After the market close on August 12, Tencent Holdings released its Q2 results for the period ended June 30, 2026: revenue reached RMB 204.78 billion, a year-on-year increase of 11%, returning to double-digit growth after several quarters; gross profit was RMB 118.43 billion, with the gross profit margin holding steady at 58%. However, single-quarter capital expenditure surged to RMB 52.78 billion, a massive year-on-year increase of 176%, driving free cash flow directly down to negative RMB 13.8 billion.
On one hand, gaming, advertising, and Channels continue to provide steady cash flow; on the other hand, money is pouring into AI infrastructure like a breached dam. However, from the author's perspective, these are not the focal points. The real highlight is that Tencent's AI aircraft carrier is about to set sail.
Looking at the Big Picture: The RMB 52.8 Billion Is Not a Routine Annual Expenditure
Tencent President Martin Lau made the stance very clear: "Capital expenditure directed towards AI-native businesses is mainly a one-time concentrated investment this year and next. Please do not assume that we will maintain the same level of massive investment every year. Model training is a fixed cost; we only need to reserve sufficient computing power. Inference computing power will only be continuously increased when it can generate substantial returns; otherwise, the current scale of investment will be capped."
James Mitchell, Chief Strategy Officer and Senior Executive Vice President of Tencent, also clarified a similar viewpoint: the business needs to be accounted for in two separate ledgers. The mature core businesses, including gaming, advertising, social networks, and payment, feature strong operating leverage and self-circulating cash flow. The AI-native businesses, including proprietary models plus WorkBuddy, XiaoWei, and Yuanbao, along with newly built computing power, represent one-time startup costs for this year and next, not perpetual cash burning.
Management revealed that fund allocation is divided into two major priorities: first, advancing the R&D of the large-parameter Hunyuan Hy4. Currently, the official version of Hunyuan Hy3 ranks among the top three globally in Token consumption on OpenRouter, with daily call volume reaching seven times that of the preview version, and Hy4 is expected to be released within the year; second, ensuring the supply of inference computing power for WorkBuddy, CodeBuddy, XiaoWei, and Yuanbao.
WorkBuddy: Leading the Domestic B2B Agent Arena, Leaving the Runner-Up Far Behind
If 2026 is the "first year of office agents," then the data from Analysys's "China Office Agent Platform Market Insight" report in Q2 is quite striking: WorkBuddy's PC visits reached 20.97 million in June, ranking first domestically, surpassing the sum of the second and third places; among the top ten in monthly visits, Tencent-affiliated products occupy four seats—WorkBuddy, CodeBuddy, QClaw, and Marvis.
Tencent-affiliated products hold a combined market share of 53.8% in the AI office agent desktop market.
The figure below shows the Baidu Index for the three major Agent platforms—WorkBuddy, QoderWork, and Trae—as of August 9, with WorkBuddy taking the clear lead.
The core highlight is not just the leading visit scale, but rather three key advancements: after integrating Hy3, the task resolution rate increased from 72% to 90%; in July, Token revenue reached RMB 119 million, nearly doubling month-on-month, with a market share of 35%, and the gross profit margin for paid users has caught up with the overall level of Tencent Cloud; the enterprise version has been implemented in over 50 industry scenarios, including Guangdong government affairs and Shenzhen Port, demonstrating genuine deployment capabilities.
Martin Lau provided a clear and high-level positioning for the product: WorkBuddy is not just an add-on plugin for Tencent Docs and Tencent Meeting, but a universal productivity workstation for the AGI era. The product's underlying architecture is built on a scheduling framework, with Hunyuan as one of the core foundations rather than the only one; DingTalk, Feishu, and Coze are all targets it aims to capture.
WeChat's "XiaoWei" vs. Doubao: Refighting the Battle Between Channels and Douyin for the AI Consumer-Facing Entry Point
"XiaoWei," WeChat's native AI assistant that underwent grayscale internal testing in June, was compared by Martin Lau during this earnings call to the mobile transition from "QQ to WeChat," stating that "XiaoWei is the second leap in WeChat's AI evolution."
To understand XiaoWei, one must first understand the differences between it and Doubao, which is practically a replica of the Channels vs. Douyin battle:
The script for Channels back then was: the recommendation algorithm could not beat Douyin, but "social recommendation + closed-loop within WeChat" forcibly tore open a path, creating a public-private domain mismatch with the high-frequency, immersive Douyin.
Notably, the per capita usage time of WeChat Channels in Q2 2026 still grew by 20% year-on-year.
XiaoWei is replicating a similar development path: at the model parameter level, it does not blindly benchmark against the performance ceiling of Doubao. WeLM adheres to the development philosophy of "sufficiency first, balancing privacy, and deep cultivation of WeChat native scenarios." But it holds 1.432 billion WeChat MAUs, 973 million Mini Programs MAUs, and the WeChat Pay ecosystem. This scenario barrier is a unique advantage that the standalone Doubao app can hardly reach.
Developers only need to check the authorization box in the AI capabilities section of the Mini Program management backend, and XiaoWei can invoke the corresponding services with a single sentence. The path logic is exactly the same as how Channels revitalized the Official Accounts and Mini Programs ecosystem back in the day.
Martin Lau stated frankly during the meeting that XiaoWei will not divert advertising and transaction GMV. Instead, it can connect the chain of "user command trigger—merchant agent response—Mini Program closed-loop transaction," driving WeChat to iterate from a mobile operating system to an AI operating system.
In the Middle Layer Between Models and Applications, Tencent Cloud Is Becoming the Cash Flow Foundation for the RMB 52.8 Billion Capex
The financial report shows that the daily Token call volume of the large model service platform TokenHub has exceeded 25 trillion, growing five times within two months. Behind this astonishing growth rate is the pull from internal businesses such as WorkBuddy, as well as the strong demand from external customers for GPU leasing and MaaS (Model as a Service) services.
Management explicitly stated that from the end of 2026 to 2027, as GPU/ASIC reserves are put in place, Tencent Cloud will increase its efforts in bare metal leasing and MaaS output. This will not only provide cash flow support for AI businesses but also offer strong defensiveness during industry demand fluctuations.
Final Thoughts
Pony Ma divided Tencent's AI layout into a three-tier architecture: the intelligence layer relies on Hunyuan Hy3 to consolidate the foundation and depends on Hy4 to push the boundaries of cutting-edge technology; the application layer sees WorkBuddy, CodeBuddy, and others seizing leading domestic positions, XiaoWei reshaping the WeChat traffic entry point, and Yuanbao holding its ground in the consumer-facing standalone app track; the infrastructure layer plans to invest RMB 52.8 billion in capital expenditure to drive the efficient conversion of model and application traffic into revenue.
Therefore, this Q2 performance is not a narrative of Tencent being dragged down by AI investments and facing profit pressure. In essence, the company is utilizing the phased capital expenditure over the two years of 2026–2027 to secure the three core entry tickets for the AI era—models, applications, and computing power—all at once. Currently, WorkBuddy has verified its visit scale and Token commercialization revenue. During its grayscale testing phase, XiaoWei is highly likely to replicate the growth path of Channels, and Tencent Cloud's MaaS business is solidifying the fundamental performance base.