Tencent executive Tang Daosheng addressed AI strategy criticism in a June 5 media interview, calling the competition a marathon not a sprint. Tang, Tencent Group Senior Executive Vice President and CEO of Cloud & Smart Industries Group, systematically responded to market questions about whether Tencent has been 'slow' in AI development. He acknowledged phase-based differences in progress across Tencent's diverse business units but emphasized the company's long-term approach. The interview covered Tencent's AI agent deployments, computing power challenges, and the impact of former OpenAI researcher Yao Shunyu, who joined Tencent at year-end at age 28.
Tang Daosheng Calls AI Competition a Marathon
Tang Daosheng opened his response by stating, "I remember Shunyu mentioned on stage that the term 'second half' is somewhat overused; now it looks more like a marathon, a longer-term competition." He acknowledged that over three years have passed since ChatGPT's release, during which the industry has undergone dramatic changes. "Tencent's business ecosystem is very diverse and we do many things. I also think it's difficult to ensure every sector is the industry's most advanced. It's normal for different businesses to move faster or slower at different stages," Tang said.
Tang cited Tencent's response to this year's early AI agent wave as an example of rapid execution. "Looking at it another way, for example, this year's early wave, Tencent is also recognized as having the fastest response in the domestic market, and now WorkBuddy is also the most popular product in this track," he stated. He added that Tencent's product philosophy is to "persist through cycles when you determine something is valuable."
Yao Shunyu Brings Model-Product Integration Focus
Yao Shunyu's arrival brought three fundamental changes to Tencent AI, according to Tang. First, it drove coordination between models and products. "From previously when Hunyuan was very concerned with external rankings, it directly changed to using product user experience as the core metric," Tang explained.
Second, Yao significantly improved data quality. "Our data seemingly was abundant, but not high enough quality. Early on, before training Hunyuan 3, a large part of his work was improving data quality, including cutting a lot of data that seemingly could bulk up volume but actually had little help or was even harmful to model training," Tang said.
Third, Yao introduced a simplification philosophy. Tang stated, "If you don't understand the importance of data quality and just blindly pursue more Tokens, then you can't make the decision to cut data." He explained that under Scaling Law influence, complex model architectures with many tricks make scaling difficult, while simpler architectures with sufficient computing power and parameters allow data to fully demonstrate the model's potential capabilities. Tang credited Yao with "great merit" for Hunyuan 3's progress despite not being a very large model.
Currently, approximately 80% of Yuanbao users are using Hunyuan 3, with product retention rates showing clear improvement. Tang disclosed that the Yuanbao and Hunyuan teams will soon move to the same building to facilitate communication and alignment.
Tencent Releases 20+ Vertical Scenario AI Agent Tools
Tencent released efficiency AI agent tools covering over 20 vertical scenarios, with WorkBuddy and CodeBuddy as flagship products. Tang stated, "Tencent has always been very focused on product experience, meeting user needs, and providing value to users. These goals all need products as carriers for users to obtain this value, so when people look at Tencent, they generally say Tencent is a product company. This is in our team's DNA; I don't think there will be much change in the AI era."
Tencent adopted an open model strategy for AI agents. "Today, for CodeBuddy and WorkBuddy, we also adopt an open model strategy. Because these general tools need to support various scenarios for different enterprises and users, we hope to give model selection rights to users," Tang explained.
Regarding WorkBuddy's relationship with Enterprise WeChat, Tang said the two will coexist and develop together. "Enterprise WeChat will focus more on internal person-to-person communication, person-to-service communication, or directly calling OA with some approval processes. But we can also envision future work modes where more is human-AI collaboration; we hope WorkBuddy can provide a more natural AI-native product experience," he stated.
Tencent Faces GPU Supply Bottleneck
Tang acknowledged multiple times that Tencent currently faces a severe computing power supply bottleneck. "In the past few quarters' financial reports, quite a few investors have asked related questions. We have consistently been in a state where infrastructure computing power is not quite enough. With limited resources, we tilt toward internal needs, including Hunyuan training, WeChat needs, meeting needs, etc. Yuanbao also consumes quite a lot of computing power resources," Tang stated.
He explained that the actual GPU computing power allocated to cloud services for customers across industries has benchmark cases but cannot fully cover all customer needs. "Over the past two or three years, we still prioritize serving internal products well. Actually, internal products are also serving external users, so indeed for Tencent, this priority is somewhat higher than renting out GPUs," Tang added.
Tang expressed anticipation for more domestic computing power in the second half of the year. "We very much look forward to more domestic computing power coming in the second half of the year to support cloud business. As more domestic computing power comes in the second half, while meeting internal needs, we can also serve external parties. This is our current plan," he said.
On whether Tencent would increase investment in proprietary chip development, Tang stated, "First, doing chip design ourselves doesn't solve the production capacity problem. Because I deal with many chip manufacturers and partners, I believe no company today has sufficient production capacity to meet today's market demand, so these two things are actually separate matters. Our current approach or this ecosystem combination strategy actually allows us to cooperate with more chip manufacturers and makes everyone very willing to embrace Tencent as a computing power showcase benchmark."
Tencent Sets No Commercialization Targets for AI Agents
Tang clearly stated that Tencent AI business currently prioritizes refining product experience over pursuing commercialization revenue. "For AI agents like WorkBuddy and CodeBuddy, we're still in the investment period; we haven't set commercialization targets for the Buddy team," Tang said. "Agent call volume is not a commercialization metric; it's a usage metric. Commercialization is not our current focus; we still need to refine the product well, serve more users, and prove this is a tool that can create value for everyone and improve work efficiency."
He acknowledged that commercialization serves as a necessary regulator. "Because computing power resources are limited, how to screen out those who most need this product and most recognize the value it creates—value worth paying for to obtain computing power—I think is also something Agent products need to consider in their development process," Tang explained.
On the industry's large model price war, Tang stated that the overall industry trend hopes Token inference costs continuously decrease, which helps popularization and applying AI capabilities to more scenarios. However, different model specifications will have different pricing strategies. "Many manufacturers now make models of different specifications; those with relatively fewer parameters can meet scenarios with higher cost-effectiveness requirements, but at the same time, some particularly difficult problems need larger models with higher costs, and everyone's pricing strategies will differ accordingly," he said.
Tang acknowledged that in AI and cloud service competition trends, Tencent is still in the investment and product-building stage. "Competitors are indeed ahead of us in commercial planning; our style is very different," he stated.
FAQ
What did Tang Daosheng say about Tencent's AI development pace on June 5?
Tang Daosheng stated in a June 5 media interview that AI competition is a marathon rather than a sprint. He acknowledged phase-based differences in Tencent's AI progress across its diverse business units but emphasized the company's long-term approach. Tang cited Tencent's rapid response to this year's early AI agent wave as evidence of the company's execution capability, noting that WorkBuddy became the most popular product in its track.
What changes did Yao Shunyu bring to Tencent AI after joining?
According to Tang Daosheng, Yao Shunyu brought three fundamental changes: driving model-product coordination by shifting Hunyuan's focus from external rankings to user experience metrics; significantly improving data quality by cutting low-value training data; and introducing a simplification philosophy that prioritizes simpler architectures with sufficient computing power over complex models with many technical tricks. Currently, approximately 80% of Yuanbao users are using Hunyuan 3, with improved retention rates.
What computing power challenges does Tencent face for its AI business?
Tang Daosheng acknowledged that Tencent faces a severe GPU supply bottleneck, with infrastructure computing power consistently insufficient. The company prioritizes internal needs including Hunyuan training, WeChat, meetings, and Yuanbao over renting GPU computing power to external cloud customers. Tang expressed anticipation for more domestic computing power in the second half of the year to meet both internal needs and external cloud service demands.
