Tencent Holdings released its second-quarter results, posting revenues of RMB 204.79 billion, an 11% year-on-year increase. Non-IFRS operating profit reached RMB 75.64 billion, up 9%, and excluding the impact of new AI products, this figure grew 19% to RMB 86.1 billion. Following the earnings release, key executives including Chairman and CEO Pony Ma, President Martin Lau, Chief Strategy Officer James Mitchell, and CFO John Lo held a conference call to discuss the results and field analyst questions. The following is a summary of the key Q&A session from the call.
Addressing AI Investment and Returns
Analyst Robin Zhu from Bernstein inquired about the significant capital expenditure spike, noting it reached approximately RMB 53 billion in the quarter. He questioned how to view AI-related costs, whether new revenue from AI could cover these investments, and the potential impact on profitability in the coming quarters. James Mitchell responded that while leasing computing power to third parties could quickly recover depreciation costs and generate short-term gains, Tencent is pursuing a different, longer-term strategy. The company will allocate the vast majority of new computing power to developing its own large models and AI applications, aiming for industry leadership and a dominant domestic market position. The long-term goal is to translate top-tier AI capabilities into superior business returns, for example, through token sales via WorkBuddy. Martin Lau elaborated, explaining that Tencent's business can be viewed in two parts: the mature core business, which provides stable growth and operating leverage, and the new AI-native business, encompassing self-developed models, applications, and new computing infrastructure. Capital expenditure can also be divided accordingly: one part supports the traditional business and is self-sustaining, while the other is a concentrated, one-time investment specifically for the AI business, including training and inference computing power. This investment has clear upside, as demand for computing power rental is strong. In fact, computing power ordered months ago could now be resold at a significant profit. However, Tencent believes that prioritizing self-development of models and applications before renting out spare capacity is the optimal path to building a large-scale, profitable AI business.
The Strategic Role of WorkBuddy
Robin Zhu also asked about the competitive landscape for AI-powered office tools, specifically how WorkBuddy will compete with similar tools developed by other tech companies, and whether it is essentially enterprise software integrated with Tencent Meeting and Docs or a new platform. Martin Lau clarified that WorkBuddy is fundamentally a new platform—a highly flexible workbench designed for the general AI era, capable of executing tasks for office workers, freelancers, and small business owners. Its core value lies in an orchestration framework that helps users call upon multiple large models and smart agents to solve complex tasks. Long-term, WorkBuddy will integrate a wide range of models and plugins from various developers, aiming to solve efficiency problems by matching the best model or skill for each task, thereby controlling costs. The Hunyuan model will be one of the key models integrated, but not the only one available on the platform.
Xiaowei's Potential and Monetization
Analyst Kenneth Fong from UBS raised questions about Xiaowei, Tencent's new AI assistant. He asked about initial user feedback from testing, the challenges faced, and its monetization potential. He expressed concerns that an AI agent might merely shift transactions from manual operations in mini-programs to an AI interface, increasing computing costs without generating substantial new Gross Merchandise Value (GMV), while also potentially cannibalizing high-margin advertising inventory. Martin Lau dismissed these risks, stating that AI will make the WeChat ecosystem more intelligent, allowing users to complete transactions and handle tasks automatically. He drew an analogy to the evolution from PC-era QQ to mobile-first WeChat, which multiplied the ecosystem's value. The AI era represents another massive opportunity for WeChat to evolve into an AI-centric ecosystem. With Xiaowei, users will simply issue a command, and the AI will execute complex tasks, creating a revolutionary experience for users and merchants. Tencent's self-developed visual language model is designed with privacy, cost, and WeChat-specific AI requirements in mind, aiming to adapt the entire ecosystem for the AI era at a controlled cost. Initial testing of the Xiaowei prototype has reinforced confidence in this vision.
AI Cloud Pricing and Margins
Kenneth Fong also inquired about the gross margins of Tencent's AI Cloud business, given the rapid price decline of large model tokens in China and the price-sensitive nature of the domestic cloud market. James Mitchell explained that while token pricing in China is indeed low, the underlying production costs are also extremely low, often lower than market estimates. Even at low prices, token sales can generate positive gross margins. The gross margins for WorkBuddy's paying users and the Model-as-a-Service (MaaS) business are currently comparable to Tencent Cloud's overall margin. The overall blended margin for WorkBuddy is lower due to subsidies for free users aimed at capturing market share. Furthermore, the competitive pricing pressure in the domestic cloud market has eased in recent months due to rising hardware costs, particularly for memory. Tencent Cloud has raised prices and reduced discounts, making the market environment significantly less competitive than before.
Hunyuan 4's Differentiation and CapEx Allocation
Analyst Ronald Keung from Goldman Sachs had two questions. First, how will the upcoming Hunyuan 4 differentiate itself from the cost-effective Hunyuan 3, which excels at agentic tasks, in the increasingly crowded trillion-parameter model race? Martin Lau responded that Hunyuan 3 is a small-to-medium-sized model with broad applicability, offering performance comparable to much larger competitors while focusing on real-world business scenarios over benchmark scores. This philosophy will continue with Hunyuan 4, which will be larger in parameter size, surpass even larger competitors in performance, and provide a significant leap in utility, ushering in a new phase of stronger AI capabilities. All of Tencent's AI products are co-developed with its models, so Hunyuan 4 will bring substantial upgrades across the board. The long-term plan involves iterating towards the industry frontier, eventually creating a multi-tier model matrix with various parameter sizes and costs for different products and scenarios.
James Mitchell answered the second question regarding the allocation of AI capital expenditure. He confirmed that the top priority is training the next generation of Hunyuan models. The second most important use is providing inference power for Hunyuan, DeepSeek, and other models on WorkBuddy. The strategic focus on WorkBuddy is to drive the adoption of this key application while generating immediate cash flow. While there is a lag between cash collection and revenue recognition, cash revenue is growing rapidly. Later this year or early next, with sufficient GPU capacity, Tencent will also scale up its bare-metal GPU leasing and MaaS offerings. However, token billing through WorkBuddy is seen as the most long-term and stable source of commercial value, which is why it receives priority for computing power allocation.
The Future of WeChat Agents and Advertising Growth
Analyst Alicia Yap from Citi asked about the concept of agent-to-agent transactions within Xiaowei, whether it represents the long-term vision for a fully autonomous agent ecosystem in WeChat, and the role of on-device inference. Martin Lau explained that agent-to-agent transactions are the long-term vision. In the future, users will instruct their personal AI agents to complete complex tasks, and merchants will have their own agents. These agents will be able to interact and transact automatically. Regarding on-device inference, it will be a phased process. Initially, a hybrid cloud-device architecture will be used, with more tasks moving to devices as local hardware improves. This trend will ultimately increase the value of software and models, as the capital expenditure for computing power will be shared across the ecosystem. Tencent is already preparing its technology for this shift.
James Mitchell addressed the question on advertising growth, noting that its 22% growth rate is subject to fluctuations. While in-app advertising from mobile games contributed a couple of percentage points this quarter, the macro economy and consumer spending also create uncertainty. Despite this, Tencent's ad business continues to outpace the broader market, driven by three factors: ongoing implementation of AI-powered ad targeting, growth in core inventory like WeChat Video Accounts, and the early-stage, high-conversion closed-loop ad models that support rising ad unit prices.
Capital Allocation Priorities
Analyst Alex Liu from BofA Merrill Lynch asked about capital allocation priorities for the next 12-24 months, given the simultaneous increase in buybacks and capital expenditure. James Mitchell stated that capital allocation is dynamic. If investing in computing power to develop models and power WorkBuddy offers superior returns, Tencent will allocate more cash there and reduce buybacks. Martin Lau added a crucial point: the heavy AI-related capital expenditure is a one-time, concentrated investment primarily for this year and next. It should not be assumed to be a recurring annual expense. Model training is a fixed cost, and inference power will only be scaled up if the business generates sufficient returns. The funding for this one-time investment will come from operating cash flow, balance sheet cash, and investment assets, ensuring a prudent and manageable scale.
Hunyuan's Strategic Value and AI Spending Management
Analyst Alex Yao from J.P. Morgan questioned the strategic value of achieving a truly frontier-level, large-scale Hunyuan model, given that Hunyuan 3 prioritizes cost-efficiency. He asked what new commercial value a frontier model could unlock that Hunyuan 3 cannot. Martin Lau clarified that Xiaowei's strategy is distinct from Hunyuan's. Xiaowei uses a dedicated visual language model focused on privacy and cost. However, a top-tier Hunyuan model would create significant value by: 1) Building a large-scale token billing business; 2) Empowering WorkBuddy to handle more complex and higher-value tasks for clients; and 3) Enabling a tiered model matrix on a top-tier foundation, with specialized models for different cost points and tasks, all of which can be profitable.
On managing AI-related spending, Martin Lau explained that the strategy is flexible and prudent. Spending is managed within a fixed percentage of group profit, but Tencent will increase investment if a clear breakout opportunity is seen. The company views AI as a long-term track and is prepared for long-term investment. Critically, there is a safety net: even if they stop developing applications, simply renting out the computing power would be profitable, which gives them the confidence to invest. James Mitchell added that funding is dynamically reallocated within a total budget ceiling, as seen by the shift in spending towards the surging WorkBuddy from other AI product lines between Q1 and Q2.
AI Investment, Xiaowei's Role, and Profitability Timeline
Analyst Gary Yu from Morgan Stanley asked where Xiaowei fits in the computing power priority list and when overall profit growth, including AI investments, will outpace profit growth excluding these investments. Martin Lau stated that Xiaowei's annual operating budget is manageable, lower than the historical spending on Yuanbao. The cost is expected to be quickly covered by business returns as the product experience improves. Regarding a specific timeline for profitability, the company will not provide precise quantitative guidance. However, the overall investment philosophy remains disciplined, with a ceiling on spending. The safety net of being able to lease spare computing power ensures the investment risk is well-controlled.
Comments