WeChat + Deepseek: The turning point of 2C applications

Author: Botaijin; Source: Consensus Crusher

I was very excited when I woke up in the morning. I saw such a big thing as soon as I opened my eyes. It was already 1 am Beijing time when I woke up in the morning, but I still caught a lot of industry friends to talk about the impact of the subsequent products.

Let's talk about why this matter is so big.

First of all, it means that the trial and error cost of the product has been greatly reduced, and large-scale trial and error can be started:

• We have seen from the supply chain that Tencent has added orders for 100,000 to 200,000 H20s. Now we see that the WeChat version of Deepseek has a clear purpose.

• Each H20 can support 500 Deepseek full-blooded users at the same time, which means that 100,000-200,000 H20s can support 50 million-100 million users online at the same time (there are many disputes among readers, we will reply in the comments and tomorrow's article), which basically meets the usage of the first batch of WeChat Deepseek users and exceeds the number of ChatGPT's simultaneous online users.

• Currently, the cost of H20 is less than 10,000 US dollars, corresponding to 15,000 US dollars for a cluster. Assuming 100,000-200,000 cards, it is about 2 billion US dollars; according to AWS's latest depreciation definition modification, GPU can be used for 5 years, and the average annual depreciation cost is 400 million US dollars.

• 400 million US dollars is equivalent to 1% of Tencent's NonGAAP profit this year. It only takes 1% of the profit to do such a big thing, while if it had used the equivalent model of OpenAI before, it would have cost 5-10%, so Deepseek can afford the cost of trial and error, but GPT4o is a cost that Chinese companies cannot accept, not to mention O1/O3.

• The trial and error cost of 400 million US dollars is lower than the trial and error cost of operating Weishi 7 years ago (I was one of the earliest employees of Weishi at that time), and Tencent's profit 7 years ago was only 30% of the current one. Deepseek has reduced the cost to a level that the giant can fully afford to trial and error without affecting its main business.

This may be the reason why it is now and not before. It is equivalent to giving everyone a gpt4o+o1 that is 5-10 times cheaper. You just need to take it...

For articles about how Tencent is thinking about investment now, you can read our articles that were liked by Tencent’s general manager in the past

Tencent: An efficiency machine that no longer FOMO

In addition, Tencent has begun to use Deepseek in knowledge base (IMA) and marketing scenarios. Today, a friend mentioned to me that domestic e-commerce companies have also begun to use Deepseek to optimize personnel, including customer service automation, negative review management, copywriting optimization, and long-tail keywords. In addition, domestic ERP companies serving e-commerce sellers have also begun to productize these functions.

Catch a few use cases:

In our article "Verification Logic of AI Grand Narrative", we discussed Jevons Paradox:

• "But in 2B scenarios, customers have stronger payment capabilities, and better models will bring more usage, and the elasticity brought by price may be limited. For example, the Salesforce Agentforce product currently has a general customer discount of 20%-30%, and the cost reduction of the model is unlikely to encourage customers to offer a 10% discount to stimulate more volume. However, the improvement of the model's capabilities can not only bring an increase in usage, but also a higher ASP."

• "But in most 2C scenarios, cheaper prices mean lower trial and error costs and can cover more customers, there is no problem at all."

Verification logic of AI grand narrative

So after 2C has drastically reduced its costs, this Jevons Paradox opportunity, which is ten times better than 2B, has emerged.

Let's talk about what major changes will be made at the product level.

First, the WeChat version of Deepseek may be much easier to use than Deepseek itself:

• Friends who have tested it can already search for public account articles in real time.

• Those familiar with the domestic 2C ecosystem know that WeChat public accounts and Xiaohongshu content are basically closed loops, and the content of WeChat public accounts is dozens of times the volume of Xiaohongshu content, so the quality is naturally much higher than that of public searches.

• Of course, it is still a beta version, and there are opportunities for continuous improvement.

In the PC scenario, WeChat version of Deepseek can naturally extend to the browser scenario:

• ChatGPT and Doubao are likely to be browsers in the next step, because browsers are the easiest to carry Agent, including the latest Operator scenario, which will also be implemented in browsers. DeepResearch is also based on search and browsing, but users are not aware of it.

• The WeChat search browser is already very frequent on PCs, and WeChat itself is the most direct representative product of dialog boxes. The WeChat version of Deepseek+AI CoT Agent will greatly simplify the future interactive interface. The future browser and even the operating system may be just a window.

• This is the largest front-end traffic collection point and the core distribution position.

In the mobile scenario, according to the tone of WeChat, it will not be very radical. It is more of a trial and error process. It is not too late to go ahead when the technology and scenarios are more mature. Just figure out what form and experience to interact with.

Some scenarios can be done first:

• For example, Agent helps you read, tracks your various public account subscriptions, and then summarizes a daily or weekly report, and can even interact,which will be extended to a personal knowledge base in the future.

• For example, you can do group operation management and add it to corporate WeChat. Many e-commerce operations are already doing this,which is very suitable for the current corporate WeChat + video account e-commerce scenario.

• Another example is the New Year scene that just passed, where you have to send greetings one by one. Can you let AI rewrite the closest friends into sincere blessings according to my tone based on the chat conversations and label relationships of different people? On the other end of the blessing, it is already difficult to feel whether it is you or the Agent.

As for the group chat Agent or personal assistant for ordinary users that we want more, it is highly likely that it will not be so fast or urgent according to WeChat's extremely high compliance requirements.

I guess that in Long Ge's vision, WeChat Agent is not only a tool to improve efficiency, but also a tool to build a content ecosystem and a tool ecosystem, so that everyone can provide creativity to others, improve efficiency, and serve everyone.

This may be what WeChat wants to do.