Summary of Key Points
Tencent’s financial report for Q2 2026 shows impressive results (revenue growth of 11%, adjusted profit growth of 9%), but behind these figures lies a contradiction: substantial capital expenditure on computing power, which increased by 176% to 52.78 billion yuan, resulting in negative free cash flow until the exclusion of prepaid AI computing power payments. The management has clarified the priority for using this computing power: first for training self-developed models → then for inference within their own AI applications → and finally, any excess capacity will be rented out. They even forgone the short-term opportunity to earn a 30% profit by reselling computing power orders, betting on the high profits that the long-term AI business can generate. The next 12-18 months are crucial for verification; success or failure will depend on whether the models prove effective and whether the internal applications can make efficient use of this computing power. Otherwise, the billions invested in computing power could become a long-term source of depreciation pressure.
Why Doesn’t Tencent Earn That 30% Quick Profit?
During the conference call, Liu Chiping mentioned, “We could earn a profit of over 30% on the prepaid computing power orders before reselling them.” However, Tencent chose not to do so. This isn’t due to foolishness but a clear distinction between short-term arbitrage and long-term business strategies:
- Short-term Arbitrage: It’s like buying and selling concert tickets at a profit due to temporary scarcity or differences in delivery times; once the surge passes, the profit disappears.
- Long-term Business: Using computing power to develop AI models (such as Hunyuan) and applications (like WorkBuddy) creates a continuous source of revenue, akin to a “money-making machine.”
The management’s logic is that if they rent out all the computing power now to earn quick profits, they will fail to build their own competitive advantage in AI. By forgoing short-term arbitrage, they aim for greater long-term benefits.
Is Renting Out Computing Power a Viable Option?
Tencent claims they can rent out excess computing power, but this option is not guaranteed. Three factors need to be considered:
1. Internal Demand: If Hunyuan training, WorkBuddy inference, and WeChat AI are constantly using up all the computing power, there won’t be any surplus available for rental. The fact that the WeChat AI team can’t even obtain domestic computing power indicates a current shortage, so having excess capacity is a long-term goal.
2. Depreciation: GPUs have a useful life of 2.5-3 years; if the models don’t prove effective, the hardware will start depreciating quickly, and rental income might not cover the depreciation costs.
3. Market Conditions: Global GPU production will increase in 2026-2027, likely leading to lower rental prices. By then, rental earnings could be significantly reduced.
Tencent also has a “regulatory mechanism” in place: The investment in computing power for inference is linked to business returns. If the applications don’t generate sufficient revenue, they will reduce their purchasing of computing power, thus limiting the potential benefits from this option.
WorkBuddy: The First Test Case for Tencent’s AI Business
WorkBuddy is Tencent’s most prominent AI commercialization effort, with promising numbers (20.97 million monthly PC visits and 20 million monthly active users across platforms, with similar gross margins to Tencent Cloud). However, there are concerns:
- Payment Conversion: The enterprise version costs 198 yuan per user per month, but most users are still in the free trial phase; the conversion rate from free to paid and the customer acquisition cost are not disclosed.
- Integration with Hunyuan: WorkBuddy can use third-party models (similar to Lark and DingTalk), so using it doesn’t necessarily mean using Hunyuan. Only if Hunyuan outperforms third-party models in practical tasks can WorkBuddy truly contribute to Tencent’s revenue.
On the positive side, the daily average Token usage of the official Hy3 version is seven times that of the preview version, indicating an improvement in model efficiency.
WeChat AI: The Biggest Consumer of Computing Power – or a Money Maker?
WeChat AI is Tencent’s most valuable differentiating asset. Once fully launched, it will require a large amount of computing power. Management states that the investment is lower than last year’s for Yuanbao, and costs are controllable. However:
- High Demand: The demand for computing power will far exceed the current beta phase.
- Launch Timing Uncertain: WeChat’s technical team needs more iterations on the models (including their own WeLM and various business-specific models), so the timing of a public trial or full rollout is unclear.
- Resource Allocation: WeLM is designed for the WeChat ecosystem, focusing on privacy and cost-effectiveness, and it doesn’t compete directly with Hunyuan for computing resources.
If WeChat AI is successful, it will not only utilize a large amount of computing power but also generate significant commercial value. If the launch is delayed, the computing power might remain underutilized.
The Next 12-18 Months: A Critical Period for Tencent’s AI Strategy
The management emphasizes that the bulk of investment in model training will occur this year and next. These 12-18 months will determine the direction of Tencent’s AI strategy:
- Scenario One: If the models succeed, Hunyuan closes the gap with overseas competitors, WorkBuddy’s conversion rate exceeds 15%, and WeChat AI is fully launched. All computing power will be used internally, with rental as a secondary option.
- Scenario Two: If the models are average, Hunyuan performs well in specific scenarios but not at the top level. Excess capacity will be rented out to create a mix of internal use and rental.
- Scenario Three: If the models fail, excess computing power will need to be rented out, facing lower rental income and depreciation pressures. Annual depreciation could amount to 10-13 billion yuan, with rental revenue possibly only covering 60-80 billion yuan, leaving a deficit that must be made up through profits.
There’s also an unknown risk: The TEG (Technology Engineering Group) and WeChat teams are separate, and whether the priority for allocating computing power is effectively implemented is uncertain. If there’s internal competition for resources, the strategy to prioritize model development could be compromised.
In summary, Tencent’s investment in AI is not about having the most computing power but about transforming it into tangible product capabilities. The success or failure will depend on whether the models prove effective. These 12-18 months will reveal the outcome.