Tencent Holdings Ltd (TENCENT) shares dropped 4.46% after the company released its second-quarter results, signaling market disapproval despite solid financial performance. The company reported total revenue of RMB 204.8 billion, an 11% year-on-year increase, and gross profit of RMB 118.4 billion, up 13%. However, Tencent faces a critical strategic disconnect: it is playing catch-up in foundational AI models while leading in AI products like WorkBuddy, yet its heavy AI investments have driven free cash flow negative for the first time, to the tune of RMB -13.8 billion. For a traditionally cautious company, this aggressive spending is unprecedented and almost entirely attributable to AI. According to Tencent's explanation, "excluding prepayments for computing power procurement, our free cash flow would be RMB 37.6 billion." This implies that prepayments for computing power amounted to a staggering RMB 51.4 billion (13.8 + 37.6 billion). AI investments require enormous outlays and long cycles, but revenue and profit returns lag behind. This creates a temporary divergence: the more Tencent invests in AI, the faster its technological progress may be, but the more its short-term financial performance suffers. During this transitional period, the market struggles to reach a consensus, and the stock price decline reflects uncertainty and division. When will the market accept Tencent's new AI narrative?
Where to begin
The first step is catching up on models. The previously lagging Hunyuan model is now being led by Yao Shunyu, with organizational integration helping to regain momentum. However, whether "gradually approaching SOTA" is a roadmap or a promise remains to be seen with the launch of Hy4 and Hy5. The second is product monetization. WorkBuddy has established a clear lead, but its free period will not last forever. Yuanbao and "Xiaowei" have deep user access points, but their business models remain unclear. Regarding the timeline for profitability in Tencent's AI segment, President Martin Lau cautiously stated, "At this stage, we will not provide precise quantitative targets." The third is a change in revenue structure. The value-added services from AI to existing businesses and the monetization of native AI businesses must account for a significant portion of total revenue. This earnings report feels more like a declaration of war on Tencent's old self: buybacks can shrink, cash flow can turn negative, but computing power must be secured, and models must be trained. Pony Ma once compared AI to "a leaking ship." Tencent must board that ship, or it will miss the AI new continent. Now, the company has chosen to spend RMB 51.4 billion to reinforce the hull and install a new engine. Spend big money, earn slow money. The essence of this climb is trading today's cash flow for tomorrow's ticket. More answers about Tencent's AI lie in the following transcript from the company's earnings call, which has not been reviewed by the executives. Attendees included Pony Ma (Chairman and CEO), Martin Lau (President), James Mitchell (Chief Strategy Officer), and John Lo (CFO).
Question 1: AI Investment Returns and Capital Expenditure (Robin Zhu, Bernstein)
Robin Zhu asked about capital expenditure reaching approximately RMB 53 billion this quarter, a significant increase from the previous quarter, with an annualized scale exceeding RMB 200 billion. He questioned how to view the cost of AI investments, how much of the new revenue from AI would cover these costs, and whether earnings would be continuously eroded by this investment in the coming quarters. He also asked about the payback period for AI investments, including R&D costs. James Mitchell responded that if they leased computing power to third parties, they could almost immediately recover equipment depreciation costs, as many new cloud vendors do. However, they have a different, long-term strategy. They will allocate the vast majority of new computing power to their self-developed large models, aiming to achieve industry-leading standards, while also deploying and popularizing their self-developed AI applications to secure a leading position in the domestic market. They believe that top-tier intelligent capabilities, built on leading models and AI applications, can be converted into exceptional commercial returns over the long term. For example, they can monetize by selling tokens through WorkBuddy.
Martin Lau added that the company's business can be split into two major parts. The first is the existing mature core business, which has stable growth and good operating leverage. The other is the new AI-native business being built, which includes self-developed large models, new AI applications, and supporting computing infrastructure. When looking at financial data, investors can separate the revenue and profit of the mature business, and Tencent will also disclose the investment in the AI-native business as a separate operating item. Capital expenditure can also be divided into two parts. The first part serves the traditional mature business, which can generate its own operating free cash flow. The second part is specifically for the AI-native business, representing a one-time concentrated investment to procure training and inference computing power, as well as hardware. The core reason for the large-scale procurement of computing power is the need to start the entire AI business line. This investment has a clear upside: their large models and new AI applications are performing well, and market demand for computing power leasing is strong. If all computing power were placed in Tencent Cloud for external leasing, it could generate significant revenue and a good return on capital expenditure. In fact, a batch of computing equipment ordered a few months ago could now be resold for a profit of over 30%. However, they believe the long-term strategy is to use the computing power for self-developed models and applications first, then lease the remainder. This will allow them to build a large, profitable AI-native business with strong cash flow and returns.
Question 2: WorkBuddy's Positioning and Competitive Landscape (Robin Zhu, follow-up)
Robin Zhu asked about the market landscape for AI-powered office scheduling tools, as various tech companies are developing their own, and how Tencent plans to compete with them. He also questioned whether WorkBuddy is an enterprise software complementing Tencent Meeting and Tencent Docs, or a new platform for the entire industry's AI capabilities. Martin Lau responded that WorkBuddy is essentially a new platform, a highly flexible workbench tailored for general AI. Its core value is executing tasks, addressing all productivity needs for office workers, individual operators, and small businesses. The platform has a scheduling framework that helps users call on multiple large models to solve complex tasks. Over the long term, many models will serve users on WorkBuddy, and developers will create a vast number of skill plugins. The platform's core goal is to solve efficiency pain points by integrating all available tools and models. At this stage, they act as a scheduler, matching users with the most suitable model and optimal skills to ensure task completion while controlling costs. Hunyuan will be one of the models on WorkBuddy. As long as Hunyuan can efficiently solve a large number of user needs, it will become the core underlying model, but it will not be the only model available.
Question 3: WeChat AI "Xiaowei" Progress and Monetization (Kenneth Fong, UBS)
Kenneth Fong asked about the development progress of "Xiaowei," the initial user feedback from prototype testing, and the challenges encountered. He also asked about its net monetization potential, expressing concern that AI agents simplifying transaction paths might merely shift existing manual transactions in Mini Programs to agents, increasing computing costs without generating new Gross Merchandise Value (GMV). He also questioned whether shorter transaction paths could cannibalize high-margin ad impressions. Martin Lau dismissed these risks. He believes AI will make the WeChat ecosystem smarter, allowing users to automatically complete transactions, browse content, and handle daily tasks. This will make the already rich ecosystem even more valuable. He drew an analogy: in the PC era, QQ was just a social communication tool. In the mobile internet era, WeChat was born, magnifying QQ's overall value by over ten times because it was a mobile-first product. Now, with the AI era, WeChat's ecosystem faces a second huge growth opportunity, evolving into an AI-centric application and ecosystem. Instead of manually typing and clicking, users will simply give a command to Xiaowei, which will automatically complete the transaction. This will provide a disruptive user experience and empower all merchants in the WeChat ecosystem. Their self-developed vision-language model was designed with user privacy, operating costs, and adaptation to the WeChat AI environment in mind, aiming to meet all existing user needs. If they can deliver this experience at a controlled cost, they can fully adapt WeChat to the AI era. The ecosystem will continue to expand, and with WeChat's existing advertising, payments, and e-commerce monetization systems, it will naturally create massive new value. After the prototype testing of Xiaowei, they are increasingly confident in this vision.
Question 4: AI Cloud Gross Margin and Price Competition (Kenneth Fong, second question)
Kenneth Fong asked about the gross margin of Tencent's AI cloud business compared to traditional infrastructure cloud and platform cloud services, given the ongoing price declines for Token pricing in China and the industry's rapid commoditization. He also inquired about the future trend of gross margins as AI adoption increases. James Mitchell acknowledged that Token pricing is relatively low in the Chinese market, but the underlying production cost is also extremely low, far below common perceptions and external estimates. Even with low prices, the Token business can maintain positive margins because costs are controllable. He pointed to the gross margin of WorkBuddy paid users and the Model-as-a-Service (MaaS) business, which is already on par with Tencent Cloud's overall gross margin. The overall comprehensive gross margin for WorkBuddy is lower due to subsidies for free users to gain market share, but the margin for the paid customer segment is very good. While the domestic cloud market is indeed price-competitive, the environment has changed significantly in recent months. Hardware raw material costs, especially for memory, have been rising. Consequently, they have raised prices for customers across the entire Tencent Cloud product line since May and have also significantly reduced customer discounts. Overall, the price competition pressure in the domestic cloud market is much less than in the past.
Question 5: Hunyuan 4 Differentiation and Capital Expenditure Allocation for Cloud (Ronald Keung, Goldman Sachs)
Ronald Keung asked two questions. First, given that Hunyuan 3 focuses on high cost-performance and strong agent capabilities, what is the differentiated positioning of the upcoming Hunyuan 4? As the trillion-parameter model space becomes increasingly crowded, how will Hunyuan 4 build its competitive moat? Second, noting that US tech giants are shifting their strategic focus from applications to building up their cloud computing businesses, how does Tencent plan to allocate its AI-related capital expenditure? At what point will it prioritize cloud as a high-return area? What are the similarities and differences between Tencent Cloud's strategy and that of overseas tech companies? Martin Lau responded that Hunyuan 3, which is a small-to-medium parameter model by current industry standards, has a very wide range of applications. It has two core characteristics: first, its performance can match or even exceed that of much larger competitors; second, its R&D focus is on real-world business scenarios, not just benchmark scores. This makes it far more practical than competing models of similar or even larger size. This R&D philosophy will be fully carried over to Hunyuan 4. Hunyuan 4 will have significantly more parameters, performance surpassing larger competitors, and a qualitative leap in practicality. This will help them enter a new phase, providing users with stronger intelligence. All their AI products are co-developed with the large models. When Hunyuan 4 launches, its entire product suite will see major upgrades in functionality and practicality, bringing significant growth to related businesses. This is their iterative roadmap, with Hunyuan 4 being a milestone, followed by Hunyuan 5. Through continuous iteration, they will approach and eventually reach industry-leading standards. Once they achieve that, they will build a multi-tiered gradient model matrix to serve different costs and user needs. These differentiated models will be adapted for various product lines like code and office, enriching functionality while balancing performance, capability, and execution speed.
James Mitchell addressed the second question by stating that in the coming months, the top priority for capital expenditure will be training larger, more powerful Hunyuan models. The second most important use will be providing inference computing power to support Hunyuan, DeepSeek, and other third-party models integrated into WorkBuddy. Their core strategic goal with heavy investment in WorkBuddy is to drive the adoption of this strategically important application while continuously collecting scenario feedback to optimize the large model and the entire Tencent ecosystem. This product also generates immediate cash revenue. From an accounting perspective, most user spending on WorkBuddy is subscription fees. Similar to games, there is a time lag between cash collection and revenue recognition, but cash revenue is growing rapidly and will gradually convert into Tencent Cloud's reported revenue this year. By the end of this year or early next year, their GPU computing power reserves will be sufficient to scale up two external businesses: bare-metal GPU leasing and MaaS sales. However, among all monetization channels, the Token-based billing business from WorkBuddy will create the most long-term and stable commercial value. This is the core reason for prioritizing computing power allocation to it at this stage.
Question 6: Agent Transaction Closure, On-Device Inference, and Advertising Growth Momentum (Alicia Yap, Citigroup)
Alicia Yap asked about the concept of agent-to-agent transaction closure, whether this is the long-term direction for WeChat's autonomous agent ecosystem, and the exploration of on-device inference for Xiaowei. She also questioned the benefits and challenges of on-device inference and why Tencent chose a self-developed vision-language model for Xiaowei instead of using a third-party model. She also had a brief follow-up on advertising revenue growth, which had accelerated to 22% in the quarter, asking if AI advertising tools and automated systems could sustain this momentum and what long-term incremental benefits Hunyuan 3 could bring to the advertising business. Martin Lau explained that the agent-to-agent transaction closure is their long-term vision. In the future, users will give commands to Xiaowei and various AI agents to complete transactions. Instead of manually browsing and operating Mini Programs, users will simply give a complex command to their personal agent, which will automatically complete the entire transaction. Most Mini Program merchants will also deploy their own merchant agents. In the long run, merchant agents can connect directly with user agents. Ultimately, every user will have a dedicated AI agent, and agents can interact and complete transactions automatically. This is a future scenario they are building towards in stages. Regarding on-device inference, it will be rolled out in phases. In the short term, a hybrid cloud-device architecture will be used. Over time, as local hardware capabilities improve and model efficiency increases, more inference tasks will move to the device. This is the norm in the computing and smartphone industries. However, in the early stages of AI infrastructure, models require immense computing power, and current local hardware is insufficient for large-scale AI inference. In the future, as phones and PCs get more powerful GPUs, more inference will move on-device. At that stage, the value of the software and the model itself will increase significantly, and the return on investment for models and applications will improve because the capital expenditure for computing power will be shared across the entire ecosystem, not borne solely by the model developer. This trend is inevitable, and they have already started preparing the relevant technology.
James Mitchell addressed the advertising question, noting that growth rates have fluctuated in the past and will continue to do so, as they cannot be simply extrapolated linearly. He mentioned that in-app advertising in mobile games, while a drag on overseas game revenue, contributed two percentage points to the ad segment's growth this quarter. This is a new category for the industry, and its future impact is uncertain. External factors like macroeconomic pressures and consumer spending also affect ad spending. Despite this, their ad business growth has consistently outpaced the domestic market and will continue to do so. This is supported by three factors: first, the ongoing rollout of AI-powered precision advertising; second, the steady increase in user time and ad inventory on core traffic sources, especially video accounts; and third, the high-conversion closed-loop ad model, which is still in its early stages and will drive ad unit prices higher over time.
Question 7: Buybacks and Capital Allocation Priorities (Alex Liu, BofA Securities)
Alex Liu asked about how investors should view Tencent's capital allocation priorities over the next 12 to 24 months, given the increased share buybacks since May and the significant acceleration in capital expenditure, which is still in its early stages. James Mitchell stated that the capital allocation strategy is dynamic and flexible, changing with the market environment. If they believe that increasing computing power capital expenditure to develop models, lease computing power, or sell MaaS will generate superior returns, they will allocate more cash to CapEx and correspondingly reduce the scale of share buybacks, maintaining a dynamic balance. Martin Lau added that the capital expenditure for the AI-native business is primarily a one-time concentrated investment this year and next. Investors should not assume they will maintain this level of spending annually. Model training is a fixed cost; once sufficient computing power is secured, it doesn't require continuous annual increases. Inference computing power requires sufficient reserves to support Token monetization and leasing, but they will only continue to invest if the business generates satisfactory returns. If returns are not as expected, the existing investment scale will be capped. Future incremental CapEx will be directly linked to the revenue generated by the business. The funding for this one-time computing power investment cannot be viewed solely through operating cash flow; it must also consider cash reserves on the balance sheet, investment assets, and new operating cash flow. These factors together will determine the funding arrangement for the initial computing power investment.
Question 8: The Commercial Value of Hunyuan Reaching Industry Leadership (Alex Yao, JPMorgan)
Alex Yao asked about the top-level strategy for Hunyuan. Given that Hunyuan 3 focuses on cost-effectiveness rather than extreme hardware performance, what additional commercial value would a truly industry-leading, larger, and more expensive model create that Hunyuan 3 cannot? He asked if this would be a stronger WeChat agent, better ad performance, or increased revenue from enterprise customers, and whether these potential benefits could justify a significant increase in model training investment over the next 12 months. Martin Lau clarified that the R&D positioning and strategic route for Xiaowei's vision-language model are completely separate from Hunyuan. Xiaowei's underlying model is designed around user privacy, adapting to all intelligent interaction needs within the WeChat ecosystem, and controlling operating costs. However, Hunyuan reaching industry-leading standards can bring multiple commercial values. First, it would enable a large-scale Token-based paid business. Second, it would empower WorkBuddy to handle more complex and higher-value-added office services, creating more value for clients. Their vision for WorkBuddy is not just a tool that replicates manual office operations. They are continuously exploring high-value-added scenarios to create incremental value and even revenue for clients. Achieving this would unlock numerous new business models. Once Hunyuan reaches an industry-leading level, they can also create a tiered model matrix, developing specialized models for different cost brackets and tasks on top of the advanced foundational technology. Since they control the models and inference costs, each layer of products can be profitable. This is their long-term goal for the next generation of large models.
Question 9: AI Investment Control: From RMB 8.8 Billion to RMB 10.5 Billion, When to Scale Up or Turn Profitable (Alex Yao, second question)
Alex Yao asked about how management controls AI-related product investment, which rose from RMB 8.8 billion in Q1 to RMB 10.5 billion in Q2. He asked if there is a fixed spending cap, a revenue target threshold, or a strategic disregard for short-term returns. He also asked what specific user data, revenue growth, or profitability signals would prompt the company to scale up investment further or transition a product from pure investment to profitability. Martin Lau stated that the AI investment strategy is flexible and dynamic. They will maintain prudent investment, but if they see clear explosive growth opportunities, they will increase spending. The overall investment will be controlled within a fixed proportion of the group's profit, but if they can foresee substantial returns, they will increase investment. They believe AI is a long-term race and will continue to invest, with commercial returns gradually materializing. Crucially, they have a fallback plan: even if they stop developing their own applications, simply leasing the computing power to third parties would be profitable. This gives them the confidence to invest for the long term. James Mitchell added that they dynamically adjust funding priorities within a total budget cap. The allocation between Q1's RMB 8.8 billion and Q2's RMB 10.5 billion showed a significant change: they found WorkBuddy experiencing explosive growth and shifted funds heavily towards it while cutting budgets for other AI products. This dynamic adjustment mechanism will continue to be used.
Question 10: Xiaowei's Computing Power Tier and AI Business Profitability Timeline (Gary Yu, Morgan Stanley)
Gary Yu asked two questions about AI investment. First, he understood the priority order for investment is model training, WorkBuddy, and then cloud computing leasing. Where does Xiaowei fit in this computing power allocation hierarchy? Second, he asked about the commercialization cycle and profitability visibility for various AI innovation businesses, and when the overall operating profit growth, including AI investment, will exceed the profit growth excluding AI investment. Martin Lau replied that the annual budget for Xiaowei's ongoing operations is lower than the normalized investment scale for Yuanbao in previous years, and the overall cost is entirely controllable. As the product experience improves, business returns will be realized quickly and will soon cover the cost. Regarding the timeline for profitability, he stated they will not provide precise quantitative targets at this stage. They have described their overall judgment on the AI business and related investments. The investment scale will have an upper limit, maintaining Tencent's consistent prudent and disciplined management style. However, if they believe they can build a large-scale, long-term profitable AI business, they will also invest appropriately. They always have a fallback plan: they can lease idle computing power on Tencent Cloud to quickly generate revenue, profit, and investment returns, making the investment risk completely controllable.
Comments