Summary of Key Points
Tencent has recently adjusted its AI strategy, shifting from a focus on "fundamental models + applications" to an emphasis on "computing power + agent fleets." On one hand, the company is investing heavily in computing power (with cash outflows related to this exceeding 110 billion yuan in a single quarter) to address the issue of "computing power scarcity." On the other hand, it is building a range of agent products tailored for specific use cases, such as WorkBuddy, CodeBuddy, and Xiaowei, leveraging the strengths of its WeChat ecosystem to gain a competitive advantage. Meanwhile, its former core businesses, Yuanbao Chatbot and Hunyuan Large Model, have been temporarily marginalized, with resources being reallocated towards agent development. Tencent aims to avoid direct competition with ByteDance and Alibaba in the chatbot and fundamental model areas, adopting a "pack of wolves" strategy to find its own unique path in the agent market.
1. Investment in Computing Power: From "No Cards Available" to "Stepping on the Gas Pedal"
Tencent's free cash flow turned negative for the first time last quarter, not because it couldn't generate revenue, but because it spent all its money on acquiring computing power. The details are as follows:
- Delivered Equipment: Capital expenditure amounted to 52.8 billion yuan (a year-on-year increase of 176%), which includes installed GPUs and servers.
- Actual Cash Payments: 59.3 billion yuan was spent on capital expenditures.
- Locking in Future Computing Power: AI prepayments totaled 51.4 billion yuan, with payments made in advance to secure future GPU resources.
The combined cash outflows for computing power in a single quarter exceeded 110 billion yuan—this is more than Tencent's annual capital expenditure in any year from 2019 to 2023 (the highest being 37 billion yuan).
Why such a significant investment? Last year, Tencent claimed it had sufficient computing resources; however, by March this year, it admitted that it could no longer purchase GPUs and had to rely on third-party services. Now, the company is using its financial strength to bridge the gap and even considers renting out its computing power for profit (Liu Chiping mentioned that some of the computing capacity could generate a 30% return). Nevertheless, the priority remains: first to refine the Hunyuan model, then to support agents like WorkBuddy, and only later to consider leasing the resources.
2. Agent Fleets: A "Pack of Wolves" Strategy to Cover All Workplace Scenarios, with WorkBuddy as the Flagship
Tencent does not aim to create a single, all-encompassing agent; instead, it develops multiple products for specific use cases, covering the needs of workplace users throughout the day:
- Flagship Product: WorkBuddy – An office AI tool that has exceeded 20 million monthly active users and 13 million daily active users during its public beta period. It has already started commercialization (with a gross profit margin similar to that of Tencent Cloud services) and is considered the company's leading AI product.
- Supporting Products: CodeBuddy, an AI programming tool that complements WorkBuddy; Xiaowei, which integrates with WeChat to help users summarize messages, update their social profiles, and manage mini-programs, effectively making WeChat the "command center" for these agents.
- Additional Tools: QClaw (for remote computer control), Marvis (PC system operations), Dayuan (AI for enterprise WeChat), Miora (creative design tools), etc.
- External Partnerships: Investments in Manus (to expand into cloud-based agents and overseas markets) and Pragmatik Labs (founded by Lin Junyang, to acquire the next generation of agent technologies).
All these products are integrated with Tencent's ecosystem. For example, WorkBuddy connects with Tencent Document and Meeting services, allowing users to access the entire suite of tools directly through WeChat—a unique advantage for the company.
3. Yuanbao and Hunyuan: From Leading Roles to Supporting Ones
Tencent's former core products are now playing a supporting role:
- Yuanbao Chatbot: Its features will be split into components for use in agents, such as sharing chat capabilities with WorkBuddy.
- Hunyuan Large Model: While it offers good value for money, it is not the only option for agents. WorkBuddy uses multiple models (including competitors like DeepSeek and GLM), and WeChat's own WeLM is preferred due to privacy and cost considerations. Liu Chiping mentioned that Hunyuan 4 is under development, but if it fails to catch up with the latest standards (SOTA), it may impact the performance of agents.
The reasons for this shift are practical: Yuanbao has only 50 million monthly active users, compared to DouBao's nearly 400 million; Hunyuan still lags behind the leading models. Tencent prefers to focus on its agent strategy and leverage its ecosystem.
4. Differentiation from ByteDance and Alibaba: Avoiding Direct Competition, Focusing on an "Ecosystem + Fleet" Approach
The AI strategies of these three giants differ:
- ByteDance: Builds from fundamental models towards chatbots and then agents (e.g., DouBao is powered by ByteDance's large model).
- Alibaba: Expands from cloud infrastructure to enterprise agents, leveraging Alibaba Cloud’s computing power and customer base.
- Tencent: Avoids direct competition in chatbots and fundamental models, focusing on a "computing power foundation + agent fleet" strategy, using WeChat’s 1.4 billion monthly active users and mini-programs to create a competitive advantage.
Although Tencent is a latecomer in the computing power investment sector (with annual budgets of less than 200 billion yuan compared to ByteDance’s 200 billion yuan and Alibaba’s over 380 billion yuan), it has seen the fastest growth in the second quarter, indicating a significant commitment to this strategy.
5. Challenges: Agent Performance Limitations and Model Dependency
Despite the enthusiasm for agent development, there are challenges:
- Limited Capability for Complex Tasks: Agents can handle simple tasks (e.g., document creation, spreadsheet management), but struggle with more complex processes involving multiple systems.
- Model Performance Impact: The performance of agents depends on the underlying models. If Hunyuan 4 and Hunyuan 5 fail to match the latest standards, their capabilities will be weaker compared to those of ByteDance and Alibaba.
- Ecosystem Dependency: Agents are closely integrated with WeChat; however, if users change their habits (e.g., stop using WeChat for work), this integration could become a disadvantage.
However, Tencent has no choice but to pursue its agent strategy, as it represents an opportunity to leverage its 20 years of ecosystem development. As Ma Huateng said: "The old boat was leaking; we’ve switched to a new one. Now that we’re on board, I hope the new boat can sail faster."
Conclusion: How Fast Can Tencent’s New AI Strategy Succeed?
Tencent’s approach is pragmatic: it does not compete with rivals on individual aspects (fundamental models, chatbots) but uses a combination of computing power, agent fleets, and an integrated ecosystem. Whether this strategy will be successful depends on two key factors: the sustainability of computing power supply and the ability of the Hunyuan model to handle more complex tasks. After all, even the largest fleet cannot win battles without effective weapons. Nevertheless, fighting on its own home ground gives Tencent a better chance of success compared to competing in other companies’ territories.