In the second quarter, Tencent's capital expenditures increased nearly threefold year-over-year to RMB 52.8 billion, all directed toward AI infrastructure construction. Management explicitly stated that computing power rental can be profitable without incurring losses as an alternative option. The gross margin of Work Buddy's paying users has already matched the overall gross margin of Tencent Cloud, and Yuanbao has also achieved healthy margins. WeChat ecosystem AIization is strategically positioned as a key vehicle, with long-term potential for a 10x value amplification. Hunyuan will continue to iterate toward industry-leading standards.Source: Wall Street Journal
Tencent is entering a critical phase where AI investment and commercialization are being advanced in parallel: its core businesses continue to generate profits, providing financial support for intensive investments in AI infrastructure and native applications.
On August 12, during Tencent's Q2 2026 earnings call, President Liu Chiping stated that the company is making substantial investments in AI computing power and has already identified clear upside potential for returns. On one hand, several AI-native applications have performed well; on the other, leveraging computing power for cloud rentals is also expected to generate significant revenue and improve capital expenditure efficiency. For certain computing power orders placed previously, the resale price could exceed the original purchase price by more than 30%.
Tencent is building its AI infrastructure with unprecedented intensity. Capital expenditures in the second quarter rose to RMB 52.8 billion, significantly exceeding the expected RMB 32.1 billion and nearly tripling the RMB 19.1 billion recorded in the same period last year; free cash flow stood at negative RMB 13.8 billion. The company stated that, excluding advance payments for AI computing power purchases, free cash flow would have been RMB 37.6 billion. This indicates that the current cash flow pressure stems primarily from upfront investments in AI infrastructure, not from a deterioration in traditional business operations.
This round of investment primarily supports the upgrade of the Hunyuan model, inference demands for Work Buddy and CodeBuddy, WeChat’s AI strategy, and growing cloud service needs from external customers. From management’s perspective, AI infrastructure is a foundational investment for future business growth: first, build intelligent capabilities through state-of-the-art models; then, transform these capabilities into user demand via AI applications; ultimately generating commercial returns.
Regarding the pace of AI investment, Liu Chiping said the current situation remains highly dynamic. Tencent will maintain a relatively cautious approach until a clear opportunity for significant growth emerges; once such an opportunity is identified, the company will increase its investment—such as with WorkBuddy—while maintaining a long-term focus on returns. Over time, the economic benefits of AI initiatives will gradually materialize and ultimately lead to profitability.
In terms of commercialization, Tencent is exploring ways to monetize through AI applications such as selling tokens, while retaining GPU rental as an “alternative option.” Liu Chiping stated that if Tencent were to directly shift its business model to renting computing power, it would not only avoid losses but also generate profits. Tencent has always had this alternative option, giving it “greater confidence.”
Work Buddy: The AI "coordinator" has a clear positioning, and the paid segment has already achieved a healthy gross margin.
At the AI application level, management highlighted Work Buddy, which is not a single model but rather an AI "coordinator" that integrates various models and skills to select the most appropriate solution for user tasks, balancing task quality with cost efficiency.
Hunyuan will be one of the models supported by Work Buddy, but not the only option. If Hunyuan leads in performance and ratings on actual tasks, it has the potential to become the primary model on the platform. Meanwhile, more models and skills provided by future developers will be integrated to meet diverse user needs.
This means Tencent is not trying to rely on a single model to handle everything, but rather aims to enhance the overall efficiency of its AI services through combinations of models, skills, and applications.
From a profitability perspective, Work Buddy has shown healthy commercialization signals. James Mitchell, Chief Strategy Officer of Tencent, stated that the gross margin of Work Buddy’s paying users is already comparable to Tencent Cloud’s overall gross margin. Work Buddy’s overall gross margin is slightly lower, primarily due to the inclusion of free users, as Tencent continues to subsidize to expand market share; however, the paid segment has already generated a “quite good” gross margin.
Hunyuan surges toward SOTA, with model and product evolving in tandem
Regarding the models, management revealed that Hunyuan 4 has reached a key milestone, and the next phase will involve upgrading to Hunyuan 5, with continued efforts to approach the state-of-the-art (SOTA) level in the industry.
The management stated that once SOTA is achieved, Tencent will establish a matrix of models covering various sizes, meeting the needs of different users across a range of cost and performance levels. This strategy emphasizes matching model sizes with specific use cases, rather than solely pursuing "large models."
Meanwhile, Tencent is driving the co-design of models and products. As Hunyuan 4 and Hunyuan 5 continue to evolve, the capabilities, feature richness, and execution speed of related products will also improve in tandem. For Tencent, the deep integration of model capabilities with application experience will be a critical step in transforming AI investments into commercial returns.
WeChat Mini Program: AI-Driven Transformation of the WeChat Ecosystem
For WeChat's AI assistant, Xiao Wei, currently in internal testing, Liu Chiping positioned it as the key vehicle for WeChat's transition into the "AI era," comparing its potential to the tenfold value increase WeChat achieved over QQ.
In the long term, Tencent aims for users to simply issue natural language commands, allowing WeChat Mini Program to execute transactions and operations on their behalf; at the same time, merchants and mini-programs within the WeChat ecosystem will also deploy their own agents to enable automated interactions between user agents and merchant agents.
Regarding cost control, Liu Chiping stated that the ongoing inference costs for Wee will be lower than Tencent’s previous investments in Yuanbao, and emphasized that the design of its VLM (Vision-Language Model) was intended to balance privacy protection and cost efficiency within the WeChat environment.
Regarding concerns that agent-based trading might erode ad inventory, he believes that WeChat's AI transformation will generate incremental value rather than simply replace existing business models.
Yuanbao: From Subscription Revenue to Token Commercialization
The commercialization of Yuanbao is also beginning to reveal a clearer economic model. Tencent management disclosed that the current gross margin of Yuanbao's paying users is now comparable to the overall gross margin of Tencent Cloud.
Management also noted that, due to the revenue recognition mechanism similar to that of gaming business in the subscription model, there is a timing difference between current cash inflows and revenue recognition in the financial statements. Therefore, the cash revenue growth observed today will gradually translate into reported revenue growth for Tencent Cloud over the course of the year.
Among the three cloud business opportunities—GPU bare-metal leasing, Model-as-a-Service (MaaS), and Yuanbao Token production—management believes that Token production offers the most enduring economic value and therefore has the highest priority. For free users, the company continues to expand its market share through subsidies, while paying users already provide a healthy gross profit margin base.
Capital Allocation: Dynamic Balance Between AI Investment and Buybacks
Facing a significant increase in AI-related capital expenditures, Tencent does not view share repurchases and AI investments as a fixed trade-off, but rather emphasizes that capital allocation will be dynamically adjusted based on returns.
Management stated that if AI capital expenditures demonstrate significantly better returns than stock buybacks, the company will increase its capital spending accordingly.
Liu Chiping emphasized that investment in AI infrastructure is essentially a finite, upfront capital expenditure, not a linear annual increase. Therefore, evaluating this expense should not rely solely on current free cash flow but must also take into account cash on hand, portfolio value, operating cash flow, and reasonable debt capacity.
For Tencent, the core logic behind its current AI investments is shifting from "how much to invest" to "how to generate returns." As compute leasing, token production, and AI-native applications gradually commercialize, Tencent aims to transform its earlier infrastructure investments into a new growth curve.
Full transcript of the phone call
Good afternoon, good evening. Thank you for your patience. Welcome to the Tencent Holdings Limited Second Quarter 2026 Earnings Conference Call. I am Huang Wendi from Tencent’s Investor Relations team. At this time, all participants are in listen-only mode. Following the management presentation, we will conduct a question-and-answer session for participants joining by phone. If you would like to ask a question, please press 5 on your phone. If you are joining via Tencent Meeting or the meeting app, please click the “Raise Hand” button in the lower-left corner. Please note that today’s conference call is being recorded.
Before we begin the presentation, we would like to remind you that it contains forward-looking statements subject to various risks and uncertainties that may not materialize in the future. Information regarding general market conditions is sourced from multiple external sources outside Tencent. This presentation also includes certain unaudited non-IFRS financial metrics, which should be considered supplementary and not as a substitute for the group’s financial performance metrics prepared in accordance with IFRS.
For a detailed discussion of risk factors and non-IFRS measures, please refer to our disclosure documents in the Investor Relations section of our website. Now, let me introduce the management team on the webinar. Tonight, Mr. Ma Huateng, Chairman and CEO, will begin with a brief overview. Mr. Liu Chi Ping, President, will provide a strategic review, Mr. James Mitchell, Chief Strategy Officer, will present a business review, and Mr. Luo Shuohan, Chief Financial Officer, will conclude with a financial discussion. We will now hear from Mr. Ma Huateng before beginning the Q&A session.
Ma Huateng, Co-founder, Chairman of the Board, and Chief Executive Officer:
Thank you, Wendy. Good evening, everyone. Thank you for your participation. Entering the third quarter of this year, we have made significant progress in building a new Tencent powered by AI, across intelligent capabilities, applications, and infrastructure. On the intelligent side, the production version of HunYuan 3 provides users with a high-performance, cost-effective general-purpose model and lays the foundation for the HunYuan model family to achieve state-of-the-art capabilities in the future. On the application side, our WorkBuddy AI productivity service and CodeBuddy AI coding tool are achieving breakthrough user growth and are now clearly the leaders in this space in China. On the infrastructure side, we have substantially increased our procurement of computing resources, enabling us to convert application and model usage into future revenue. Meanwhile, we continue to strengthen our existing services through steady growth in marketing service revenue, several recently successful game launches, and the rapid growth in video views on WeChat Video Channels. Looking at our second-quarter financial results, total revenue reached RMB 205 billion, an 11% year-over-year increase. Gross profit was RMB 118 billion, up 13% year-over-year. Non-IFRS operating profit was RMB 76 billion, up 9% year-over-year. Excluding new AI products, non-IFRS operating profit was RMB 86 billion, up 19% year-over-year. Non-IFRS net profit attributable to equity holders was RMB 68 billion, up 9% year-over-year. Turning to our key services, the combined monthly active accounts for WeChat and WeChat reached 1.4 billion, growing both year-over-year and quarter-over-quarter. In digital content, Tencent Music Entertainment Group’s acquisition of Himalaya has strengthened our audio content and unlocked new synergies with ecosystem IPs—including阅文集团 (China Literature). In gaming, our new title “Luoke Kingdom: World” ranked number one among all new games launched in China this year by both average daily active users and total revenue. In cloud services, WorkBuddy recently ranked number one in China among AI productivity services based on monthly interactions. I will now hand the floor to Martin for a strategic review.
Liu Chi-ping, President:
Thank you, Pony. Good evening and good morning, everyone. Today, we’d like to share the latest progress on our overall AI strategy. Tencent’s existing businesses are growing steadily, driven by their inherent moats and AI enablement. As discussed earlier this year, our moats stem from factors such as network effects, supply chain depth and value-add, IP, low churn rates, regulatory requirements, and proprietary data. Beyond these moats, we are further deploying AI to enhance returns across businesses including gaming and advertising. As a result, our existing businesses provide strong financial support for our new AI initiatives. Regarding our new AI efforts, we’ve made significant progress in building a solid foundation, including a significantly improved, market-leading cost-performance foundation model—HunYuan 3—market-leading AI productivity tools WorkBuddy and CodeBuddy in China, and the consumer-facing gateways Yuanbao and Weixiao, designed to drive broader AI adoption.
We have observed increasing potential for attractive financial returns from our differentiated proprietary products, including HunYuan, WorkBuddy, and Xiaowei. Over time, we have become increasingly confident in our large-scale AI investments, as they offer not only substantial upside potential but also clear downside protection. Our AI investments are primarily focused on AI infrastructure; in the worst-case scenario—which we believe is unlikely—we can choose to lease out this infrastructure via Tencent Cloud at cost recovery or even better terms. Now, let’s turn to the individual components. First, regarding the HunYuan foundational model, the full production release of HunYuan 3 has been highly successful, delivering significant performance improvements over the HunYuan 3 preview version. By leveraging feedback loops from the product team to enhance the quality and diversity of post-training data, and by scaling up reinforcement learning, HunYuan 3 has achieved marked improvements in task completion rates and meaningful reductions in hallucinations and error rates.
The enhancements in Qwen3 are most evident in its agent capabilities and product experience. The model’s significant performance improvements in reasoning, agent tasks, and long-context handling deliver clear advantages for use cases such as coding, office productivity, financial modeling, and frontend design. These advancements have accelerated user adoption and driven growth in external customer demand, validating its practical utility—evidenced by Qwen3’s daily average token usage across all channels during its paid period increasing by approximately six times compared to the preview version. Additionally, according to token usage rankings on OpenRouter, Qwen3 has consistently ranked among the top three globally. The production version of Qwen3 performs robustly, laying the foundation for the Qwen model series to achieve state-of-the-art capabilities in the future while delivering the cost-efficient performance users need today.
We have been integrating Hunyuan into our products, significantly impacting WorkBuddy. Hunyuan enables complex agent workflows, improving task success rates and reducing completion times. For Yuanbao, Hunyuan delivers leading execution quality in information retrieval, data processing, document workflows, and daily decision-making. In gaming, we are leveraging Hunyuan to create AI teammates and conduct code reviews, including for our flagship game, Peacekeeper Elite.
We have deployed Hunyuan to support our official account’s AI assistant and the developer tools for our mini-program. Meanwhile, product integration continuously feeds real-world product usage and domain feedback into model training, improving Hunyuan over time. Our model-product co-design approach enables Hunyuan to validate model accuracy, identify and handle edge cases, enabling faster model iteration and continuous performance improvements. After establishing a new rapid model iteration system and validating it with Hunyuan 3, we are accelerating model enhancements. We are expanding more powerful reinforcement learning to significantly upgrade the model after pre-training, and we are also upgrading its multimodal capabilities.
More importantly, we are training a larger-parameter model, Hunyuan 4, which is expected to be released later this year. By accelerating technological iteration and pushing the boundaries of model intelligence, we believe Hunyuan’s capabilities will reach state-of-the-art levels. We are confident that building a large, valuable AI-native business for Tencent will generate significant returns. The reason we invest in our own foundational models is that through co-design across our applications, models, and computing infrastructure, we can achieve better unit economics, more innovative features, and greater access to intelligent value.
In the early stages of AI adoption, we continue to focus on the application level, where our AI-powered office productivity workspace, WorkBuddy, and coding tool, CodeBuddy, have achieved breakthrough success in terms of capabilities and user growth. They are China’s leading office productivity services based on monthly interactions. WorkBuddy serves as an all-in-one workspace that coordinates multiple agents to handle complex tasks end-to-end.
Users can remotely control WorkBuddy via WeChat, Enterprise WeChat, and VPC, and access over 70,000 skills from the Tencent Cloud Skills Center. Beyond rapid user adoption, WorkBuddy achieves high retention and strong user willingness to pay because it directly enhances user productivity. It also attracts a growing and more vibrant developer community by embedding skill payments and proportional payouts directly into task flows, compensating developers when skills are invoked. This progress supports our belief that significant opportunities remain untapped in productivity markets, including coding and existing office scenarios. We are currently focused on investing in market education and expanding our market leadership. Over time, the product’s economics will become increasingly attractive, as enhanced premium features will accelerate paid user growth, while we reduce token costs through improved agent efficiency, reasoning efficiency, and model optimization.
Given the widespread adoption of Tencent applications such as WeChat, Enterprise WeChat, and Tencent Meeting by businesses, WorkBuddy offers us a new way to monetize enterprise relationships. On the consumer side, we recently launched the prototype of Xiao Wei, which delivers an embedded, context-aware AI experience within WeChat, leveraging WeChat’s social graph, knowledge graph, rich merchant information, and payment capabilities. Xiao Wei is powered by WeChat’s custom VLM model, designed with a focus on user privacy, WeChat-specific use cases, and cost efficiency. Xiao Wei helps users navigate and extract insights from WeChat’s diverse content ecosystem in a personalized and efficient manner. It also leverages WeChat’s unique mini-program ecosystem to help users discover products, make purchasing decisions, and place orders, laying the foundation for agent-to-agent transaction cycles. Although the prototype is technically capable of handling advanced agent workflows, as a safety measure, we currently configure Xiao Wei to require user intervention and multi-step confirmation. Xiao Wei will be rolled out gradually to a broader user base as we advance several core initiatives to enhance the user experience.
These initiatives include enhancing Xiao Wei’s conversational, memory, and recommendation capabilities, expanding service and content integrations, scaling our AI infrastructure, and upgrading our systems to support a significantly larger user base. As we upgrade WeChat for the AI era, we can achieve this in a cost-efficient manner, and we believe AI will accelerate the growth of the entire WeChat ecosystem over time, thereby accelerating its monetization and delivering attractive returns. Turning to Yuanbao, we are focused on enhancing its capabilities and user experience, particularly in search, speech recognition, and text-to-speech functions. We are also improving its ability to meet consumers’ broader long-tail AI needs, including multimodal generation. Yuanbao plays a key role in the co-design of Hunyuan, as its conversational use cases generate valuable feedback that helps refine our Hunyuan model series. Over time, features developed and refined by Yuanbao can become atomic capabilities for other Tencent products, such as WorkBuddy, CodeBuddy, WeChat, and QQ Browser. Now, please welcome James.
James Michel, Chief Strategy Officer and Senior Executive Vice President:
Thank you, Martin. This quarter, total revenue increased by 11%, with contributions from social networks at 16%, domestic games at 23%, international games at 9%, marketing services at 21%, and fintech and enterprise services at 30%. Gross profit increased by 13%, with gross profit from value-added services up 14%, marketing services up 21%, and fintech and enterprise services up 9%. Revenue from value-added services amounted to RMB 98 billion, an 8% year-over-year increase. Revenue from social networks rose 1% year-over-year to RMB 32 billion, primarily driven by increased sales of in-app game items, partially offset by a 6% decline in long-form video subscription revenue.
However, our exclusive series "The Lead" is the most-watched series across all video platforms in China. In the second quarter, audio subscription revenue increased by 8%, driven by higher average revenue per user for music and a richer content library. By integrating Himalaya, we enabled users to access QQ Music with a single click from their video accounts, enhancing discovery of new music. In May, we completed the acquisition of Himalaya. By incorporating Himalaya into the Tencent ecosystem, we can strengthen the resilience of Tencent Music Entertainment Group, deepen content supply relationships between Wenxue City and Himalaya, and provide users with new content formats such as audiobooks and podcasts. Domestic game revenue increased by 17%, primarily driven by "Delta Action," the PC version of "Valorant," the mobile version of "Valorant," and "Luoke Kingdom: World." International game revenue declined by 1%, but increased by 4% on a constant currency basis, as revenue growth from "Mingchao" and the PC version of "Valorant" was offset by declines in revenue from two Supercell games. In communications and social networking, total time spent on Video Channels grew by over 20% in the second quarter, benefiting from a richer content offering, enhanced interactivity, and the introduction of a new multi-variable content ranking system.
We expanded content appealing to younger users by partnering with game studios, music labels, and TV show IPs, while providing creators with new revenue-sharing opportunities and broadening the group of creators who can earn directly from video accounts. Gross merchandise value (GMV) for mini-stores increased, with significant growth in GMV from WeChat’s centralized e-commerce gateway page. For mini-store merchants, we introduced marketing tools such as lotteries to help them enhance brand awareness and drive product discovery; for mini-store consumers, we enhanced rewards for returning customers to increase customer lifetime value and thereby boost merchants’ customer lifetime value.
In the domestic gaming sector, Delta Action achieved its highest-ever average daily active users in the second quarter, driven by the Burst Fest event, the game’s first professional esports finals, and a global 20-vs-20 tournament. On the production side, the Delta Action team has integrated AI into multiple workflows, including using data agents for performance analysis and employing Hunyuan 3D models for asset generation. The PC version of Valorant also reached its highest-ever average daily active users in the second quarter, benefiting from the Skirmish Ascension mode featuring rotating progressive weapons and the Summit map with destructible walls. The game expanded its reach through influencer collaborations during the Ground City event and promotions across more than 10,000 internet cafes. Among new releases, Rock Kingdom: World ranked fifth by average daily active users and eighth by total revenue among all mobile games launched in China during the second quarter, making it the highest-ranking new game released year-to-date. Since its launch, the game has maintained a rapid content delivery pace, adding 100 creatures and expanding into seven new regions. On July 9, we released Runaway Evolution. This game is an adaptation of the PC survival open-world crafting game Rust, which has consistently ranked among Steam’s top 20 most-played games over the past eight years due to its unique high-risk, high-reward gameplay—where players compete in week-long matches to be the last one standing. Runaway Evolution adapts this gameplay for the Chinese market by offering mobile and PC platforms and introducing sandbox safe zones for new players.
In our international games, League of Legends' daily active users increased year-over-year in the second quarter, primarily driven by the ARAM Mayhem mode. We launched League Classic, a nostalgic mode that re-engages longtime fans by recreating early-game mechanics, including pre-rework champions, classic runes, and the original Summoner’s Rift map layout. Warframe’s daily active users grew year-over-year, and total revenue reached an all-time high this quarter, supported by the new Wolf-themed Prime Warframe and the new story arc, Jade Shadow’s Constellations. Arrow’s Puzzle Escape, a maze-clearing game developed by Miniclip subsidiary Lessmore, was the most downloaded mobile game globally in the second quarter. Arrow’s success demonstrates that Miniclip’s studio family, backed by Miniclip’s publishing expertise, can collaboratively innovate and lead the development of new casual game genres. Arrow is monetized through in-app advertising; therefore, its revenue is reported under the marketing services segment rather than the international games sub-segment. When adjusting for Arrow and other in-app advertising game revenues, our international games year-over-year revenue growth rate would be 4 percentage points higher than the reported figure. In marketing services, revenue increased 22% year-over-year to RMB 44 billion, driven by higher cost per thousand impressions and increased impression volume. Most major categories increased their marketing spend on our platform, including e-commerce, internet services, and local services. We enhanced our AI Marketing+ end-to-end execution capabilities to better support closed-loop mini-store and mini-series advertisers.
For example, AI Marketing+ now enables mini-store owners to automatically select products for promotion, generate ad creatives related to those products, and run intelligent bidding to purchase inventory for these creatives. We have significantly expanded the parameters of the advertising AI recommendation system to more precisely capture user interests, thereby improving ad conversion rates. Video account ad impressions have grown rapidly year-over-year, driven by higher video views and improved ad load rates, although ad load rates remain well below the industry average for short-form video. Mini-programs are attracting increasing marketing spend from mini-drama and mini-game studios. Revenue from the fintech and enterprise services segment amounted to RMB 60 billion, representing a 9% increase. Fintech service revenue grew year-over-year, supported by growth in commercial payments, wealth management, and consumer lending services. In commercial payments, the number of transactions increased year-over-year, while the decline in average transaction value narrowed. In wealth management, aggregated client assets grew year-over-year, benefiting from the growing adoption of automated investment strategies and thematic index funds.
In terms of enterprise services, although we are still addressing capacity constraints, our cloud revenue growth rate accelerated from a high single-digit year-over-year increase in the first quarter to a low double-digit increase in the second quarter, driven by AI-related demand, international expansion, and increased usage and pricing of general cloud services. AI-related demand translated into revenue growth from GPU leasing, model-as-a-service, and increased usage of WorkBuddy and CodeBuddy tokens. Our international cloud business expanded rapidly, with skills developed through CodeBuddy enabling us to migrate customers to the cloud faster than ever before—for example, serving a leading telecommunications company in Indonesia. Now, please welcome John.
Luo Shuohan, Chief Financial Officer and Senior Vice President:
Thank you, James. In the second quarter of 2026, total revenue amounted to RMB 204.8 billion, representing an 11% year-over-year growth. Gross profit was RMB 118.4 billion, up 13% year-over-year. Operating profit reached RMB 67.3 billion, an increase of 12% year-over-year. Interest income was RMB 4.2 billion, up 2% year-over-year. Financial costs amounted to RMB 3.0 billion, compared to RMB 3.9 billion in the same period last year, reflecting favorable foreign exchange movements and lower interest expenses due to reduced average interest rates.
Our share of losses from associates and joint ventures in the second quarter of 2026 amounted to RMB 10 billion, primarily reflecting adjustments arising from the fair value revaluation of convertible redeemable preferred shares of an unlisted investee, driven by an increase in the investee’s valuation; this item has been excluded from our non-IFRS profit. On a non-IFRS basis, our share of profit from associates and joint ventures for the quarter was RMB 6.4 billion, compared to RMB 6.3 billion in the same period last year. Income tax expenses increased by 3% year-over-year to RMB 11.7 billion. On a non-IFRS basis, operating profit was RMB 75.6 billion, up 9% year-over-year. Excluding the new AI product, operating profit was RMB 86.1 billion, up 19% year-over-year. Net profit attributable to equity holders was RMB 68.4 billion, up 9% year-over-year. Diluted earnings per share were RMB 7.433, up 9% year-over-year.
Next, let’s look at gross margins. The overall gross margin for the second quarter was 58%, an increase of 1 percentage point year-over-year. By segment, the gross margin for value-added services rose 4 percentage points year-over-year to 64%, driven by a shift in revenue mix toward internally developed games with higher profit margins. The gross margin for marketing services was 57%, a slight decrease of 0.3 percentage points year-over-year, as revenue growth supported by our AI-driven marketing capabilities was largely offset by increased depreciation and operating costs related to expanding our AI infrastructure to improve advertising and content recommendations. The gross margin for fintech and enterprise services was 52%, remaining largely stable year-over-year. Regarding operating expenses, sales and marketing expenses amounted to RMB 11.9 billion, up 26% year-over-year, due to increased marketing spending to support our gaming business and drive adoption of AI-native products. Research and development expenses increased 35% year-over-year to RMB 27.2 billion, primarily reflecting higher R&D investments to support full-year model enhancements, the WeChat AI initiative, and the development of AI capabilities across products and services.
General and administrative expenses, net of R&D expenses, decreased by 1% year-over-year to RMB 11.5 billion. As of the end of the quarter, we had approximately 116,000 employees, representing a 4% year-over-year increase and a 1% quarter-over-quarter increase, primarily driven by growth in personnel for gaming and technology platforms, including AI-related roles. Our non-IFRS operating margin for the second quarter was 36.9%, a decrease of 0.6 percentage points year-over-year. Excluding AI products, our non-IFRS operating margin was 42%, an increase of 2.8 percentage points year-over-year. In summary, I will highlight key cash flow and balance sheet metrics.
Operating capital expenditures amounted to RMB 51.8 billion, representing a 190% year-over-year increase and a 66% quarter-over-quarter increase, as we accelerated investments in AI infrastructure to support model enhancements throughout the year, inference demands for WorkBuddy and CodeBuddy, the WeChat AI initiative, and the development of AI capabilities across products and services, while also meeting growing external demand for our cloud services. Non-operating capital expenditures totaled RMB 1 billion. Free cash flow was negative RMB 13.8 billion, reflecting substantial investments in AI infrastructure and AI-related advance payments, as well as seasonally lower total gaming revenue. Excluding advance payments for computing resource procurement, free cash flow would have been RMB 37.6 billion. Net cash position stood at RMB 58.2 billion, compared to RMB 146.9 billion as of March 31, 2026, reflecting capital expenditure payments of RMB 59.3 billion this quarter and dividend payments of RMB 41.6 billion for 2025. Thank you.
Q&A Session
Wendy Huang, Director of Investor Relations:
Thank you, John. We now begin the Q&A session. (Operator instruction) The first question is from Robin Zhu of Bernstein. Robin, your line is open.
Robin Zhu, analyst:
Thank you, Wendy. Thank you to management for the opportunity to ask a question. I’d like to ask about your recent quarterly capital expenditure of RMB 53 billion, which increased from the prior quarter and annualizes to over RMB 200 billion. If we simply multiply this by four, how should we think about the resulting depreciation and amortization costs? To what extent do you believe these will be offset by incremental revenue generated from AI investments? Or will this erode profits over the next few quarters? It would be very helpful to hear your thoughts on the payback period, particularly including the associated R&D costs. Thank you.
Tencent management:
Thank you for your question, Robin. Given the surge in computing demands and the resulting rise in rental costs, we could almost immediately recoup our depreciation expenses by leasing computing resources to third parties, as many new cloud businesses do, and achieve substantial returns in a short time. However, in reality, we are playing a different game or executing a broader strategy: we are allocating a significant portion of our new computing resources toward building our own models to achieve state-of-the-art performance and deploying and promoting our own AI applications to attain market leadership in China. We believe that by delivering the superior intelligence enabled by our state-of-the-art models and leading-market AI applications, we can, over time, convert this superior intelligence into exceptional economic returns—such as through token sales via the WorkBuddy app. This is the path we have chosen. To elaborate further on this, I believe you can currently view Tencent’s business as comprising two parts: one is our existing or legacy business, which is already generating solid growth and possesses a degree of operating leverage. This is the high-quality growth trajectory we have been building and will continue to pursue; the other is the new AI-native business we are constructing, which, as James mentioned, involves our own models, the new applications we are developing, and the corresponding computing infrastructure. Financially, you should focus on the revenue and profits from our core existing business, while we separately disclose investments in the AI-native business as an operational item. When you look at capital expenditures, I believe they too fall into two categories: one relates to our existing business—historically, this was simply the operating cash flow generated along with associated capital spending, and this segment still maintains strong cash flow generation; the other portion of capital expenditure relates to the new AI-native business, which essentially represents the upfront investment required to acquire computing resources for model training, prepare for inference demands, and secure additional resources to build our AI computing and AI cloud business.
Anonymous spokesperson:
So this is essentially how it works—we invested in all these computing resources because we needed them to launch our business. At the same time, when we made these investments, we saw clear upside potential, as our models are performing well, our new applications are performing well, and we currently have substantial demand for computing resources. If we can lease out these computing resources through Tencent Cloud, it would generate additional revenue and deliver significant returns on our capital expenditures. In fact, for some prepayments and computing orders we made months ago, we are now selling the capacity at profits exceeding 30% above what we originally paid. But we believe that by using these computing resources to build our own models and applications, and then allocating excess capacity for leasing, over time we will build a highly significant AI-native business that will generate substantial profits, high cash flow, and strong returns for Tencent. This is how we currently think about the business.
Robin Zhu, analyst:
Understood. Thank you. If I may ask one follow-up question about WorkBuddy: I’d like to hear your perspective. After all, every AI lab inherently has an incentive to develop its own type of application. What are your thoughts on how the market will divide between first-party and third-party applications? How do you plan to position WorkBuddy to compete with these first-party apps? And in your view, is WorkBuddy intended to be an enterprise software product on par with Tencent Meeting and Docs, or is it meant to become a new platform-level product in the future AI market? Thank you.
Anonymous spokesperson:
I believe it is indeed a new platform—a highly flexible foundation for AI agent workflows. Its core mission is to address all productivity needs of office workers and individuals running their own businesses, such as sole proprietors. Beneath it will be a platform that helps users leverage the capabilities of different models to solve their agent-related challenges. Over time, numerous models will serve users through WorkBuddy. A wide variety of skills, developed by different developers over time, will be created specifically to tackle productivity issues. The platform itself will then integrate various tools and models to accomplish these tasks. As coordinators, we select the most appropriate models and skills to help users resolve their problems, ensuring tasks are completed flawlessly. At the same time, these solutions will be delivered in a highly cost-effective manner. Qwen will be one of the models offered by WorkBuddy. However, if Qwen can effectively and exceptionally solve many user problems, it will become one of the primary models within WorkBuddy—though not the only one.
Robin Zhu, analyst:
Thank you very much.
Wendy Huang, Director of Investor Relations:
Thank you, Robin. We’ll now take the next question from Kenneth Fong of UBS.
Kenneth Fong, Analyst:
Hello, good evening management, thank you for accepting my question. I have a question regarding Xiao Wei’s development. Could management share any initial feedback or challenges from the Xiao Wei pilot phase? From a business perspective, how should we assess its net monetization potential? Specifically, as agents streamline the transaction path, we are concerned this may simply shift existing transaction volume from users’ traditional self-executed mini-program transactions to agents, incurring higher computational costs without significantly increasing net gross merchandise volume. Additionally, the shortened user transaction journey with Xiao Wei could also reduce exposure to high-margin advertising inventory risk. Thank you very much.
Anonymous spokesperson:
Well, I believe all the risks you mentioned are irrelevant, because we believe that when AI makes the WeChat ecosystem smarter—enabling users to execute transactions, explore content, and manage their daily lives through extensive AI-driven assistance—the already rich and powerful WeChat ecosystem will become even more valuable to users. So, if you think about how QQ was a communication and social tool in the PC era, and then when we entered the mobile era, WeChat emerged and essentially amplified QQ’s value by more than tenfold, because it was empowered by and optimized for the mobile era. Similarly, when we look at AI, we believe the WeChat ecosystem has another massive opportunity—to be first empowered by AI and, over time, become an AI-first application and ecosystem. When this happens, users will enjoy incredible experiences. Right now, you have to type and navigate through clicks. In the future, if you simply give one instruction to Xiao Wei and it can execute transactions and carry out your commands on your behalf, that will be an unbelievable experience for users. It will also be an incredible empowerment for the entire ecosystem. Therefore, we believe that if we can deliver this experience, if we can control the cost of delivery, and if you look at the design of VLM—it was built for privacy, cost efficiency, and to ensure it can fulfill all user needs within the WeChat environment—then we can truly empower WeChat for the AI era at controllable costs. When this happens, the WeChat ecosystem will expand, translating into massive value based on today’s monetization mechanisms within WeChat. I believe this is the future we see, and with the rollout of prototypes, we’re becoming increasingly confident that this will happen.
Kenneth Fong, Analyst:
Thank you, Martin. I have a follow-up question regarding AI cloud. Given that domestic API token pricing favors rapid commoditization and the Chinese cloud market is structurally price-sensitive, how should we view Tencent AI Cloud’s current profit margin situation, for example, compared to PaaS products in the industry? As AI adoption scales up, how should we anticipate future changes in profit margins? Thank you.
Anonymous spokesperson:
Yes, it's true that token prices in China are low, but the cost to produce tokens in China is also extremely low—far below common perceptions or external estimates. As a result, the token business can maintain a positive gross margin under these low token prices due to the minimal costs. If you look at the gross margin of paying users (WorkBuddy) or our Model-as-a-Service offering, our current gross margin is already comparable to Tencent Cloud’s overall gross margin. Of course, WorkBuddy’s overall gross margin is lower because we subsidize a portion of free users to drive market share and growth. However, among paying users, we are already achieving a quite healthy gross margin. Regarding your broader concerns, while the Chinese cloud market is indeed highly competitive, the landscape has changed significantly over the past few months due to rising input costs—particularly memory costs. Consequently, we have been increasing our pricing to customers. In May, we implemented a comprehensive price increase across Tencent Cloud, and beyond these headline price hikes, we have also substantially reduced discounts. As a result, the overall pricing environment in China’s cloud market is no longer as challenging as it once was. Thank you.
Wendy Huang, Director of Investor Relations:
Thank you, Kenneth. We’ll now take the next question from Ronald Keung of Goldman Sachs.
Ronald Keung, Analyst:
Thank you to Pony, Martin, James, John, and Wendy. I have two questions. First, regarding the HunYuan model: after the progress made with HunYuan 3, it has demonstrated strong cost efficiency and excellent performance in agents. How will HunYuan 4 differentiate itself, given that the landscape in the 3-trillion-parameter class is likely to become increasingly crowded over the coming months? What specific category, niche market, or differentiation strategy is HunYuan 4 targeting? Second, regarding capital expenditure and our focus on AI initiatives. Looking at some of our peers, they have shifted their strategies toward large-scale cloud businesses. I’d like to hear when or at what stage we might place greater emphasis on the cloud business as a potential high-return-on-investment opportunity worthy of increased capital allocation. Additionally, could you share how Tencent Cloud compares and contrasts with our U.S. counterparts, some of whom have shifted focus from applications to cloud? Thank you.
Anonymous spokesperson:
If you look at Qwen3, right? Even by today’s standards, Qwen3 is a very small model, but it is actually used extremely widely. So I believe Qwen3 has several key characteristics: first, it can match or even outperform larger models; second, it focuses on real-world use cases rather than just benchmark scores. As a result, in practical applications, it has become significantly more useful than many models of similar or even larger sizes. We believe this is also the principle we will apply to Qwen4. Therefore, when Qwen4 is launched, it will be a larger model capable of outperforming even bigger models, and it will be even more useful than Qwen3. We believe this will truly take us to the next stage, enabling us to deliver better intelligence to many of our users. Keep in mind that we also have products co-designed with our models. So when Qwen4 arrives, the products powered by Qwen4 will become significantly more powerful and useful than they are today, delivering substantial improvements to the products they drive. I think this is the path: Qwen4 is just another step, right? Then we’ll move on to Qwen5. As we continue to advance, we will approach SOTA and at some point definitely reach it. Once we achieve SOTA, we will also have a range of models in different sizes, capable of addressing diverse user needs at varying model tiers and cost efficiencies. Meanwhile, we will have multiple models co-designed with different products. This will help us enrich product features, strengthen both models and products, and ensure fast execution speeds. So this is our vision for Qwen4 and the subsequent Qwen5. Regarding your second question about allocating capital expenditure across different use cases, including Tencent Cloud: the most direct and primary use of capital expenditure in the coming months is to train larger and better Qwen models, as Martin discussed. But an important secondary use case is providing inference services for the Qwen models behind WorkBuddy, as well as for Deep Seq and other models. The primary goal of this initiative is to drive adoption of applications we consider strategically important and to establish critical feedback loops for our models and the broader ecosystem. It also brings the beneficial side effect of generating upfront revenue. From an accounting perspective, most user spending on WorkBuddy comes in the form of subscription fees—similar to gaming and some of our other businesses—there is a significant time lag between when users pay in cash and when that cash is recognized as reported revenue. However, we are already seeing substantial growth in cash revenue today, which will translate into reported revenue growth for Tencent Cloud for the remainder of this year. By year-end and into next year, we will also have sufficient GPU/ASIC capacity to strengthen Tencent Cloud’s offerings of bare-metal GPU rentals or model-as-a-service. Among these opportunities—GPU rental, model-as-a-service, and producing tokens for WorkBuddy—we believe producing tokens for WorkBuddy offers the most enduring economic value, which is why we are prioritizing it today. Thank you.
Ronald Keung, Analyst:
Thank you, Martin and James.
Operator:
Thank you. We’ll now take the next question from Alicia Yap of Citigroup.
Alicia Yap, analyst:
Hello, good evening management. Thank you for accepting my questions. Congratulations on the solid performance. My first question is about WeChat Mini Programs. Could management elaborate on your comments regarding the agent-to-agent transaction loop? Does this concept align with the long-term vision of a fully autonomous agent ecosystem within WeChat? Management also emphasized your exploration of on-device inference for Mini Programs. What are the challenges and benefits of this approach? Is this on-device inference method another reason why a proprietary VLM model is better suited to empower Mini Programs compared to external models? A quick follow-up question regarding your marketing services revenue: the growth rate accelerated to 22% this quarter. Should we expect continued upgrades and automation of AI advertising technology to further support this growth momentum? Additionally, do you anticipate any further future benefits from deeper integration with the HunYuan 3D model? Thank you.
Anonymous spokesperson:
Yes. Regarding agent-to-agent transactions, I envision a future in which many users will delegate their instructions to micro-agents and agents, conducting transactions over time. In the past, the WeChat ecosystem involved users interacting with content and with mini-programs themselves. In the future, if users can send complex instructions to an agent, that agent could begin assisting them in executing transactions. Many mini-programs and many merchants will have their own agents, and over time, these agents can interact with users’ agents. In the long term, every user may have their own agent, capable of interacting with others to execute transactions. I believe this is essentially what the future will look like, and we are gradually building the architecture to make it happen. Regarding on-device inference, I think it will happen incrementally, and only in the long term will most inference occur on-device, correct? But at some point, it’s not hard to imagine certain inference actually taking place on-device, while other inference occurs in the cloud. Over time, as on-device computing becomes more powerful and models become more efficient, more inference will shift to users’ devices. I believe this will return to the norm in the computing industry. If you consider the computing industry and the smartphone industry, most computation—CPU—is already performed on-device, while the cloud handles only a small portion. But in this initial phase of AI infrastructure, most computation occurs in the cloud because it requires immense power, and the challenges of equipping devices with sufficient computing capacity at an affordable and energy-efficient cost have not yet been fully resolved. That’s why everything currently happens in the cloud. But there will come a moment when increasing GPU capabilities are integrated into everyone’s phones and computers; at that point, more and more inference will shift to devices, returning us to an era where software and models become even more critical, and the returns from running models and applications will rise because the capital expenditure for computers will be shared across the entire ecosystem, not just borne by model companies. I believe this will inevitably happen at some point, and we are preparing for it. Regarding your question about marketing services, our advertising revenue growth has fluctuated in the past and will continue to do so in the future. I won’t simply extrapolate linear trends. There are several reasons. One is the flip side of my earlier comment about in-app advertising games dragging down international gaming revenue growth (relative to what it otherwise might have been)—this quarter, those same in-app advertising games contributed approximately 2 percentage points to advertising revenue growth. These in-app advertising games are a new business for Tencent and, to some extent, a new product globally; therefore, we have less clarity about the growth trajectory of their contribution compared to our traditional marketing services revenue. Additionally, Chinese consumers and the advertising market remain volatile, with potential economic or consumption headwinds affecting advertising trends. That said, we have consistently outperformed the overall Chinese advertising market and are confident we will continue to maintain a significant lead. Given the upside potential from our AI-driven ad targeting, the strong growth rate of engagement—particularly in our key video inventory—and the fact that we are still in the early stages of evolving toward closed-loop advertising—which will drive higher ad pricing—our outlook remains positive. Thank you.
Operator:
Thank you. We’ll now take the next question from Alex Liu of Bank of America Merrill Lynch.
Alex Liu:
Thank you for accepting my question. I have only one question. We’ve noticed that Tencent has recently increased its share repurchases, with the buyback program beginning in May, alongside a significant acceleration in capital expenditures. We understand this still places us in the relatively early stages of the AI investment cycle. Given this, how should investors view Tencent’s capital allocation priorities over the next 12 to 24 months? Thank you.
Anonymous spokesperson:
I believe, or I know, that capital allocation will be dynamic and reflect the environment we observe. Therefore, if we determine that capital expenditures offer superior returns—such as increasing our computing resources and then leveraging those resources to build models, leasing the computing resources through WorkBuddy tokens, or offering them as model-as-a-service—then we will direct more cash toward capital expenditures and potentially reduce cash used for buybacks. However, this will be a dynamic situation.
Unknown participant:
Thank you, Alex.
Anonymous spokesperson:
Additionally, I still want to emphasize that when we look at the capital expenditures allocated to building AI-native businesses, it’s more like a total amount we plan to invest this year and next, and I think people shouldn’t assume it will be this high every year, because the model-building component is largely fixed cost—you need to secure sufficient computing resources upfront, but you won’t need to invest more each year. Regarding inference computing, yes, we need adequate resources to generate tokens and build out our compute business. But we will only continue investing if we see substantial returns. If not, then this amount is essentially what we’ll invest, and any additional capital expenditures will be tied directly to the returns generated by this business. To fund this initial compute investment, we shouldn’t measure it solely against our operating cash flow; instead, we should consider our cash on the balance sheet, our investment portfolio, operating cash flow, and prudent debt capacity. All of these will be used to fund this initial phase of compute investment.
Wendy Huang, Director of Investor Relations:
Thank you, Martin, for your additional insights on capital expenditures and returns. We’ll now move to the next question, from Alex Yao of JPMorgan.
Alex Yao, Analyst:
Thank you to management for accepting my question. My first question concerns the Hunyuan flagship strategy. Hunyuan 3 competes on cost efficiency rather than raw capability. If you successfully build a truly state-of-the-art model, it would need to be larger and more expensive to run—what specific business value would this create compared to what Hunyuan 3 currently cannot deliver? Whether it’s a more capable WeChat intelligent agent, stronger ad performance, or enterprise customers—how does this opportunity justify a significant increase in training expenditures over the next 12 months?
Management:
Alright, let’s be very clear about this. The WeChat model, strategy, and positioning are fundamentally different, correct? WeChat’s agents don’t truly require or rely on the latest advancements and SOTA capabilities of HunYuan. As we’ve emphasized multiple times, WeChat’s design is fundamentally user-privacy-centric, focused on delivering all necessary interactions and agent technologies within the WeChat ecosystem while prioritizing cost efficiency. That’s its positioning. Now, achieving SOTA will enable us to build a highly significant tokens business, and it will also empower WorkBuddy to perform more challenging and higher-value services and operations for users. One of our key focuses with WorkBuddy isn’t just that it’s enterprise software—we’re not merely replicating everything people already do today. Instead, we’re continuously seeking value-added use cases that allow us to deliver additional value and returns to users, and in some cases, even help them earn more money. If we can achieve this, we unlock numerous business models. I believe this is precisely what we can accomplish with SOTA models. Furthermore, once we reach SOTA, we can begin developing a range of other specialized models designed to execute specific tasks across different cost-efficiency curves—all still on the cutting edge. This will allow us to meet users’ diverse needs for intelligence, with varying costs at each level, while still generating profit because we control the models, the inference costs, and the compute. This is our vision for what future generations of models can achieve.
Alex Yao, Analyst:
Thank you, Martin. My follow-up question is about the economic impact of the AI product. The drag from the new AI product increased from approximately RMB 8.8 billion in the first quarter to approximately RMB 10.5 billion this quarter. Could you explain how you are managing these investments? Are you doing so based on spending limits,