Tencent's Dowson Tong commented on Shunyu Yao, Hunyuan 3, and Yuanbao

Text by | Wang Yuchan

Edited by | Zhang Yuxin

On June 5th, what attracted the most external attention at the Tencent Cloud AI Industry Application Conference?

Undoubtedly, it was the dialogue between Tang Daosheng and Yao Shunyu.

At this conference where a series of Agents covering more than 20 vertical scenarios were launched, since the products were too To B and there was no mention of the "WeChat AI" that everyone was most concerned about, almost all the external attention was drawn to that dialogue.

Tang Daosheng, the Senior Executive Vice President of Tencent Group and the CEO of the Cloud and Smart Industry Group, asked Yao Shunyu, the Chief AI Scientist of Tencent and the person in charge of Tencent's Hunyuan Large Language Model and AI Infra, in the dialogue, "Many people say that Tencent is slow in AI. Do you think we are really slow?"

Yao Shunyu replied, "It seems that this should be a question I ask you."

Yao Shunyu then mentioned two criteria for judgment - 1. Is AI a short - term or long - term game? 2. Will it be a linear or diversified game?

Yao Shunyu, a 27 - year - old so - called "genius teenager from Tsinghua University" who once worked at OpenAI, has joined Tencent for half a year. His thinking is almost the same as Tencent's AI thinking. His conclusion is: First, AI is a long - term game and has just entered the second half; second, Coding Agent is very important.

As Yao Shunyu described, Tencent has launched a series of Coding Agent products, tools for efficiency improvement that serve both To B and To C. In contrast, Yuanbao, which cost Tencent a large amount of marketing expenses at the beginning of the year, was almost invisible at the conference.

In terms of user activity, Yuanbao has been far left behind by Doubao. And that series of Coding Agent products seem too professional and not very appealing.

But they are indeed closer to making money.

Yao Shunyu reports directly to Liu Chiping, the President of Tencent. At the Q1 earnings conference call, Liu Chiping said:

"The ability to find high - value use cases is at least as important as getting a large number of DAUs and user hours, and may even be more important. In the AI world, you must find high - value use cases, not just focus on DAU. The difference between the AI revolution and the Internet revolution is that AI is about 'intelligence', and the value of intelligence is reflected in how much people are willing to pay for it. In the field of AI, each service delivery incurs a relatively high cost."

Enterprise customers are the "high - value use cases" mentioned by Liu Chiping. Improving efficiency is the main service mode of Tencent.

But there are still problems. Token consumption is still expensive, and Tencent's computing power is still in short supply. Customers may "improve efficiency but also increase expenses". Although Tencent has found the gold mine, is there enough "shovels"?

Backstage at the conference, Yao Shunyu did not face the media again, but Tang Daosheng was interviewed by several media outlets including 36Kr Smart Wave. Somewhat unexpectedly, he said that "current commercialization is not our focus" and "we have not set commercialization goals for the Buddy team".

Tang Daosheng talked about topics such as business models, computing power shortage, and the "horse - racing mechanism", and also evaluated Yao Shunyu, but avoided questions about "WeChat AI".

"Tencent's business is very diversified and it does many things. It's hard to ensure that every segment is the most advanced in the industry. So it's normal for different businesses to progress at different speeds at different stages," Tang Daosheng said. "Looking at Tencent's successful businesses over the past nearly 28 years, not all of them had a smooth journey. They all experienced ups and downs. Tencent's philosophy is that when you are clear that a product is valuable, you should be able to persevere through the entire cycle."

"Commercialization is not our focus"

Q: Tencent Cloud previously judged that simply charging by token and API is not a long - term and healthy business. If shifting from resource sales to charging by task results and business value, what is the path? How to balance the contradiction between current token - based revenue and long - term value - based pricing?

Tang Daosheng: I think both are possible. In fact, they are just different charging methods.

There is no clear indication whether the API method based on tokens is charging below or above cost. Charging above cost is a sustainable business model. I believe that in some industries and business scenarios, it is possible to charge by effect. However, to achieve certain effects for a certain thing or task, there are many factors involved.

If a product only makes a partial contribution to the final result, it is very difficult to fully charge for the value it provides based on the effect. But you can't deny that partial value is not value. So in such a delivery situation, I think it's quite difficult and prone to disputes to charge purely based on results.

Q: How does Tencent Cloud help AI application companies avoid the gross - margin trap where the more active users are, the higher the inference cost?

Tang Daosheng: The marginal service cost of the mobile Internet is relatively low. So we can establish a business model through advertising, the attention economy, e - commerce, or driving certain trading behaviors, at least ensuring that your revenue is higher than the cost. But for AI - native services, with today's high operating and inference costs, it's very difficult to cover users' usage costs purely through the advertising model.

Especially when users may have different consumption levels for different questions and tasks, it's even more difficult to have a stable return and ensure that advertisers will pay for the uncertain operating costs when users don't pay.

So, given the strong correlation between today's token costs and task complexity, ToC charging - whether it's subscription - based, based on the number of input tokens, or more detailed, with different prices for different numbers of input and output tokens, and different prices depending on whether there is a cache - involves very complex transactions. The cost consumption varies greatly, and it's difficult to monetize using the same attention - based model.

So, I think if the inference cost of AI products remains at this level, they are still likely to be used in scenarios with high commercial value, where you can make a profit, or where the new productivity they bring can offset the higher costs you would incur without them.

Q: Now many people are talking about the token economy. Will Tencent or Tencent Cloud have key commercialization indicators for assessment, such as token call volume and industry penetration rate? Will there be such indicator assessments?

Tang Daosheng: The call volume of Agents is not a commercialization indicator. It is a usage indicator. Currently, commercialization is not our focus. We still need to polish the products well, serve more users, and prove that it is a tool that can create value and improve work efficiency for everyone.

But we will have a business model, which is a regulator. Since computing resources are limited, how to screen out users who have the greatest need for this product and recognize that the value it creates is worth paying for computing power is also something to be considered in the development process of Agent products.

Q: The price per token of the Hunyuan model has also decreased. At the same time, DeepSeek and Xiaomi are also significantly reducing prices. Domestic chips will be available in the second half of this year. Please judge what the curve of the price decline of large models will be like, and what impact it will have on the pricing of Tencent's AI products. Is there a possibility of price reduction?

Tang Daosheng: I'm not in a position to comment on the price strategies of other model manufacturers. However, the general trend in the industry definitely hopes that the inference cost of tokens will continue to decrease. This will help with popularization and enable the use of AI capabilities in more scenarios to solve more and more difficult problems. This is a major development trend.

But at the same time, I also see some trends. According to the scaling law, the more parameters and the larger the model, the performance and capabilities still improve. So many manufacturers are now developing models of different specifications. There are models with relatively fewer parameters to meet scenarios with higher requirements for cost - effectiveness and lower ROI inference costs. But there are also some very difficult problems that require larger models, which of course have higher costs, and the pricing strategies will also be different.

Q: The cloud business segment achieved full - year profitability in 2025. In the AI era, WorkBuddy is very popular and has a large number of users. Will we conduct profit or ROI assessments on AI Agent products, or will we continue with strategic promotion and investment? Since our computing power and cost expenditures are increasing significantly, will there be further pressure in this regard?

Tang Daosheng: Tencent has multiple business tracks, and each track has many different products, which are in different stages. For AI agents like WorkBuddy and CodeBuddy, they are still in the investment stage. We have not set commercialization goals for the Buddy team.

However, at the same time, since we have received a lot of interest from enterprise customers in WorkBuddy and CodeBuddy, there is a clear business model in the enterprise scenario. But that is more of a normal extension of the cloud business, which originally served enterprises.

I think WorkBuddy today is a bit like Tencent Meeting a few years ago. It has both ToC and ToB attributes. We will continue to leverage its C2B capabilities to build a sustainable service system.

Q: Recently, some competitors have plans to charge on the C - end for their large models. What are Tencent's commercialization plans for its C - end large models? How do you view the clear commercialization plans of competitors on the C - end?

Tang Daosheng: There are different commercialization paths in the industry. I think the current goal is to improve the product experience, find the differentiated positioning of Yuanbao, and serve more users.

Q: Will MaaS or TokenHub be the main goal this year? Is there an expectation or goal for MaaS in terms of revenue?

Tang Daosheng: It is definitely a very important part. It is a high - growth segment that has experienced rapid growth. It is the fuel for agent products. So I believe that today's consumption of many MaaS tokens is also related to our deployment of WorkBuddy on the enterprise side.

I'm still very optimistic because the market demand is very strong. As I mentioned earlier, limited by the supply of computing power, we are also looking forward to - since token services are also a carrier of computing power, as computing power becomes more abundant, I believe this will be a huge new growth point for the entire cloud market.

Q: In terms of MaaS, we recently saw that some competitors have set ambitious revenue targets. How do we view the competition with other cloud providers in MaaS, and what are our own goals?

Tang Daosheng: Each company has different development styles and rhythms. Tencent prefers to let products and data speak.

Q: Tencent Cloud's AI business is still in the investment stage. In this long - distance race or marathon, how can we prove to the market that our AI investment has a quantifiable expectation? When can the investment in AI Agents cover the costs?

Tang Daosheng: It's hard to predict the specific time. The AI business is currently in the strategic investment stage.

Q: Tencent Cloud has achieved large - scale profitability, but there was some pressure on the growth rate in the first quarter. In the future development, while ensuring a certain profit margin and a certain level of in - house technology utilization, is there any further consideration for the market scale that you can share?

Tang Daosheng: I can't predict future revenues. Certainly, the team also has a relatively positive growth target.

About Yao Shunyu and Hunyuan 3

Q: Why did Tencent choose Yao Shunyu at that time? And what changes has Shunyu brought to Tencent's AI since he came to Hunyuan and Tencent's AI department?

Tang Daosheng: There should be no doubt about choosing Shunyu. He is a very influential expert in this field. Even in the full communication before he came, we could fully feel his professionalism. Indeed, the understanding of the AI - native generation is very different from what we used to have.

After he came, he brought great value to Yuanbao. He actively promoted the co - design of the model and the product. Originally, Hunyuan attached great importance to external benchmarks. Later, it directly changed to using the product user experience as the North Star indicator.

At the same time, although we may have a lot of data, the quality is not high enough. So in the early stage, many of his tasks before training Hunyuan 3 were to improve the data quality, including identifying and no longer using a lot of data that seemed to be able to increase the quantity but actually had little help or was even harmful to model training.

I think having a correct understanding of the development of AIGC models can lead to more appropriate decisions. For example, if you don't understand the importance of data quality and just blindly pursue more T of tokens, you won't be able to make the decision to cut data.

Moreover, according to the scaling law, if you want to make the model very complex with many tricks, it will be difficult to scale; or if you want to scale a lot of complex architectures, you have to make the architecture simpler to ensure sufficient computing power and parameters, and that your data can fully reflect the potential capabilities of these model sizes.

I think he has done a lot of things to simplify complexity. For Hunyuan 3 Preview, although it is not a very large model today, it has made great progress compared to before. There is no doubt that he has contributed a lot. So, the cooperation between Yuanbao and Hunyuan in the past half - year has achieved more progress than in a much longer period before.

Q: After the launch of Hunyuan 3 Preview, the token call volume has doubled compared to the 2.0 version. When will the official version be launched, or when will the next - generation new model be introduced to the public?

Tang Daosheng: Stay tuned. Don't put too much pressure on Shunyu. The work is actually in full swing.

Q: In the current strategy, which contributes more to the growth of Tencent Cloud's AI at this stage, the iteration speed of the basic model or the engineering implementation ability of Agents?

Tang Daosheng: If you usually use WorkBuddy, you will also find that its automatic mode calls different models for different problems. Today's agents can help us solve many problems encountered in office scenarios to a large extent because the capabilities of the models have developed to the current level. With the same tool, if you use the models from two or three years ago, you definitely won't be able to achieve today's results. So I still think that the iteration of models is crucial for the development of agents.

As I mentioned earlier, the automatic mode calls different models for various reasons. Tencent has always taken an open attitude towards the construction of agents and AI product solutions. We are very willing to cooperate with different model manufacturers. For example, last year, Yuanbao had a deep integration with DeepSeek.

Today, for CodeBuddy and WorkBuddy, we also adopt an open - model strategy. Since these general - purpose tools need to support various scenarios of different enterprises and users, we hope to give users the right to choose models. Of course, Hunyuan is also continuously iterating, and we also have goals for model capabilities. Many times, when customers use our agents, they are also particularly interested in calling Hunyuan. We will still adopt a relatively open strategy to develop the AI agent business.

"Yuanbao is undoubtedly an important product"

Q: There are still many expectations for the co - design between Yuanbao and Hunyuan 3. It was also mentioned this morning that the key to the cooperation between the two sides is to build trust. May I ask what the cooperation rhythm between the two sides is like? As a C - end product, what are Yuanbao's core positioning, growth goals, and KPIs?

Tang Daosheng: In fact, the cooperation between the two sides is getting closer. Recently, they will even move to the same building, which will facilitate communication and alignment. We can see that about 80% of Yuanbao's users are already using Hy3 preview, and there has been a significant increase in the product's retention rate. Now, many different services in Yuanbao are supported by Hy3 preview, including the latest AI voice recognition and dialect recognition, which are all based on the base model of Hy3 preview to train other models.

One aspect of Yuanbao's KPI is