On August 12, after Hong Kong market close, TENCENT released a second-quarter earnings report that appeared solid: total revenue of 204.785 billion yuan, up 11% year-on-year, and Non-IFRS operating profit of 75.64 billion yuan, up 9%. Both figures slightly exceeded market expectations, suggesting the stock should have risen. However, the next day, TENCENT shares closed nearly 4.5% lower. The market wasn't celebrating. Instead, it was fixated on another number: capital expenditures of 52.784 billion yuan, a staggering 176% increase year-on-year, nearly 2.8 times the amount from the same period last year. Free cash flow turned negative for the first time since listing, at -13.8 billion yuan. For its new AI products alone, the company is burning over 100 million yuan daily. What exactly is TENCENT betting on with this massive spending?
While all three tech giants are pouring money into AI, their strategies have diverged. Alibaba treats computing power as a business, generating revenue by renting out GPUs. ByteDance is going all-in on its own Doubao and AI applications. TENCENT's choice is the most distinctive: selling computing power ranks only third in its capital expenditure priorities. TENCENT isn't in the business of selling computing power; it's betting on controlling the entry point.
TENCENT clearly outlined where its money is going during the earnings call, with a very specific order of priority: first, training larger and better Hunyuan models; second, providing inference computing power for the Hunyuan model behind WorkBuddy, as well as DeepSeek and other models; third, and only then, its cloud business. Models first, WorkBuddy second, cloud third. This is not a path of "renting out GPUs for quick profits." TENCENT is prioritizing this investment for its own models and AI-native applications. Chairman Ma Huateng opened the call by stating that WorkBuddy, based on monthly interaction volume, has already ranked first among China's AI productivity services. More directly, Chief Strategy Officer James Mitchell confirmed, "After confirming that WorkBuddy is breaking out, we decisively shifted resources towards it while lowering the priority of other new AI products in our portfolio." The most obvious example of these "other new AI products" is Yuanbao. Once the favored "son" of TENCENT's consumer AI strategy just a year ago, it has now been clearly downgraded to an experimental project. TENCENT has shifted its bet from "building a chatbot" to "building an AI-native entry point."
The core business isn't broken; AI is simply consuming the incremental profits. Many people's first reaction is that "TENCENT's profits are being dragged down by AI." But a closer look at the books tells a different story. There's an overlooked number in the financial report: if you exclude new AI products like Hunyuan, Yuanbao, WorkBuddy, and CodeBuddy, TENCENT's Non-IFRS operating profit would be 86.1 billion yuan, a year-on-year increase of 19%, while the reported growth rate is only 9%. Subtracting 75.6 billion from 86.1 billion gives exactly 10.5 billion yuan – this is the amount TENCENT burned on new AI products in a single quarter. Note that this doesn't include the 52.784 billion yuan in capital expenditure; it's purely current spending on training, inference, and subsidizing free users. Last quarter it was 8.8 billion yuan, and this quarter it's another 1.7 billion yuan higher, with the bill growing larger. In other words, TENCENT's traditional businesses are still releasing profit elasticity, but AI is eating up that incremental gain. The core business is also performing well. Advertising revenue reached 43.565 billion yuan, up 21.8% year-on-year, accelerating for three consecutive quarters (17%, 20%, 22%). Time spent on WeChat Channels grew by 20%, and its ad load rate is still far lower than peers, meaning there's plenty of ammunition left. Domestic gaming revenue was 47.3 billion yuan, up 17% year-on-year, with daily active users for "Delta Force" and "Valorant" on PC hitting new highs. So, the market's real concern boils down to just one question: How long will it take for this 100-million-yuan-a-day bill to start generating visible returns?
What has the 100 million daily investment bought? Let's look at what's already been built – honestly, it's more than many people think. The official version of Hunyuan Hy3, launched in July, saw its daily token consumption increase sixfold compared to the preview version, and it has consistently ranked among the top three globally on OpenRouter. The larger Hy4 is expected by the end of the year, followed by Hy5. The term "SOTA" (state-of-the-art) was mentioned at least three times by management during the call – it's clear TENCENT is driven by ambition. WorkBuddy is already starting to generate revenue. Third-party data is even more compelling: according to Analysys, WorkBuddy's PC version had 20.97 million visits in June, ranking first among similar domestic products, surpassing the combined total of the second and third place. President Martin Lau positioned it as a "flexible workspace for agentic AI," targeting office workers and autonomous operators like "one-person companies." TENCENT has even embedded a "skill payment and revenue sharing" mechanism within the developer community's task flows – developers whose skills are called upon can share in the earnings. This is no longer a strategy of "making an AI assistant"; it's about building an ecosystem that can collect rent. The gross margin for paid users is already approaching the overall level of TENCENT Cloud, but because it's still subsidizing free users, the overall gross margin is relatively low. Management also explained a specific accounting detail: subscription payments from users for WorkBuddy have a significant time lag from when cash is received to when it's recognized as revenue on the financial statements. "Cash collection is climbing significantly," they said. Translation: WorkBuddy's money is already coming in; it's just not yet visible on the financial statements. The most telling sign of "burning money reflecting true demand" was an offhand detail from management: some computing power ordered and prepaid a few months ago, if resold at current prices, could yield a profit of over 30%. "But TENCENT would never do that." The fact that computing power prices are rising instead of falling indicates one thing: cheap models in China haven't reduced demand for computing power; the cheaper the models, the more people use them. Martin Lau also left a stronger card on the table: "If we directly adjusted our business model to renting out computing power, TENCENT would not only not lose money but could actually be profitable. This backup plan is always there, giving TENCENT considerable confidence." In short, this high-stakes gamble has a safety net – in the worst-case scenario, they can rent out the GPUs to recoup their investment.
The real battleground is not the present, but WeChat. This is the true focal point of this massive bet. TENCENT's biggest asset is WeChat, with 1.439 billion monthly active users. Management repeatedly emphasized one term during this earnings cycle: "AI-first." President Martin Lau drew an analogy: QQ was a social tool for the PC era; WeChat amplified it more than tenfold because WeChat was mobile-first. In the AI era, the WeChat ecosystem has a similar opportunity, first to be empowered by AI, and then to evolve into an AI-first ecosystem. On the product side, this manifests as "Wei Xiao" – an AI assistant powered by the customized WeLM model, currently in a small-scale grayscale test with its capabilities "held back" (requiring user intervention and multi-step confirmation). An analyst from UBS asked the most incisive question of the entire call: if agents shorten transaction paths, users enjoy the experience and stop watching ads, what happens then? Martin Lau's response was confident: "As long as we can deliver this experience and control the delivery cost, the WeChat ecosystem will expand and generate substantial value based on the existing monetization mechanism." First, let users have a great experience. As for how to make money, TENCENT has plenty of methods.
In the short term, valuations are under pressure; in the long term, the curve is worth watching. How to interpret this bet on the stock market? For TENCENT itself, Goldman Sachs, UBS, Jefferies, and JPMorgan all maintained positive ratings but lowered short-term earnings forecasts and target prices. Goldman Sachs cut its target from 700 to 670 Hong Kong dollars, UBS from 780 to 770, Jefferies to 750, and JPMorgan to 690. JPMorgan's assessment best captures the current sentiment: "Valuation is not high, but without a clear catalyst and auditable AI revenue, it's difficult for the stock price to be revalued quickly." For the computing power supply chain, TENCENT is the player that raised its full-year capital expenditure forecast from 170 billion to 250 billion yuan. This number itself is a bellwether for AI computing power demand – if it increases its investment, demand for servers, optical modules, and GPUs in the chain has a solid foundation. For Hong Kong tech stocks, this earnings report also served as a stress test: if even a cash machine like TENCENT has to delay profit realization for AI, the valuations of other smaller companies still burning cash become even harder to justify. In the AI application space, valuations will likely be under pressure together in the short term, until someone first demonstrates auditable revenue. Four things are worth watching next: whether the Hunyuan Hy4 model at the end of the year can truly deliver; WorkBuddy's user retention and payment rates; when WeChat's "Wei Xiao" will move from grayscale testing to scale; and whether the daily token volume of TENCENT Cloud's TokenHub (already 25 trillion, up fivefold in two months) can continue to grow. The triggers are clear: the performance of Hunyuan Hy4 at year-end, WorkBuddy's retention and payment data, and when WeChat's "Wei Xiao" will expand its grayscale test. The failure conditions are also clear: if capital expenditures continue to rise above expectations while the AI revenue curve remains elusive, market patience will run out first. Management also offered a reassuring note: this investment is a "lump sum startup investment for this year and next," not an annual increase of the same amount. Subsequent additions will be tied to returns. In other words, they won't burn money mindlessly. Ultimately, TENCENT's current confidence stems from its ability to afford this burn. Whether the money is sufficient has never been the issue. The real question is whether this 100 million yuan a day can ultimately forge a visible revenue curve that everyone can see. If it succeeds, it's a ticket for the next decade; if it fails, the market may not be so patient.
