Tencent's formula for success may not be flashy, but it is consistently sound.
Last Friday, a dialogue between Tang Daosheng and Yao Shunyu once again pushed the narrative that "Tencent is slow" into the spotlight. When asked why the outside world perceives them as slow, the two shared a knowing smile but offered no answer. As this notion shifts from industry chatter to market consensus, Tencent appears to have yet to fully grasp its potential impact on the company.
The most direct manifestations are on two fronts: first, the media's persistent pessimism regarding Tencent's AI narrative; second, the volatility and decline in its Hong Kong stock price in the first half of the year, which saw the company temporarily fall from its status as a Chinese internet "evergreen" to being labeled a legacy stock. Capital expenditure tells a clear story: while global tech giants readily invest a year's worth of operating cash flow into AI capex, Tencent allocated only 35% of its operating cash flow to AI last quarter.
Simultaneously, WorkBuddy has become the top productivity agent in China's Claude Code-like market, Hy3 Preview has ranked first in OpenRouter call volume for three consecutive weeks, CodeBuddy is used by 90% of Tencent engineers and has cut coding time by 40%, and Tencent Cloud achieved its first full-year profit in 2025.
Caught between external criticism of being "slow and unaware" and steady progress in product deployment, Tencent's AI strategy seems to operate outside the mainstream industry narrative. This gap between perception and reality forces a re-evaluation: why did Tencent seem slow in the first half of the AI race? And what is Tencent competing for, if not speed, while others scramble for launches, chase leaderboards, and burn through capital?
Is Tencent Unaware of Its Own Pace?
Criticism of Tencent's "slowness" centers on three dimensions: a late start on foundational models, a lack of breakout C-end products, and a non-aggressive approach to computing power investment.
The industry's rush to release large language models in 2024 was essentially a sprint at the end of the "method-hunting" first half, before true value creation had begun. After the Hunyuan team completed its architectural rebuild in 2026, Tencent focused on solving two core problems: rebuilding the infrastructure for large models and improving data quality.
The release of Hy3 Preview is a result of this phase, prioritizing optimization of infrastructure, data systems, and real-world scenario adaptability over chasing extreme parameter counts. Yao Shunyu stated plainly, "Today, building a large model itself is relatively straightforward; algorithms are no longer the core bottleneck. The real difficulty lies in building solid infrastructure and thoroughly developing data and evaluation systems."
In hindsight, the strategy of prioritizing a strong foundation over a rushed launch helped Tencent avoid the industry-wide "leaderboard trap" during subsequent iterations of its Hunyuan model. Many domestic LLM vendors became trapped in a cycle of optimizing models specifically for benchmark tests, leading to impressive scores but poor performance in real applications.
In his dialogue with Tang Daosheng, CEO of Tencent's Cloud and Smart Industries Group, Chief AI Scientist Yao Shunyu clearly defined the second half of the AI race: "Before last year, decades of AI development were about finding methods to solve problems. But with the maturity of pre-training and post-training, we now have a 'universal hammer.' The harder part is finding the good problems worth solving."
As for the frequent jibes about Tencent's Yuanbao having only a fraction of Doubao's monthly active users, one must ask: in C-end conversational AI, is high traffic truly an asset or a liability?
Doubao's rapid growth was built on sustained spending, as each user interaction incurs token inference costs, leading to exponentially higher computing power consumption with scale. Foreign media reported that Doubao's aggressive spending strategy once prompted ByteDance to raise its 2026 capital expenditure forecast to $70 billion, significantly impacting overall profitability.
In this context, Doubao's push for C-end subscriptions appears more a reactive choice driven by revenue pressure. Data from tracking firm aicpb.com shows that after Doubao was reported to be introducing subscription options, its May MAU fell by 6.1 million—a rare decline since its 2023 launch.
Globally, conversational AI has yet to find a sustainable path to mass-market commercialization. While OpenAI's ChatGPT mobile app surpassed 1 billion MAU, its 2025 net loss reached $8-9 billion, with a projected 2026 loss of $14-15 billion. In contrast, Anthropic, with only about 10-15% of OpenAI's total platform active accounts (134 million in March 2026), achieved an annualized revenue run rate of approximately $45 billion in May 2026. Its ARPU is roughly 10-12 times that of OpenAI, a result of strategically contracting from the mass market to focus on high-value enterprise and professional service scenarios.
Tencent has effectively followed Anthropic's path in China. Yuanbao did not blindly chase C-end user scale but instead serves as a testing ground for the Hunyuan model and a C-end entry point, refining foundational model capabilities while focusing its real efforts on the high-value productivity agent sector.
The evolution from CodeBuddy to WorkBuddy exemplifies Tencent's product logic: first honing a mature Agent core in the code domain, which is formally verifiable and has clear value, then migrating it to general scenarios like office work and design.
Thus, Tang Daosheng emphasizes, "AI is a marathon, not a sprint." Yao Shunyu confidently adds, "If the second half of AI has just begun, then past speed is meaningless. What matters is the ability to be honest with oneself and maintain long-term patience."
The Core Tencent Strategy: Competing on Real-World Implementation
In Tencent's view, large models themselves do not generate commercial value; only models implemented in real scenarios to solve actual problems hold value. This philosophy centers on the "Co-Design" model repeatedly mentioned by Tang and Yao: deep collaboration between model teams and product teams to jointly define problems, design solutions, and evaluate outcomes.
Co-Design is not simply about model teams providing API interfaces to product teams. Yao Shunyu summarizes it with three core points.
First, the model itself must be robust. Pre-training is a universal foundational task unrelated to specific products and must be executed to the highest standard; its improvements deliver continuous value to all downstream tasks. Post-training must abandon leaderboard-chasing, building evaluation systems based on real product scenarios. "Practical value far outweighs leaderboard value" is a core tenet of Tencent's AI team.
Second, building trust between model and product teams is the hardest part of Co-Design. Model teams aim for maximum capability, while product teams seek to precisely meet user needs—goals that are naturally divergent. Tencent's solution is empathy and deep integration: the Hunyuan team dispatched core post-training personnel to support Yuanbao's post-training work, prioritizing its technical support even during incomplete pre-training phases.
Third, fully leverage the generalization capability of large language models. Unlike traditional AI, LLM-era capabilities are transferable. The chat and search capabilities refined through Co-Design between Yuanbao and Hunyuan can be directly applied to products like the Intelligent Marketing Assistant and WorkBuddy. Office scenario data accumulated by WorkBuddy can, in turn, feed back into iterating Hunyuan's general capabilities.
Consequently, while the industry remains preoccupied with "whose model is smarter," Tencent has shifted its focus to Agents. Yao Shunyu believes, "Agents, or Coding Agents, are a foundational capability that must be contested. A Coding Agent is essentially Turing-complete; when you can control file systems and containers, you possess a complete automation system."
Within Tencent's AI system, the evaluation system is the crucial link connecting models and products. Yao Shunyu states, "External benchmarks have value but are highly susceptible to overfitting. Truly valuable evaluation comes from real-world user feedback."
Unlike the industry's general reliance on third-party leaderboards, Tencent has established an internal evaluation system based on product scenarios, with core value in three areas.
First, it uncovers fundamental experience flaws and security risks that third-party benchmarks miss. A key purpose of releasing Hy3 Preview was to gather large-scale real user feedback to fix latent vulnerabilities not exposed in benchmark tests.
Second, it captures the distribution of real-user prompts. Benchmark questions are often precise and logically clear, but real-user queries are frequently ambiguous, logically jumpy, and involve multi-turn dialogues. Analyzing prompts from Yuanbao and WorkBuddy users allows Tencent to more precisely optimize model instruction-following and multi-turn dialogue coherence.
Third, real-scenario user demands can drive iterative changes in technical research direction. For instance, strong user demand for long context in Yuanbao directly spurred Tencent's technical breakthroughs in long-context learning, while WorkBuddy users' need for complex task automation effectively promoted the enhancement of agents' complex task planning abilities.
Many misjudge Tencent's AI strategy because they still measure the AI industry with traditional internet thinking, believing in a winner-takes-all logic where traffic is the moat. However, while the internet's moats are traffic and network effects, the core moats in AI are scenario depth, data quality, and organizational synergy.
Today, the phrase "whoever owns the scenario owns the future" is being validated in AI's second half.
When joining Tencent, Yao Shunyu remarked, "The most important thing in AI's second half is context. As models become increasingly adept at transforming complex inputs into standardized outputs, the competitive barrier shifts to the most primitive input: do you know what this person is doing, do you know the various pieces of information about the enterprise."
TENCENT possesses China's richest and most complete ecosystem of scenarios: WeChat is a super social gateway covering 1.4 billion Chinese users; Tencent Docs and Tencent Meeting are national-level office tools; its gaming business reaches billions globally; and Tencent Cloud serves millions of enterprise clients. These scenarios generate massive, high-value structured and unstructured data daily, providing unique contextual data support for large models.
More importantly, Tencent's scenario ecosystem is interconnected and synergistic. Through gateways like WeChat and WeCom, Tencent can deeply embed AI capabilities into users' daily work and life, achieving seamless AI integration. For example, a user sending a natural language instruction in WeChat could trigger WorkBuddy to automatically draft a preliminary plan in Tencent Docs, initiate an online discussion via Tencent Meeting, and finally sync the result to a WeCom approval workflow.
At a higher level, competition in the AI industry ultimately becomes competition between organizations. Large model R&D requires deep, cross-disciplinary collaboration; agent products need rapid iteration and agile experimentation. This places extremely high demands on a company's organizational structure and cultural DNA.
Yao Shunyu admits that the most important reason for choosing Tencent was its culture of "candor, trust, and low ego." He recalls, "Chatting with members of the top management for the first time, everyone was very honest—what was done well, what wasn't—very direct, with no covering up. Tencent is a company that operates on trust, not just KPIs, which is crucial for AI work."
Under this cultural influence, Tencent Cloud's teams exhibit two distinct organizational characteristics.
First, products like WorkBuddy and CodeBuddy adopt elite small-team models of 3-5 people. Team members have significant autonomy for quick decision-making and iterative experimentation, free from complex hierarchies and reporting lines. Tang Daosheng comments, "In today's AI R&D, ideas are more important than processes. Small teams can respond more flexibly to market changes and validate product directions faster."
Second, traditional waterfall development processes have been restructured. Boundaries between product managers, engineers, designers, and testers have blurred; everyone participates in requirement definition, architecture design, and quality evaluation. Engineers are no longer just coding but have become conductors directing AI. Testing has also moved forward from the final stage to the requirements phase, with automation via AI.
Additionally, Tencent has built a complete pathway from individual efficiency gains to organizational capability upgrades. Liu Yi, Vice President of Tencent Cloud and head of WorkBuddy, summarizes this in four core steps: cultivating super individuals, accumulating organizational assets, integrating into production workflows, and building AI governance systems.
Of course, the industry's biggest pain point remains unclear commercialization paths. Tencent's AI commercialization is understood to operate on three levels. First, AI serves Tencent's own business to improve R&D, operations, and management efficiency. Second, AI capabilities are delivered to enterprise clients via Tencent Cloud through MaaS, SaaS, and other services—unlike traditional cloud services, AI services offer higher stickiness and added value, potentially yielding higher margins for Tencent Cloud. Third, Tencent provides partners with computing power, models, tools, and traffic support, while partners handle scenario implementation and client expansion, sharing profits per agreed ratios. This model can rapidly expand the market and enrich Tencent's AI ecosystem.
Regarding the industry's prevalent per-token pricing model, Tang Daosheng stated, "Pure per-token billing is not a sustainable long-term business model. If AI creates tangible value for clients, they will naturally pay for the outcome. Tencent is not opposed to per-token pricing but views it as one foundational billing method, combined with per-task, outcome-based, and other models to meet diverse client needs."
The Cost of Slowness and Cards Yet to Be Played
Acknowledging Tencent's strategic resolve does not imply an absence of problems. In fact, Tencent's pace has come at a cost, which will continue to affect its competitiveness in AI's second half.
For instance, Tencent started late on foundational models. After Yao Shunyu joined, he essentially "rebuilt" Hunyuan. While Hy3 Preview shows significant progress, a gap remains compared to top international models like GPT-4o and Claude 3 Opus.
This gap is evident in two areas: first, the comprehensiveness of general capabilities, where top international models still lead in complex reasoning, multimodal understanding, and long-context processing; second, exploration of cutting-edge technologies, where Tencent's investment and accumulation in AGI foundational research trail companies like OpenAI and Anthropic.
In his dialogue with Tang Daosheng, Yao Shunyu conceded, "China is not doing enough cutting-edge exploration today. We need to inject more of that spirit of frontier exploration into our organizational DNA."
A more critical bottleneck is computing power, currently Tencent's biggest challenge. In 2024, Tencent procured about 230,000 Nvidia H20 chips, tying with ByteDance for the largest domestic purchase volume. However, after the US imposed stricter export controls on H20 chips in April 2025, Tencent could no longer procure high-end training chips in bulk. Mass production of domestic Ascend 950 training chips is not expected until Q4 2026, a distant solution for an urgent need.
Tang Daosheng admitted, "For the past two to three years, our computing power has been in a tight balance. Limited resources had to be prioritized for internal core product needs."
Thus, insufficient computing power will be the core bottleneck constraining Tencent's AI development over the next 18-24 months, slowing the training and iteration of the Hunyuan model, tightening supply for products like WorkBuddy and Yuanbao, and impacting enterprise client demand on Tencent Cloud. In response, Tang stated in the interview that Tencent is increasing investment in diversifying its computing power supply.
It is noteworthy that while scenarios and data are Tencent's strengths, its multi-business group (BG) structure also creates internal collaboration barriers. Although the Hunyuan team has achieved cross-BG resource mobilization, significant obstacles remain in integrating products and data across different BGs. For example, WeChat's social data, gaming user data, and cloud enterprise data are not yet interconnected, hindering synergistic force. This restraint in data integration is understood to be primarily due to compliance considerations.
Tang Daosheng is aware of this issue and has begun promoting the integration of Tencent's AI products. Liu Yi explicitly stated, "In the future, WorkBuddy will become the unified entry point and platform for Tencent Cloud AI agents, integrating scenarios like code, office work, and design."
Currently, WorkBuddy already integrates multiple mainstream large models including Hunyuan, DeepSeek, GLM, Kimi, and MiniMax, allowing users to choose freely based on needs. This not only meets diverse user requirements but also reduces dependency on any single model, dispersing risk.
However, Tencent still holds the card of WeChat AI, which could become a market-disrupting "trump card" at any moment. Industry sources indicate WeChat is collaborating with several smartphone manufacturers to open its social relationship chain capabilities to system-level AI assistants. This means users could potentially access an AI chat window via a swipe on the home screen, with instructions automatically calling upon millions of WeChat mini-programs, potentially making WeChat the world's largest agent platform.
Therefore, in the AI marathon, Tencent is not racing for momentary speed but betting on scenario implementation and depth of value. Unlike many companies attempting to build a single AI super-app, Tencent's strategy is "AI everywhere"—embedding AI capabilities into all existing products to make each one smarter.
The core advantage of this strategy is the ability to fully leverage Tencent's existing massive user base and mature scenario ecosystem for rapid, large-scale AI deployment. For instance, WeChat's ad delivery system is fully integrated with the Hunyuan model, significantly improving ad precision and conversion rates. Video Channels' content recommendation algorithms and creation tools are continuously optimized with AI. Features like automatic meeting minutes and real-time translation in Tencent Meeting have become indispensable core functions for users.
In summary, the foundation of Tencent's AI strategy is a continuation of its successful "latecomer" logic from the internet era: from QQ Mail to WeChat to Video Channels, Tencent has never been the first mover but has consistently caught up through sustained product iteration and powerful ecosystem advantages. Whether this strategy succeeds again in AI's second half depends on Tencent's ability to break through computing power bottlenecks, dismantle internal data barriers, and invest more resources in frontier technology exploration.
Comments