Summary of Key Points
Tencent's capital expenditure in the second quarter soared to 52.8 billion yuan (a 65% increase quarter-on-quarter and a 176% increase year-on-year), accounting for 26% of its total revenue, far exceeding the market's forecast of 32.1 billion yuan. This led to its free cash flow turning negative for the first time (-13.8 billion yuan). The majority of this investment was focused on AI, covering the entire ecosystem from models (Hy3/Hy4) and applications (WorkBuddy office AI) to infrastructure (computing power and memory). Tencent recognizes that the AI industry chain has created a positive feedback loop: better models attract more users, which in turn increases demand for computing power, and this increased demand generates revenue. However, the company also faces challenges such as insufficient computing power, its models lagging behind industry standards, and intense competition from rivals like ByteDance and Alibaba, so it is accelerating its efforts to catch up.
1. Tencent's Sudden Heavy Investment in AI: Where Did the 52.8 Billion Yuan Go?
Tencent's capital expenditure in the second quarter far exceeded expectations, amounting to 52.8 billion yuan—more than 20 billion yuan more than what the market anticipated. This means that for every 100 yuan it earns, it invests 26 yuan in AI-related hardware (such as servers and chips) and software (model training). Bernstein analysts estimated that at this rate, Tencent would spend 200 billion yuan per year, which is several times higher than Alibaba's capital expenditure of 26.9 billion yuan in the previous quarter.
The money was primarily allocated to three areas:
1. Model Training: The usage of tokens for the Hy3 model (the number of texts/instructions processed by the model) has increased by six times compared to the preview version, and the upcoming Hy4 model will require a significant amount of computing power.
2. Application Deployment: WorkBuddy, the office AI product, has become very popular, so resources have been shifted towards it at the expense of other AI projects.
3. Infrastructure: Tencent signed a 20-billion-yuan DRAM supply agreement with memory manufacturer ChangXin to secure server memory and purchased additional computing power. Some of these computing power orders can even generate a 30% profit when sold, but Tencent is reluctant to sell them due to its own high demand.
Why is Tencent so willing to invest? Because it sees AI as an investment that can generate returns: better models attract more users, increasing the demand for computing power, which in turn improves efficiency and creates a profitable cycle of "investment → growth → further investment."
2. WorkBuddy Becomes a Dark Horse: Has Tencent Changed Its Focus to Office AI?
Tencent, known for its strength in consumer internet services (such as WeChat and gaming), has unexpectedly taken the lead in the office AI sector. WorkBuddy is the most popular AI productivity tool in China, with monthly traffic exceeding the combined traffic of its two main competitors.
What makes it successful?
- Strong User Base: Despite being still in a free trial phase, it has a high user retention rate, and users are willing to pay for the service.
- Resource Allocation: WorkBuddy has received significant attention during phone meetings (compared to other AI products, which were mentioned only once), indicating that resources have been prioritized for its development.
- Feedback Loop: A large number of users use WorkBuddy for real work tasks (such as writing reports and organizing data), providing valuable feedback that helps improve the model.
However, challenges arise as ByteDance and Alibaba have identified WorkBuddy as a major competitor and are adjusting their organizational structures to compete in this market. Tencent has not yet clarified how it plans to compete with these rivals, especially regarding their own office AI solutions.
3. Computing Power Becomes a Valuable Asset: Can't Use It All, but Still Makes Money?
Computing power has become a critical asset in the AI era, not just a necessity but also a profitable resource:
- Elon Musk's xAI has sold unused computing power at a premium, and Meta is receiving numerous high-priced requests for computing power.
- Tencent can earn a 30% profit on some of its computing power orders, but it cannot sell them because it needs the power for its own model training and WorkBuddy operations.
The situation in China is particularly challenging: while overseas companies can purchase computing power at market prices (for example, Anthropic), domestic companies may struggle to secure sufficient capacity even with additional funding. Therefore, Tencent must prioritize its needs, with model training taking precedence over other cloud services.
4. Models Remain a Weakness: Is the Hy4 Model Not Up to Par?
Tencent's AI models are still behind industry leaders:
- The current Hy3 model has fewer than 300 billion parameters, while leading domestic models have reached 3 trillion and are moving towards 10 trillion.
- The upcoming Hy4 model is not considered the "state-of-the-art" (SOTA), and Tencent's iteration speed cannot keep up with competitors.
- Yao Shunyu, responsible for the models, has requested that industry rankings be ignored for at least a year after the new model is released to avoid embarrassing comparisons.
This is one of the reasons why Tencent is investing heavily in computing power: it needs more power to train larger and better models to catch up with the industry.
5. The "Micro-AI" within WeChat: Could It Become the Next Big Thing?
Tencent is focusing on "micro-AI" solutions that integrate into its WeChat ecosystem, allowing the AI to remember conversation context. This technology is currently in beta testing and aims to evolve into something similar to how WeChat transformed from QQ.
Tencent's expectations for this project are high:
- It is positioned as the next major advancement in mobile internet services, similar to how WeChat replaced QQ.
- The ultimate goal is to enable interactions between merchant and user AI entities, as well as among different AI systems.
However, progress has been slow due to the large number of WeChat users (1.4 billion) and the need for sufficient computing power to handle their requests, while also considering privacy and user experience.
In summary, Tencent is accelerating its investment in AI, but its rivals have already taken the lead in acquiring essential resources (computing power and advanced models). Now, Tencent must not only spend more money but also face challenges related to the scarcity of computing power and the difficulty of catching up with industry standards. This AI race has just entered a phase of accelerated competition.