou should have public financial reports. My impression is that its Q1 revenue was around RMB 600m-700m, but I don’t know much about its MAU.

[00:17:13] Do you have insights into AI application adoption in the office and financial sectors?

Expert [00:17:13]: For office apps similar to WorkBuddy, if you look at Adobe PR, its Creative Cloud has a relatively fast growth rate. There is also a segment where you originally have a flagship-level video application and then add AI enhancements. It is primarily for video editing, but it later integrated some AI capabilities, such as AI image-to-video generation and AI video editing, to enhance its competitiveness. The growth rate on the mobile end is very fast. However, although the products themselves are more competitive now, the whole market is in fierce competition and software such as image-to-video generation has somewhat squeezed the space for video editing. In addition, ByteDance has a video editing software called CapCut, which grows rapidly every year. It also has distribution platforms, making it relatively difficult to compete with it.

[00:23:11] For 3D models, do their foundation models still rely on GLM [general language model] or models such as DeepSeek V4?

Expert [00:23:11]: No, 3D models are new. For example, GLM is a semantic large AI model launched by Zhipu, called GLM-5.2 and GLM-5.1, but the video model uses the DiT architecture, namely diffusion image transformer. There will be new things in 3D. Normally, when AI generates a video, one second of video essentially generates 24 frames of images, which are combined by judging their spatio-temporal consistency, called optical flow, which is essentially planar. 3D is a bit more complex because 3D has vectors and physical models. Therefore, its architecture is different and the generated content is also different. At present, the most mainstream 3D model generations are called world models.

Client [00:24:24]: However, as for the vendors of such world models, they still distil other people's work.

Expert [00:24:24]: That's right. That still feels like a toy. 3D models might be used in industrial scenarios, but it is difficult to use them in C-end scenarios like this. As for what you mentioned about game modelling, industrial modelling or audio and video, I haven't seen them used.

[00:25:07] Are you familiar with Tencent Hunyuan's product that can generate games directly from scripts?

Expert [00:25:22]: That is essentially a text-to-code model. The competitive landscape in China is very clear. In the video field, Seedance 2.0 takes a cliff-like lead, followed by Kling AI. It does have revenue and is not making a loss. That is, while many video models operate at a loss, Kling AI is not making a loss, as it is profitable. The more compute it has, the more profit it may make.

[00:26:00] What are the gross profit margins of companies such as Kling AI and Seedance?

Expert [00:26:00]: The gross margin of Seedance 2.0 is very high. I am not sure about Kling AI, but I know that it is profitable because Kling AI and Kuaishou build their own compute.