Full text | Tencent Q2 earnings call transcript: AI is a long-term track, not preparing to lease computing power

Tencent Holdings released its second-quarter report, showing revenue of 204.79 billion yuan, a year-on-year increase of 11%; Non-IFRS operating profit was 75.64 billion yuan, a year-on-year increase of 9%; excluding the impact of new AI products, Non-IFRS operating profit increased by 19% year-on-year to 86.1 billion yuan.

After the financial report was released, Tencent Chairman and CEO Ma Huateng, President Liu Chiping, Chief Strategy Officer James Mitchell, and CFO Luo Shuo Han held a conference call to interpret the key points of the financial report and answer analysts' questions.

The following are the main contents of the Q&A session with analysts during this conference call:

Bernstein Research Analyst Robin Zhu: Thank you for the opportunity to ask questions. I would like to ask, this quarter's capital expenditure reached about 53 billion yuan, a significant increase compared to the previous quarter, with an annualized scale exceeding 200 billion yuan. How should we view the cost input related to AI? To what extent can the new revenue generated from AI investments cover these inputs? Will profits continue to be eroded by this investment in the coming quarters? Additionally, if we include R&D expenses, what does the payback period for this AI investment look like?

Mitchell: It's important to know that if we lease computing power to third parties, we can almost immediately recover the equipment depreciation costs. Many emerging cloud vendors operate this way, and we can also achieve good returns in the short term. However, we have a different layout, executing a longer-term strategy. We will use the vast majority of the new computing power for self-developed large models, striving to achieve industry-leading standards while implementing and popularizing our self-developed AI applications to secure a leading position in the domestic market. We believe that the top intelligent capabilities created based on top models and industry-leading AI applications can be transformed into excess commercial returns in the long run. For example, we can monetize through WorkBuddy by selling tokens, which is the development path we have chosen.

Liu Chiping: Let me elaborate further. At this stage, you can divide Tencent's business into two main parts. The first part is our existing mature core business, which has stable growth and good operating leverage. This is our high-quality growth foundation that we have honed over the long term, and we will continue to deepen our efforts in this area. The other part is the new AI-native business we are building, which, as James just mentioned, includes self-developed large models, new AI applications, and the newly built computing infrastructure. When looking at the financial report data, you can separately analyze the revenue and profit of the mature core business, and we will also disclose the investment in the AI-native business as a separate operating item.

Looking at capital expenditure, it can also be divided into two parts. The first part serves the traditional mature business, following the same logic as in previous years. This business itself can generate operating free cash flow, and the corresponding capital expenditure can be self-circulating, continuously creating cash flow. The other part of the capital expenditure is specifically directed towards the AI-native business, which is a one-time concentrated investment used to procure model training computing power, reserve the computing power needed for inference, and stock up on computing hardware to build AI computing and AI cloud services This is the core division logic of capital expenditure. The main reason for our significant investment in computing power is the need for computing power to launch the entire AI business line. At the same time, this investment has clear upward revenue potential: our large models and new AI applications have performed well, and there is currently strong market demand for computing power leasing. If we lease all the computing power through Tencent Cloud, it can generate substantial revenue, and capital expenditure can achieve considerable returns. In fact, a batch of computing equipment we ordered a few months ago with a deposit can now be resold for over 30% profit compared to the original purchase price. However, we believe that the long-term strategy should be to first use the computing power for self-developed models and refine our own applications, and then lease out the remaining computing power. This sequence will help build a large-scale, profitable AI-native business with excellent cash flow and return rates, which is our overall judgment on the business at present.

Bernstein Research Analyst Robin Zhu: The second question: major tech companies in the market are developing their own AI office scheduling tools. How does the management view the market landscape between proprietary tools and third-party tools? How do you plan to compete with WorkBuddy and similar products developed by other companies? In the management's view, is WorkBuddy an enterprise software that complements Tencent Meeting and Tencent Docs, or will it become a new platform that carries AI capabilities for the entire industry in the future?

Liu Chunning: Exactly, I believe WorkBuddy is essentially a new platform, adaptable to general artificial intelligence and a highly flexible workspace. Its core value lies in executing tasks, addressing the productivity needs of office workers, various individual entrepreneurs, and small and micro business practitioners. The platform is built on a scheduling framework that helps users call upon multiple large models to solve various complex tasks. In the long run, a large number of large models will be integrated into WorkBuddy to serve users, and developers from various industries will create a vast array of skill plugins. The core goal of the platform is to address various efficiency pain points by integrating all available tools and large models in the market to achieve this.

Of course, at this stage, we play the role of coordinating scheduling, matching users with the most suitable large models and optimal skills, ensuring tasks are completed perfectly while controlling usage costs. Hunyuan will be one of the large models integrated into WorkBuddy. As long as Hunyuan can efficiently meet a large number of user needs, it will become the core underlying model of WorkBuddy, but it will not be the only available model on the platform.

UBS Group Analyst Kenneth Fong: Good evening, management, thank you for taking my questions. My first question is about the development progress of Xiaowei. Can management share preliminary user feedback from the prototype testing phase and the challenges encountered at this stage? From a commercialization perspective, how should we assess its net monetization potential? We have a concern: does the AI agent simplify the transaction path, which might just transfer transactions that users originally completed manually in the mini-program to the agent, only increasing computing costs without bringing substantial new total merchandise transaction volume? Additionally, will Xiaowei's shortening of the user transaction chain squeeze the originally high-margin advertising exposure? Liu Chih-Ping: I believe that the risks you mentioned will not materialize. We believe that AI will make the WeChat ecosystem smarter. With AI capabilities, users can automatically complete transactions, browse content, and handle various daily tasks. The already rich and well-functioning WeChat ecosystem will provide greater value to users. We can make an analogy: in the computer era, QQ was just a communication tool, but with the advent of the mobile internet era, WeChat was born, and the WeChat ecosystem amplified the overall value of QQ by more than ten times because it is a mobile-first product. Now, as we enter the AI era, the WeChat ecosystem is facing a second significant growth opportunity, evolving into an application and ecosystem centered around AI. Currently, user operations require manual text input and repeated clicks to navigate pages; in the future, users will only need to give a command to Xiao Wei, and Xiao Wei will automatically complete transactions and execute all demands, providing users with a disruptive experience while empowering all merchants within the WeChat ecosystem.

Our self-developed visual language model was designed with three core goals in mind: user privacy, operational costs, and compatibility with the WeChat AI environment, which can meet all existing user needs. As long as we implement this product experience and keep operational costs within a reasonable range, we can adapt WeChat to the AI era at a controllable cost. At that time, the scale of the WeChat ecosystem will continue to expand, relying on WeChat's existing advertising, payment, and e-commerce monetization systems, naturally creating massive new value. After the prototype of Xiao Wei goes live for testing, we are increasingly confident that this development vision can be realized.

UBS Group Analyst Kenneth Fong: For the second question, I would like to ask about the AI cloud business. The pricing of domestic large model tokens continues to decline, the speed of industry homogenization is rapid, and the domestic cloud market is highly sensitive to prices overall. Compared to traditional infrastructure cloud and platform cloud businesses, how is Tencent AI Cloud's gross margin level currently? As the penetration rate of AI gradually increases, what will be the trend of gross margin changes in the future?

Michelle: The pricing of tokens in the domestic market is indeed low, but the underlying production costs of domestic tokens are also very low, far below the levels generally recognized by the market and calculated by external institutions. Even with low terminal prices, the token business can still achieve positive gross margins, primarily because costs are controllable. You can check the gross margin of WorkBuddy's paid users and the model-as-a-service business, which is currently on par with Tencent Cloud's overall gross margin. Of course, the overall comprehensive gross margin of WorkBuddy is low because we allocate a portion of free users as subsidies to capture market share and drive user growth; however, when looking solely at the paid user group, the current gross margin performance is quite impressive.

The domestic cloud market is indeed fiercely competitive in terms of pricing, but the industry environment has changed significantly in recent months, with the procurement costs of hardware materials, especially memory, continuously rising. Therefore, we have adjusted our pricing standards for customers, and in May of this year, Tencent Cloud raised prices across its entire product line; in addition to the apparent price increase, we also significantly reduced customer discount levels. Overall, the pricing competition pressure in the domestic cloud market is now much lower than in the past Goldman Sachs Analyst Ronald Keung: Thank you to the management. I have two questions. First, the Hunyuan 3 focuses on high cost-performance and outstanding intelligent capabilities. What are the differentiated positioning features of the upcoming Hunyuan 4? As the competition in the trillion-parameter large model track becomes increasingly crowded, in which specific areas will Hunyuan 4 build exclusive competitive barriers?

The second question is about the current significant strategic adjustments by U.S. tech giants, shifting from application development to enhancing cloud computing capabilities. How does Tencent plan its capital expenditures related to AI? At what stage will we prioritize more capital expenditures towards cloud business, treating it as a key investment area with high returns? What are the similarities and differences between Tencent Cloud's transformation strategy and that of overseas tech companies?

Liu Chiping: First, let's look at the current Hunyuan 3. According to current industry standards, it belongs to the medium-small parameter model category, but its application range is very broad. Hunyuan 3 has two core characteristics: first, its performance can match or even surpass competitors with much larger parameter sizes; second, the R&D focus is on real business scenarios rather than merely pursuing scores on evaluation lists. Because of this, its practicality in real-world applications far exceeds that of competitors with the same specifications or even larger specifications. We will fully carry this R&D approach into Hunyuan 4.

The parameter size of Hunyuan 4 will be comprehensively enhanced, with performance surpassing that of larger competitor models, and its practicality will achieve a qualitative leap compared to Hunyuan 3, helping us enter a new stage and providing users with stronger intelligent capabilities. Additionally, all our AI products are collaboratively developed with large models, and after the launch of Hunyuan 4, the functionality and practicality of all products equipped with it will be significantly upgraded, bringing substantial growth dividends to related businesses. This is our iterative roadmap; Hunyuan 4 is just a milestone, and we will continue to iterate towards Hunyuan 5. Throughout the continuous iteration process, we will keep approaching the industry's top standards and will definitely reach a leading level in the future.

Once we achieve industry-leading capabilities, we will build a multi-specification gradient model matrix, with different parameter scales adapted to different costs and user needs, addressing pain points in various specific scenarios. At the same time, multiple differentiated models will be adapted to different product lines such as coding and office applications, enriching product functionality while balancing model performance, product capabilities, and task execution speed. This is our long-term plan.

Michelle: Your second question is about the logic of capital expenditure allocation among different businesses like Tencent Cloud. As Liu Chiping just mentioned, the top priority for capital expenditure in the coming months is to train larger-scale, higher-performance Hunyuan series large models; the second important use is to provide inference computing power to support Hunyuan, Deep Exploration, and other third-party models integrated with WorkBuddy.

Our significant investment in WorkBuddy's core strategic purpose is to promote the widespread adoption of this strategically significant core application while continuously collecting scenario feedback to optimize the large model and the entire Tencent ecosystem; at the same time, this product can immediately generate cash revenue. From an accounting perspective, most of the user spending on WorkBuddy is subscription fees. Similar to the business logic of games, there is a long time lag between user cash receipts and revenue recognition in financial reports, but currently, the product's cash revenue is growing rapidly and will gradually convert into Tencent Cloud's financial report revenue within the year By the end of this year and the beginning of next year, our graphics processing unit (GPU) computing power reserves will reach a sufficient scale, at which point we can expand two types of external businesses: one is bare-metal GPU leasing, and the other is model-as-a-service sales. However, among all the monetization channels for computing power, the token billing business generated by WorkBuddy can create the most long-term and stable commercial value for us, which is also the core reason why we prioritize allocating computing power to it at this stage.

Citi Bank Analyst Alicia Yap: Good evening, management. Thank you for taking questions, and congratulations on the company's solid performance. The first question is about Xiaowei. Management previously mentioned the intelligent agents and the transaction closed loop of intelligent agents. Could you elaborate on this concept? Is this model the ultimate direction for WeChat to build a fully autonomous intelligent agent ecosystem in the long term? Management also mentioned exploring local inference on the device side for Xiaowei. What benefits can this solution bring, and what challenges does it face? Why is developing a proprietary visual language model more suitable for Xiaowei rather than integrating external third-party large models?

I have a brief follow-up question: This quarter, advertising service revenue growth has increased to 22%. The AI advertising placement tools and automated placement systems are still iterating. Can this system continue to support growth momentum? After the deep implementation of Hunyuan 3, what long-term incremental benefits can the advertising business gain?

Liu Chao Ping: Regarding intelligent agents and the transaction closed loop of intelligent agents, this is the long-term goal we are planning. In the future, users will issue commands and complete transactions through Xiaowei and various AI intelligent agents. In the past, within the WeChat ecosystem, users needed to manually browse content and operate mini-programs; in the future, users will only need to issue a complex command to their dedicated intelligent agent, which can automatically complete the entire transaction process. The vast majority of mini-program merchants will also deploy dedicated merchant intelligent agents, and in the long run, merchant intelligent agents can directly connect with user intelligent agents. In the longer term, each user will have a dedicated AI intelligent agent, and intelligent agents can directly interact with each other to automatically complete transactions. This is a scenario that can be realized in the future, and we are building the underlying technical architecture for this system in phases.

Now, regarding local inference on the device side: local inference will be implemented in phases, and only after long-term development will most inference tasks be transferred to run on local devices. In the short term, a hybrid inference architecture will be adopted, with some calculations done locally and some remaining in the cloud. In the future, as the local hardware computing power of mobile phones and computers continues to improve, and as large models become more lightweight and operational efficiency continues to optimize, the inference tasks carried out by local devices will increase. This actually returns to the norm in the computer and smartphone industry: currently, ordinary central processing units (CPUs) complete the vast majority of calculations locally, with the cloud only undertaking a small portion of computing tasks.

However, in the early stages of AI infrastructure development, model operation requires extremely strong computing power; currently, the local computing power, hardware costs, and energy efficiency of mobile phones cannot support large-scale AI inference, so all calculations are currently placed in the cloud. In the future, mobile phones and computers will be equipped with more powerful GPUs, at which point a large number of inference tasks will be transferred to local devices. At that stage, the value of software and large models themselves will significantly increase, and the return on investment for models and applications will also rise because the capital expenditure for computing power will no longer be solely borne by large model vendors, but will be shared by the entire ecosystem This trend is bound to come, and we have already made early arrangements for relevant technological reserves.

Michelle: Regarding the advertising business, our advertising revenue growth has experienced fluctuations in the past and will continue to fluctuate in the future. We cannot simply linearly extrapolate the growth trend, as there are multiple reasons behind it. Previously, I mentioned that in-game ads for mobile games would drag down the revenue of the overseas gaming segment; conversely, this type of mobile game advertising contributed two percentage points to the growth of the advertising segment this quarter. In-game ads for mobile games are an emerging category for Tencent and the entire industry, and their future impact on traditional advertising business growth remains uncertain.

In addition, the overall trend of the domestic new consumption market and advertising market is volatile, and pressures from the macro economy and household consumption levels will affect the pace of advertising spending. Even so, our advertising business growth continues to outperform the domestic advertising market, and we will continue to significantly lead the industry in the future. The supporting logic has three points: first, the AI advertising precision delivery system continues to be implemented, and growth dividends are continuously released; second, core traffic, especially video accounts, sees steady increases in user engagement and advertising inventory; third, the high-conversion closed-loop advertising model is still in its early development stage, which can drive advertising prices upward in the long term.

Bank of America Merrill Lynch analyst Alex Liu: I have only one question. We noticed that the company has increased its stock buyback efforts since May, while capital expenditure has significantly accelerated. We understand that the current AI investment cycle is still relatively early. I would like to ask how investors should view Tencent's capital allocation priorities over the next twelve to twenty-four months?

Michelle: The capital allocation strategy will be dynamically adjusted and flexibly changed according to the market environment. If we determine that increasing capital expenditure on computing power, self-developing large models using computing power, leasing computing power for the WorkBuddy token business, and selling model-as-a-service can yield excess returns, we will allocate more cash to capital expenditure and correspondingly reduce the scale of stock buybacks, maintaining an overall dynamic balance.

Liu Chiping: One more point I need to emphasize: the capital expenditure directed towards AI-native businesses is mainly a one-time concentrated investment this year and next year. Please do not assume that we will maintain the same level of large investments every year. Model training is a fixed cost; we only need to reserve sufficient computing power and do not need to continuously increase investment every year. Inference computing power does require sufficient reserves to support token monetization and computing power leasing business, but we will only continue to invest if this business can achieve considerable returns; if the returns do not meet expectations, the existing investment scale will be capped. Future additional capital expenditures will correspond entirely to the revenue generated by the business.

The source of funds for this one-time computing power investment cannot only be viewed from operating cash flow; it must also consider cash reserves on the balance sheet, the scale of investment assets, and new operating cash flow to maintain a stable and controllable investment scale. The above dimensions will jointly determine the funding arrangement for the initial computing power investment.

JP Morgan analyst Alex Yao: Thank you to the management team. The first question is about the top-level strategy of Hunyuan. Hunyuan 3 focuses on high cost-performance rather than extreme hardware performance. If we successfully develop a truly industry-leading large model with larger parameters and higher training and operational costs, what commercial value can it create that Hunyuan 3 cannot currently achieve? Is it a stronger WeChat intelligent agent, improved advertising performance, or increased revenue from government and enterprise clients? Can these potential gains support a significant increase in model training investment over the next twelve months?

Liu Chih-Ping: Let me clarify the logic: the research and development positioning, strategic route, and Hunyuan are completely distinct. The Xiaowei intelligent agent does not rely on the top performance of Hunyuan. We have repeatedly emphasized that the core design of the Xiaowei underlying model revolves around user privacy, adapting to all intelligent interaction needs within the WeChat ecosystem, and controlling operational costs; this is its exclusive positioning.

However, achieving industry-leading standards with Hunyuan can bring multiple commercial values: first, building a large-scale token payment business; second, empowering WorkBuddy to undertake more complex and higher value-added office services, creating more revenue for clients. Our positioning for WorkBuddy is not just as a corporate tool that replicates existing manual office operations; we continuously explore high value-added scenarios to create incremental value and revenue for users, even directly helping clients increase their income. Once this is achieved, a large number of new business models can be realized, which is the core value brought by top large models.

At the same time, once Hunyuan reaches industry-leading standards, we can derive a layered gradient model matrix, creating specialized models that adapt to different cost ranges and segmented tasks on a top-tier technological foundation, covering users' diverse intelligent needs. The operational costs of different tiered models vary, but with our self-developed models and independent control over inference costs and computing resources, each product layer can achieve stable profitability; this is our long-term goal for the next generation of large models.

JP Morgan analyst Alex Yao: Second question: How does the management control the investment in AI-related products, which increased from 8.8 billion yuan in the first quarter to 10.5 billion yuan in the current quarter? Is there a fixed spending cap, a revenue threshold, or a strategic approach of continuous investment without regard for short-term returns? Under what user data, revenue growth, or unique profitability signals will the company further increase investment or transition products from pure investment phases to profitability phases?

Liu Chih-Ping: Currently, our AI investment strategy is flexible and dynamic; we will maintain cautious investment. If we clearly see explosive growth opportunities, we will increase our investment; this is our core control logic. Overall investment will be controlled within a fixed proportion of the group's profits, but if we can predict that the business will generate huge returns, we will increase our investment.

We always believe that AI is a long-term track and will continue to lay out long-term strategies. As scale expands, commercial returns will gradually materialize, ultimately achieving profitability. More importantly, we have a bottom-line plan: even if we stop self-developing applications and simply lease computing power externally, the related business itself can be profitable, which is why we dare to invest long-term.

Michelle: We will dynamically adjust the funding allocation priorities within the total budget cap. Comparing the distribution structure of the 8.8 billion yuan AI investment in the first quarter and the 10.5 billion yuan investment in the second quarter shows significant changes: we found that WorkBuddy experienced explosive growth, so we significantly shifted funds towards it while reducing the budget for other products in the AI product line; this dynamic adjustment mechanism will be continuously implemented Morgan Stanley analyst Gary Yu: I have two questions, both regarding AI investment. I understand that the company's investment priorities are model training and WorkBuddy, followed by cloud computing power leasing. What level is Xiaowei at in terms of computing power resource allocation, given that it requires supporting computing power? The second question is about the commercialization timeline and profitability visibility of various AI innovative businesses. Can it drive overall positive growth in profitability in the short term? When will the overall operating profit growth that includes AI investment exceed the profit growth after excluding AI investment?

Liu Chih-Ping: The annual investment budget for Xiaowei's continuous operation is lower than the normalized investment scale of previous years. Overall costs are completely controllable. As product experience continues to optimize, business revenue will quickly materialize and soon cover the investment costs.

Regarding profitability guidance, we will not provide precise quantitative targets at this stage. We have fully articulated our overall judgment on AI business and related investments, and the investment scale will have an upper limit, continuing Tencent's consistent prudent and disciplined management style. However, if we anticipate that we can create a large-scale, long-term profitable AI business, we will also moderately increase our investment. We always have a bottom-line plan in mind: we can quickly lease idle computing power to Tencent Cloud to generate revenue, profit, and investment returns, with investment risks completely controllable. (End)