Summary of Key Points
The outside world generally believes that Tencent is "lagging behind" in the field of AI—starting late with basic models, failing to launch popular consumer-facing products, and being conservative with its investment in computing power. However, Tencent is actually pursuing a "marathon-style" strategy: avoiding the costly pitfalls associated with consumer-facing applications, focusing on the practical implementation of AI in corporate and professional scenarios. By adopting a Co-Design approach that deeply integrates models with products, and leveraging its own ecosystem of services such as WeChat and Tencent Meeting, it aims to build long-term competitiveness. Nevertheless, Tencent faces challenges including gaps in basic model capabilities, limitations in computing power, and difficulties in coordinating internal data. It also holds a potential "trump card" in the form of WeChat AI.
The Outside World's Perception of Tencent's Slowness: Three Main Criticisms
The criticism of Tencent's slow progress focuses on three main areas:
1. Late Start with Basic Models: While the industry was rushing to release large-scale models in 2024, Tencent didn't complete the reconstruction of its Hunyuan architecture until 2026, lagging behind its competitors by two years.
2. Lack of Popular Consumer-Facing Products: Its AI chat tool, Yuanbao, has a much lower monthly active user base compared to DouBao, leading to suggestions that its consumer-facing capabilities are weak.
3. Conservative Investment in Computing Power: Last quarter, Tencent invested only 35% of its operating cash flow in AI, which is more conservative compared to global giants that devote their entire annual cash flow to this area.
Is Tencent's Slowness Deliberate? Avoiding the Consumer-Facing Traffic Trap?
Tencent's apparent slowness is not a result of inaction; rather, it is a deliberate choice to avoid the short-term traffic pitfalls associated with consumer-facing applications:
- Consumer Facing Applications as a "Cash Burn Hole": Services like DouBao rely on heavy funding to drive growth, which has led to ByteDance increasing its capital expenditure to $70 billion in 2026 and resulting in significant losses. ChatGPT has over a billion users but still incurs annual losses, with an unclear commercialization path.
- Following the High-Value Route of Anthropic: Anthropic focuses on corporate and professional scenarios, attracting only 10%-15% of OpenAI's user base, yet its ARPU (Average Revenue Per User) is ten times higher. Tencent is following this approach, with Yuanbao serving as a testing ground for models while focusing on high-value applications such as WorkBuddy (a productivity assistant) and CodeBuddy (an engineering tool).
- Tang Daosheng's Perspective: "AI is a marathon, not a 100-meter sprint; past speed is of no significance."
Tencent's Core Approach: Integrating Models and Products
Tencent's AI strategy relies on the Co-Design approach, which emphasizes close collaboration between model development and product teams:
1. Robust Model Development: Pre-training focuses on optimizing infrastructure and data rather than simply increasing the size of the parameter set. Post-training uses real-world data for refinement.
2. Building Trust: The Hunyuan team sends key members to support Yuanbao's post-training process, ensuring product requirements are met even before the pre-training is complete.
3. Transferability of Skills: Chatting capabilities from Yuanbao can be applied to WorkBuddy, and vice versa, creating a closed loop of improvement.
4. Evaluation System: Instead of relying on third-party rankings, Tencent uses user feedback to optimize its models—for example, the release of Hy3 Preview was aimed at collecting real user questions (which are often vague or irregular) to fix issues uncovered by traditional evaluation methods.
Tencent's Strengths and Weaknesses: Scenarios as a Advantage, Computing Power as a Limitation
Strengths:
- Rich Ecosystem: Services like WeChat (with 1.4 billion users), Tencent Meeting, games, and cloud services generate massive amounts of high-value data daily.
- Agile Organizational Culture: Small teams (3-5 people) make flexible decisions without complex hierarchies, allowing engineers to shift from coding to managing AI applications more efficiently.
Weaknesses:
- Gaps in Basic Models: Compared to models like GPT-4o and Claude 3, Tencent's models lag behind in complex reasoning and multimodal understanding.
- Computing Power Constraints: Restrictions on H20 chips in the U.S. and the delayed mass production of the domestic Shengteng 950 chip will create a shortage of computing power over the next 18-24 months.
- Internal Data Barriers: Data from WeChat, games, and cloud services is not easily integrated due to compliance issues, hindering collaborative efforts.
The Unrevealed "Trump Card": WeChat AI
Tencent has an untapped asset in WeChat AI:
- It is collaborating with smartphone manufacturers to integrate social relationship data into system-level AI assistants.
- Users can access AI features with a simple swipe of their fingers and use millions of mini-programs (e.g., generating documents or initiating meetings), potentially making WeChat the world's largest intelligent platform.
- Once fully implemented, WeChat AI could significantly transform the industry by seamlessly integrating its capabilities into the daily lives of 1.4 billion users.
Conclusion
Tencent's AI strategy reflects the "latecomer advantage" seen in the internet industry (from QQ Mail to WeChat). However, whether it can succeed in the second half of the AI race depends on its ability to overcome computing power limitations, integrate internal data effectively, and realize the full potential of WeChat AI. After all, the key to success in the AI marathon is long-term value, not short-term speed.