Ma Huateng believes that many companies are too impatient, a style Tencent has never adopted. "Initially, we thought AI was a once-in-a-decade opportunity for the internet. The more we think about it, the more we feel it is a once-in-several-centuries, industrial revolution-level opportunity comparable to the invention of electricity," Tencent CEO Ma Huateng predicted at a shareholders' meeting two years ago, accurately pinpointing the disruptive value of AI technology. While he acknowledged that AI would spawn new business models and social structures, he then shifted his tone, pointing out that for an industrial revolution, producing a light bulb one month earlier is not that significant in the long-term perspective. "Currently, we are still doing some thinking. Many companies are too hasty now, which has never been our style." This statement foreshadowed Tencent's "slower pace" in the fiercely competitive AI landscape of 2024-2025. While leading players rushed to launch first and compete on model parameters, Tencent consistently failed to demonstrate model capabilities matching its scale—it wasn't until 2025 that Tencent's AI strategy evolved from individual products towards a fully integrated ecosystem:
In July 2025, Yuanbao率先 integrated with QQ Music to enable "search and listen," subsequently building a "search-chat-watch" viewing loop with Tencent Video, and also jointly launched "AI Minutes" and "AI Hosting" features with Tencent Meeting. In September 2025, Yuanbao was fully integrated into Official Accounts and Video Account comment sections, thereby penetrating dozens of Tencent's core applications and covering key scenarios like social interaction, office work, and consumption.
Even though Yuanbao significantly increased its presence within the Tencent ecosystem, the real strategic upgrade came from the latest organizational restructuring: on December 17, 2025, Tencent upgraded its large model R&D architecture, newly establishing AI Infra Department, AI Data Department, and Data Computing Platform Department—the key driver behind this adjustment was Yao Shunyu, a top talent who had just joined Tencent. Yao Shunyu, a graduate of the Yao Class at Tsinghua University, joined OpenAI in 2024 and became a core contributor to the intelligent agent product operator and deep research. After joining Tencent, he assumed the role of Chief AI Scientist in the "CEO/President's Office" (reporting to Martin Lau), while concurrently serving as head of the AI Infra Department and Large Language Model Department (reporting to Lu Shan)—holding multiple roles to concentrate efforts, effectively establishing the large model's status as the group's strategic core from a corporate level. According to information obtained, Tencent's new structure has clear divisions: the AI Infra Department focuses on core technologies like distributed training and high-performance inference, building the technical foundation for large model R&D; the AI Data Department is responsible for data and evaluation system construction; the Data Computing Platform Department builds an intelligent integration platform for big data and machine learning. On the business front, Wang Di continues as Deputy General Manager of the Large Language Model Department, reporting to Yao Shunyu; Liu Yuhong and Chen Peng were appointed heads of the AI Data Department and Data Computing Platform Department respectively, both reporting to Jiang Jie, comprehensively strengthening organizational synergy. A Tencent technology expert revealed, "After Yao Shunyu joined, the density of senior AI research talent introductions at Tencent increased significantly, with the core goal of supporting the technical iteration and scenario implementation of the Hunyuan large model." It can be said that this "belated" architectural adjustment continues Tencent's engineering advantages and can improve large model R&D efficiency; however, in this AI攻坚 battle, Tencent's shift in attitude from individual products to a corporate-level number one project was a full year slower than ByteDance's adjustment speed. Did Tencent misjudge the situation? Tencent's "slowness" might stem from an initial misjudgment of strategic priorities. In early 2024, Ma Huateng's comments on traditional businesses indicated the priority of management's focus—his attention remained concentrated on Video Accounts and the gaming business; whereas ByteDance CEO Liang Rubo had already reflected on "big company disease" at an All Hands meeting,直言 stating that ByteDance was slow to react in the AI wave. The strategic differences between the two were directly reflected in their 2024 business layouts: Tencent, on one hand,持续推进 the construction of the WeChat transaction ecosystem and Video Account iteration, while on the other hand, focused on the rebound of its gaming business; in contrast, ByteDance's core group executives personally participated in Seed team reviews, deeply主导 the AI technology roadmap, model strategy, and前沿课题 planning, quickly switching from previous sluggishness to a full-catch-up mode. Simultaneously, in 2025, the Seed team completed the integration of several departments including AI Lab, with Wu Yonghui taking the "number one position," comprehensively overseeing large model basic research and application implementation, demonstrating a strategic决心 to "win the battle in one move." Thereafter, ByteDance's large model narrative transformed into a "comeback story," and Doubao's competitive advantage gradually expanded. A large model practitioner stated that ByteDance demonstrated systematic execution capabilities in the AI field,核心体现在 three aspects: accurately grasping strategic direction, rapid iteration capability in response to the market, and efficient allocation of R&D resources—Tencent, being slower to realize,虽然 achieved the first two points,但在 the most critical resource allocation and talent configuration, remained indecisive until 2025 when it gradually拆分出独立的 LLM, Infra, and Data teams. As one of the earliest manufacturers to布局 large models, Tencent's R&D investment from 2018 to date exceeded 400 billion yuan, having established top AI teams and labs early on; but over the past two years, while major manufacturers纷纷 competed for first launches,卷 model parameters, and抢占 ChatBot application entry points, Tencent始终 gave the impression of being "in no hurry,"甚至 being questioned by the market for "completely falling behind"—since the organizational restructuring in early 2025, to Yuanbao integrating DeepSeek, Tencent had no particularly noteworthy performances in this AI battle. However, the rules of AI competition are quietly being重构: over the past year, large model technology has significantly improved inference efficiency through reinforcement learning and knowledge distillation, particularly with the "low-cost + open-source" approach represented by DeepSeek, breaking the previous reliance on the "large parameters, large computing power, large investment" path, reshaping the global AI competitive landscape and ecosystem expansion pace—this also provides an opportunity for Tencent to catch up later. Regarding this, Tencent President's Office Chief AI Scientist Yao Shunyu believes, "The AI industry is quietly entering the second half. The first half focused on model training, with manufacturers crowding to pursue parameter scale and performance breakthroughs, and technical焦点集中在 Transformer architecture, deep reinforcement learning, and large-scale pre-training; in the second half, as model capabilities mature, the competitive focus will shift to task definition, system construction, and evaluation systems—evaluation is more important than training. The value of AI needs to measure practical problem-solving ability through quantitative indicators, rather than单纯 pursuing model scores." On January 10, 2026, Yao Shunyu stated in his first public speech that the differentiation in AI application落地 in To C and To B fields is becoming increasingly apparent. "When people think of AI super applications, the first that come to mind are ChatGPT and Claude Code, which are paradigms for To C and To B respectively. Today, most people use ChatGPT in ways not much different from last year, perceiving no significant changes; but in the To B field, this change is substantial." He举例称, "Perhaps a year ago the Claude Code revolution hadn't started, but now it is reshaping the entire computer industry's work methods—people no longer write code directly, but communicate with computers in English. For To C scenarios, most users feel AI is more like an 'enhanced search engine,' often not knowing how to激发 its core capabilities; but the logic in To B scenarios is completely different: the higher the AI intelligence, the higher the productivity and商业收益, making enterprises willing to invest resources to use the strongest models." This judgment is profoundly reflected in Tencent's product落地. On the C-end side, Yuanbao maintains high-frequency iteration, updating almost daily initially at launch, with user scale steadily ranking among the top three domestic AI applications; in May 2025, QQ Browser was upgraded to an AI browser, providing 400 million users with functions like intelligent summarization, voice broadcast, and personalized subscriptions, achieving an 80% success rate among users trying the download Agent; Sogou Input Method's AI search, AI writing assistance, and AI emoji functions have penetrated hundreds of millions of users, with daily recommended emoji volume reaching hundreds of millions. On the B-end side, the Hunyuan large model has attracted over 150 enterprises to access via Tencent Cloud, covering industries like game production, e-commerce display, and film special effects, including Unity China and拓竹科技; the knowledge base product ima integrates with ecosystems like腾讯新闻 and QQ Browser, supports multi-format file imports, and can generate reports, podcasts, and other content through Agent capabilities, with MAU in September 2025 growing over 80 times compared to January, and the total number of knowledge base files exceeding 200 million. A Tencent-related source revealed that Yao Shunyu has increased discussions about future directions in recent internal meetings. "Next, internally we will emphasize a cultural shift from delivery to Co-design, promoting the integration of Infra, algorithms, and products to enhance R&D efficiency—this is also the core background of the recent large model R&D architecture upgrade,同时 we will further strengthen the deep linkage between models and products." A Comprehensive AI Transformation Simultaneously, a comprehensive AI transformation is advancing within Tencent. According to information obtained, Tencent's newly released AI CLI形态 product CodeBuddy Code has 90% of its code self-generated, supports deep integration with enterprise-level services, with capabilities comparable to Claude Code, and currently covers over 12,000 engineers internally at Tencent; by the end of 2025, the Hunyuan large model had been implemented in over 900 internal applications包括腾讯会议, WeChat, advertising, and games. Among these, advertising benefited most立竿见影. As AI large models comprehensively penetrated advertising and cloud businesses, they directly drove high growth in advertising revenue, particularly in Video Account traffic and ad inventory growth—in Q3 2025, Tencent's advertising revenue growth reached 21%, a new high in the past six quarters, with increased投放 from advertisers in major industries—this was not only due to increased ad load rates but also源于 AI-driven ad targeting technology boosting eCPM growth. To this end, Tencent specifically mentioned the "Tencent Advertising AIM+" intelligent投放 product matrix in its financial report: it supports advertisers in automatically configuring targeting, bidding, and placements, and optimizing ad creatives, thereby improving marketing ROI. Data obtained shows that through AIM+, for every 10,000 yuan advertisers spend, operational actions on the ad platform decrease by 80%, and operational actions in the creative环节减少 by 47%. Beyond businesses like WeChat, advertising, and gaming, key business lines such as Tencent E-Signature, Tencent Le Xiang, intelligent customer service, and智能办公 have also been fully AI-ized—Tencent is undergoing an AI "gene recombination" and spreading replicable methodologies to external industries. For example, in落地 scenarios, Tencent deliberately avoids "novelty" scenarios, instead切入深水区: from pathological analysis in medical imaging to intelligent cockpits for tens of millions of vehicles, testing not only the model's IQ but also the system's stability, which恰恰 is where Tencent's systematic engineering沉淀 shines.
In early 2025, DeepSeek-R1 achieved breakthroughs in long-text and reasoning capabilities, creating a short-term experience gap and winning widespread user favor, but the短期 user surge also posed significant challenges to its product experience. Tencent's Yuanbao rapidly integrated the "full-blooded version" of DeepSeek in February. It is understood that the Yuanbao team conducted extensive underlying optimizations based on DeepSeek, maintaining a "daily update level"冲刺 rhythm in the first month, iterating hundreds of versions over six months, performing deep tuning for inference in high-concurrency scenarios, effectively reducing stuttering and latency, and establishing an "open source即上线" rapid response mechanism to ensure users always experience the latest and smoothest model version. Unlike other manufacturers betting on单一 AI applications, Tencent's AI approach is more ecosystem-centric: on one hand, creating AI-native超级入口 like Yuanbao; on the other hand, accelerating the integration of AI capabilities into national-level products like WeChat, QQ, Docs, and Meeting—if Hunyuan is the kernel of Tencent AI, then products like Yuanbao, ima, and QQ Browser are the touchpoints that make the ecosystem flow seamlessly. In Tencent's规划, Yuanbao is not just a standalone APP for users but also an internal "super connector"—it is deeply embedded in high-frequency scenarios like WeChat, Tencent Meeting, and Tencent Document in a distributed manner. This "embedded" approach allows users to naturally encounter AI without changing usage habits, enabling cross-domain flow of data and capabilities, and providing the most natural, low-barrier service experience. For example, in C-end scenarios, Yuanbao launched an AI recorder,打通 internal service chains; in Tencent Meeting, it can automatically attend meetings, take minutes, and提炼 action points; in Video Account, Official Account, and腾讯新闻 comment sections, users can directly @Yuanbao to ask questions and extend discussions. More importantly, dual advantages at the experience and data layers will持续强化 users' "sense of gain," and as usage frequency increases, this advantage will become increasingly apparent. Third-party data indicates that in the next two years, enterprises deploying Agents will double, and GenAI-related IaaS expenditure growth will reach 192%. With the explosive growth in the number of Agents, enterprise demand for cloud computing infrastructure will upgrade from "resource supply" to "business value"—AI Infra must possess faster inference efficiency, more flexible tool integration, more reliable system guarantees, and more automated service capabilities. A cloud computing entrepreneur revealed that the advantage of Tencent Cloud's intelligent computing system lies in its "same-source, same-structure": there is no need to build dedicated AI facilities like专属 storage separately; it can directly leverage Tencent's accumulated cloud-native technical capabilities, seamlessly承接 intelligent computing needs with "one cloud, multiple cores," achieving deep synergy between cloud computing and intelligent computing capabilities. Especially at the AI Infra foundation level, Tencent Cloud's intelligent computing relies on深厚积累 and scenario optimization of infrastructure like cloud computing, distributed storage, and high-performance networks, significantly improving the performance and utilization of cloud resources, ultimately achieving cost reduction and efficiency gains: model startup speed increased 17 times, large-scale service scaling time reduced from 10 minutes to 34 seconds; self-developed inference engines cover text, image, and video generation models, with multi-modal inference accelerated by 4 times. It was learned that Tencent Cloud successfully achieved its profitability target in 2025, primarily through two major strategies: first, focusing on actual customer needs, launching products and tools深入业务场景, rather than单纯 pursuing token调用 indicators, a pragmatic strategy driving steady AI business growth; second, leveraging ecosystem advantages, opening AI products and atomic capabilities to partners, with revenue from ecosystem partners maintaining double-digit growth. Of course, for intelligent agents to move from the lab to production, engineering and security issues are core pain points. Tencent's newly released Agent Infra solution, Agent Runtime, integrates five components: runtime engine, cloud sandbox, context service, gateway, and security observability; among them, the cloud sandbox achieves millisecond-level startup, supporting hundreds of thousands of instances秒并发. Effectively, Tencent Cloud's intelligent computing,面向更贴近 Agent的 AI Infra, has built a complete system of "Agent Infra solution + Cloud Mate cloud expert service agent + full-link security capabilities," helping Agentic AI move from the lab to the front lines. Tencent Returns to its "Comfort Zone" Undoubtedly, entering the second half of AI, technological gaps are gradually narrowing, and the core of competition shifts to experience, engineering, and systematization—this恰恰 is Tencent's "comfort zone" for a later-mover advantage: from the PC internet, mobile internet, to industrial internet, Tencent has always managed to stage "coming from behind" performances at different development stages,核心源于 three points: First, converting infinitely reusable channel potential into product power. Simply put, this means transforming complex and obscure technology into functional products that users find easy and convenient to use—in the AI era, this means not单纯 pursuing parameter leadership, but focusing on solving the "last mile" of implementation, allowing technology to truly serve people. Second, relying on the real usage data of billions of users to establish the industry's fastest "feedback-correction" mechanism. It is understood that Debug culture is not only part of Tencent's engineer culture but has also become an important mechanism in Tencent's organizational management. Third, rapidly building a top-tier talent梯队 and flexible organizational structure through "financial capability." In past waves of development like the mobile internet and industrial internet, industry top talents have successively become the "fuel" for Tencent's progress under the allure of high salaries. The recent重组 of the Hunyuan team and the high-density talent introduction本质上 reflect this organizational flexibility.
Currently, WeChat is持续引入 Yuanbao capabilities, including using AI to enhance search functions. As an AI assistant residing in WeChat, Yuanbao is equipped with dual engines, Hunyuan and DeepSeek, capable of一键解析 Official Account articles, images, and various documents, using "@summon + precise questioning" as the core usage method, deeply integrating into the WeChat Official Account and Video Account ecosystem. Recently, the "Yuanbao" contact within WeChat underwent an important upgrade: users simply need to send text or voice messages to it, and Yuanbao can intelligently识别 tasks and time, quickly calling WeChat's native reminder function, supporting "one-sentence reminder setting." More critically, as an ecosystem connector, Yuanbao will持续深度嵌入 high-frequency scenarios of Tencent products, making AI capabilities ubiquitous. A Tencent insider分析: "Traditional Chatbot usage scenarios are often limited to specific search or creation needs, with普遍存在 bottlenecks in user retention and daily average usage time (DAU Time Spent)—whether Silicon Valley giants or domestic AI application manufacturers, the industry consensus is: Chatbot is not the ultimate application form of AI. The next step for AI lies in超越 'dialogue.' This is the direction currently explored by domestic and foreign manufacturers,但 paths are gradually diverging—the new version of Yuanbao will, on the basis of tool attributes, increase interaction depth by incorporating certain social relationships. From QQ to WeChat, the基因 of connection naturally flows in Tencent's DNA." Furthermore, Martin Lau revealed during the Q3 2025 earnings call that WeChat will eventually launch an AI agent, allowing users to complete multiple tasks using AI within WeChat—this agent will be able to understand user needs, intentions, and interests. He emphasized that WeChat, with its powerful social ecosystem叠加 shopping and payment scenarios, is almost the user's "ideal assistant." As Ma Huateng previously stated, AI is a major opportunity comparable to the industrial revolution. In the second half of AI, when technological gaps are no longer the core barrier, and user experience, engineering, and systematization become the keys to competition, Tencent's ecosystem advantages, engineering底蕴, and scenario accumulation恰恰 provide the confidence for a later-mover advantage.
Comments