In December 2025, an internal Tencent announcement nevertheless stirred up massive waves across the tech circle. Former OpenAI scientist Yao Shunyu (Vinces Yao) has formally taken office as Chief AI Scientist of Tencent’s “CEO Office,” while also leading the newly established AI Infra Department and the Large Language Model Department. Just over ten days earlier, Tencent had released the highly anticipated Hunyuan large model version 2.0. From high-profile poaching to organizational restructuring to a dense cadence of new releases, Tencent is using a series of combination punches to declare: in the “second half” of China’s AI race, it is no longer willing to be a follower.
Yao Shunyu: A Theorist of the “AI Second Half” Standing on OpenAI’s Shoulders
To understand the deeper meaning of Tencent’s adjustments this time, one must first know this 27-year-old “key figure.”
Yao Shunyu’s résumé can be described as “top of the top”: he earned his undergraduate degree from Tsinghua University’s “Yao Class” (Tsinghua Xuetang Computer Science Pilot Class), pursued his PhD at Princeton University, and after joining OpenAI in August 2024, quickly became a core member, deeply participating in the R&D work for OpenAI’s first batch of agent products in 2025—Operator and Deep Research. In May 2025, he became one of the youngest China-based honorees selected for MIT Technology Review’s TR35 list. This honor, hailed as the “Nobel Prize of the tech world,” attests to his top-tier potential in the global AI field.
But what truly propelled Yao Shunyu “out of the circle” (into broad public attention) was his “AI Second Half” theory. In his research essay The Second Half, he clearly points out: “In the first half of AI development, the industry’s core is to compete on model training, compare parameter scale, and chase performance metrics; but as we enter the second half, the center of competition will shift to value definition in real scenarios, optimization of interactive experience, and the construction of evaluation systems—the core logic is that evaluation is more important than training.”
This theory hit Tencent’s pain point precisely. In the past two years of the large-model race, Tencent Hunyuan has released more than 30 new models. The Hunyuan Image 3.0 model that was open-sourced in September even won first place in the LMarena blind test, and the total number of derived models exceeds 3,500. Yet the growth of its consumer product Yuanbao went through a “roller coaster”—in early 2025, riding the open-source tailwind of DeepSeek, monthly active users surpassed 40 million within a month, and it once overtook ByteDance’s Doubao to top Apple’s free app chart, but growth slowed in mid-year, exposing the constraint of its self-developed foundation-model capabilities.
Tencent’s advantages in the AI track lie in distributed training, inference optimization, and cloud-native elasticity—that is, the core engineering capability to land AI at scale with low cost and high stability. Tencent’s shortcomings are also very obvious: original foundation-model R&D “from 0 to 1” requires deep research accumulation and sharp insight into the cutting edge of technology, which is precisely where Tencent, compared with some top peers, has at times shown weaknesses. The value of Yao Shunyu’s background lies exactly in that he brings the “research engineeringization” experience Tencent needs most. During his time at OpenAI, he was mainly responsible for the Agent framework and post-training work, which can precisely make up for Tencent’s inadequacy on the “research-driven” side.
Tencent’s choice of Yao Shunyu is not only a talent grab, but also a grab for “ideas.” More importantly, he reports on a dual line to President Martin Lau (Liu Chih-ping) and TEG President Lu Shan. This “dual-hatted” identity breaks the internal departmental walls commonly seen inside Tencent and paves the way for integrating research and engineering.
Three Signals From the Organizational Restructuring: The “Iron Triangle” of Compute, Data, and Organization
This adjustment is not a simple personnel change, but a thorough reconstruction from a “loosely allied federation” to a “technology core parent body.”
Signal One: The AI Infra Department—Filling the Compute Shortfall
The newly established AI Infra Department is personally led by Yao Shunyu and is responsible for the technical construction of large-model training and inference platforms, focusing on core technologies such as distributed training and high-performance inference services. This arrangement carries far-reaching significance. In the past, although the Tencent Hunyuan team was a company-level project, internal participating departments included the Large Language Model Department, AI Lab, and the machine learning platform, among others. There existed internal resource competition—“competing for compute, competing for people, competing for scenarios”—which led to fragmented resources, diluted outcomes, and severe internal friction.
Tencent’s current organizational adjustment, placing both the AI infrastructure line and the large language model R&D line under Yao Shunyu, marks that the company is formally abandoning the previous traditional organizational model that separated “infrastructure” and “algorithms.” This change essentially emulates the “research–engineering integration” collaborative paradigm practiced by OpenAI, aiming—through tight collaboration within the same team—to systematically tackle the core challenges of AI development such as scaling laws, inference cost optimization, and Agent architecture design. This kind of integration folds Tencent Hunyuan’s accumulated compute platforms, cloud-native frameworks, and 3D generation assets into the new pipeline, becoming the underlying support for multimodal Agents—this can be described as “sublation” (preserving and transcending) rather than “tearing down and starting over.”
Signal Two: The AI Data Department and the Data Computing Platform Department—A Data Moat
Data is the “fuel” of large models. This time, Tencent has reorganized and reintegrated the data capabilities of the original large-model team, newly establishing an AI Data Department and a Data Computing Platform Department, led respectively by Liu Yuhong and Chen Peng, both reporting to Tencent Vice President Jiang Jie. This signifies that the integration of data cleaning, evaluation systems, and big-data platforms has been elevated to the highest priority.
AI has become the core track of global technology competition. Giants at home and abroad are all increasing investment; if Tencent does not accelerate its deployment, it may very likely be marginalized in the future technology landscape. Data is precisely Tencent’s hidden advantage—WeChat, QQ, Tencent Video, and the complex interactions generated by its games produce data that covers many dimensions, including text, images, voice, video, behavioral sequences, and more. The data quality is high and the dimensions are rich—valuable resources for training multimodal, cognition-oriented AI. If efficiently integrated, it will constitute a moat in Tencent’s AI field that is difficult to replicate.
Signal Three: Power Centralization and a Talent Reshuffle
Behind the organizational adjustment is a silent personnel revolution. It is reported that senior managers who lack deep research backgrounds but have extensive tenure may face demotion or reduced authority, while rising talents with model-research experience are being promoted into core positions. Wang Di, Deputy General Manager of the Large Language Model Department, will continue to stay in the role but will report to Yao Shunyu; data leaders such as Liu Yuhong and Chen Peng will report to Jiang Jie, forming a dual-core driving pattern of “Yao Shunyu—algorithms and infrastructure” and “Jiang Jie—data and platforms.”
Even more noteworthy is Tencent’s “spend money” talent-grab strategy. The most generous AI employer in 2024 was ByteDance, but in 2025 Tencent’s name was conspicuously added to the list. According to The Information, Tencent poached ByteDance’s AI team with double salaries; for PhD campus recruits, offering 50% higher pay is only the “baseline,” with the maximum reaching 2x. Behind this willingness to spend at all costs is the Penguin Empire’s deep anxiety on the AI track.
Tencent’s “Second Half” Anxiety: A Giant Awakened by DeepSeek
Why did Tencent choose to hit the accelerator at the end of 2025? The answer is hidden in a series of external shocks and internal reflections.
External Pressure: Squeezed by ByteDance and Alibaba
ByteDance’s Doubao has already surpassed 100 million monthly active users, Alibaba’s Tongyi Qianwen is expanding strongly in the enterprise market, and while Tencent Hunyuan is not weak technically, it has consistently lacked a killer application. The market once questioned whether Tencent’s capex in AI was aggressive enough. Although Martin Lau responded on an earnings call that he “does not think Tencent is truly behind in large models, and is satisfied with the progress already achieved,” in the face of fierce expansion by other leading big tech players, the intensity of market competition cannot be underestimated.
The DeepSeek Shock: Opportunity and Warning
In early 2025, the emergence of the open-source model DeepSeek became a turning point. On one hand, Tencent Yuanbao gained its first wave of growth by integrating DeepSeek, proving users’ hunger for high-quality models. On the other hand, this growth that relied on an external open-source model made Tencent’s leadership realize that without self-developed core capabilities, it would be constrained by others. When Yuanbao’s growth slowed in mid-year, the problem began to surface: what ultimately supports the application side is still the underlying model. Hunyuan’s reputation in the developer community was mediocre, making it difficult to establish technological identification and unique value in developers’ minds—this indirectly reflects that foundation-model capability is the cornerstone of application growth.
Internal Reflection: From “Serving the Business” to “Technology-Driven”
In the past, Tencent’s AI deployment followed a “serve business demand-driven” logic: each BG had its own model team, and resources were hard to concentrate. This adjustment explicitly shifts toward “taking the foundation model as the core axis,” building a technological “parent body.” Tencent’s determination to invest in AI had in fact surfaced long ago—Pony Ma (Ma Huateng) stated at Tencent’s 2024 annual meeting that Tencent would continue investing in compute power, strengthen resource reserves, and encourage each BG to embrace the productization of large models.
This shift can be seen in Yuanbao’s adjustment. Tencent moved Yuanbao from TEG to CSIG and placed it under the leadership of Wu Zurong, the head of Tencent Meeting. Wu Zurong is one of the rare product generals within Tencent who has experience launching a consumer product from 0 to 1 and achieving scaled growth. Putting Yuanbao under Wu Zurong also reflects Tencent’s hope that Yuanbao can learn from Tencent Meeting’s successful commercialization experience, aiming to build another phenomenon-level product in the AI assistant domain. This personnel move also reveals Tencent’s resolve—not only to make strong models, but also to produce a blockbuster at the application layer.
Hunyuan 2.0: A Technical Sword Drawn, or Strategic Make-Up Lessons?
Hunyuan 2.0, released recently, is the first results check after this organizational restructuring. According to official information, the model has significantly improved its pretraining data and reinforcement learning strategies, achieving clear gains in both reasoning and efficiency.
On the technical level, Hunyuan 2.0’s optimization direction is highly aligned with Yao Shunyu’s “Second Half” theory—it no longer simply pursues parameter scale, but instead focuses on data quality, evaluation systems, and efficiency improvements. Over the past year, the Hunyuan team released more than 30 new models, among which the 3D model has entered the world’s leading echelon. The Hunyuan Image 3.0 model open-sourced in September even won the LMarena blind test, showing its accumulation in the multimodal field.
However, the market reaction has been relatively calm. Compared with ByteDance Doubao’s aggressive promotion and Alibaba Tongyi’s deep enterprise cultivation, the release of Hunyuan 2.0 looks more like a round of “strategic make-up lessons”—making up for shortcomings in self-developed capabilities, rather than leading an industry disruption. Tencent’s algorithmic capability is not weak, but in the past it lagged half a step in organizational mechanisms. This adjustment is a belated fix, but not too late.
The real test Hunyuan 2.0 faces is whether it can support Tencent’s AI application ecosystem. At present, Tencent has already deployed large models in more than 700 internal scenarios, including Tencent Video, Tencent Meeting, and Tencent Docs, but a consumer killer app still has not emerged. In early 2025, the traffic surge Yuanbao gained by integrating DeepSeek proved the instant pull of a high-quality foundation model on products. However, the product’s functionality has long remained at the level of a “Q&A tool,” exposing shortcomings in product definition and ecosystem-building capability. The lack of platform capabilities and user-retention infrastructure—such as an Agent store or a plugin system—makes it difficult for the product to form network effects and user habits, leading to a rapid fade in popularity and poor retention. Yuanbao’s ups and downs prove that external models can only bring short-term growth; long-term competition must be built on the continuous evolution of one’s own foundation model.
Tencent AI’s Gains and Losses: A Dialectic of Moats and Stumbling Blocks
Strengths: A Rich Mine of Scenarios and Data
Tencent’s moat lies in its hard-to-replicate commercial ecosystem. WeChat’s 1.3 billion monthly active users, QQ’s differentiated user base, and content scenarios such as Tencent Video, Tencent Music, and Tencent Games together form the testing ground for AI deployment. The private-domain data produced in these scenarios, if effectively utilized, will generate a massive flywheel effect.
Yao Shunyu’s “AI Second Half” theory emphasizes “value definition in real scenarios,” which is exactly Tencent’s strength. Compared with purely technology-driven startups, Tencent understands users, products, and commercialization better. Once foundation-model capability is filled in, its explosive power at the application layer should not be underestimated.
Weaknesses: Organizational Inertia and a Personnel Ceiling
Tencent’s challenge lies in organizational inertia. As a company strong in products, its technical culture has historically leaned toward “rapid iteration and small, quick steps,” which creates tension with the long-termist research required by large models. Although this organizational restructuring centralizes power in Yao Shunyu, whether it can truly break departmental walls and achieve efficient resource allocation still needs time to be tested.
In addition, Tencent’s AI talent structure faces a “gap between generations.” Veteran executives lack large-model research backgrounds, and although rising talents have been promoted, whether they can steer Tencent’s massive ship remains uncertain. Yao Shunyu is only 27; although he is a technical genius, he still needs to prove himself in managing large-scale teams and coordinating cross-BG resources.
Opportunity: The Next Battleground of Agents and Multimodality
Yao Shunyu’s deep participation at OpenAI may allow Tencent to seize first-mover advantage in the Agent track. His “Second Half” theory clearly states that building evaluation systems is the core competitiveness, and this is exactly the key to Agent deployment. The number of Hunyuan derived models has already reached 3,500; if Tencent can embed Agent capabilities into super apps such as WeChat and WeCom (Enterprise WeChat) in the future, the imagination space is huge.
Threat: Dimensionality Reduction Attacks From Open-Source Models
DeepSeek’s success proves that open-source models are rapidly narrowing the gap with closed-source models. Although Tencent has Hunyuan, how to maintain competitiveness in the wave of open source is a huge challenge. If self-developed models cannot continue to lead, Tencent may become an “application-layer pipeline,” losing technological discourse power.
Tencent’s AI strategic adjustment is a belated all-out sprint, and also a clear-eyed self-revolution. Yao Shunyu’s arrival and the release of Hunyuan 2.0 mark that Tencent has finally put down its obsession with the comfort zone of “traffic + investment” and begun to rebuild its technical foundation with large models at the core.
The success or failure of this transformation does not lie in whether it can surpass ByteDance or Alibaba in the short term, but in whether it can convert Tencent’s unique scenario advantages into a data flywheel, and land the “AI Second Half” theory into a sustainable evaluation and iteration system. Whether 27-year-old Yao Shunyu can steer this giant and achieve a harmonious resonance between technological idealism and commercial realism will determine Tencent’s position in the technology landscape of the next decade.
The AI race has no final outcome—only a constantly refreshed starting line. For Tencent, the real “second half” has only just begun.
[Disclaimer]: The above content reflects analysis of publicly available information, expert insights, and BCC research. It does not constitute investment advice. BCC is not responsible for any losses resulting from reliance on the views expressed herein. Investors should exercise caution.