Tencent Says Capital Expenditure Will Remain Near Its Peak in 2027 as It Expands Computing Capacity Through Domestic Supply Chains
Tencent said quarterly cash capital expenditure may fluctuate near peak levels through the remainder of this year and into 2027. After completing its infrastructure investments for 2026 and 2027, financial trends are expected to improve in 2028. Reduced access to advanced computing capacity may delay the launch of its next model version, but domestically produced chips can already support efficient model inference, and Tencent continues to expand computing capacity through domestic supply chains. Tencent's capital expenditure rose 527% year over year to RMB 81.433 billion in the third quarter.
Why read this
SelectedAI score 77
Tencent disclosed details about peak capital expenditure, the inference efficiency of domestically produced chips, and constraints on model training, offering insight into how domestic computing capacity supply affects its AI business.
Topics
Body · AI translation
Tencent reported third-quarter revenue of 207.336 billion yuan, up 8% year on year, while non-IFRS operating profit fell 6% year on year to 68.073 billion yuan. Excluding the impact of AI products, non-IFRS operating profit for the quarter grew 8% year on year. Tencent’s capital expenditure rose 527% year on year to 81.433 billion yuan during the quarter, while cash payments for capital expenditure totaled 65.1 billion yuan.
Tencent Chief Strategy Officer James Mitchell said capital expenditure prepayments were unusually high in both the second and third quarters. However, Tencent has signed long-term contracts, so cash outflows for prepayments will decline, and some quarters may even see prepayment reversals. Cash capital expenditure—namely, prepayments—has also reached peak levels. Quarterly cash capital expenditure may fluctuate around these peak levels for the remainder of this year and throughout 2027, before beginning to decline as expected.
Tencent President Martin Lau added that the infrastructure investment now underway is primarily intended to launch AI-native businesses. Once the infrastructure is in place, it can continuously train new generations of models and generate tokens. Therefore, after these capital expenditure investments are completed in 2026 and 2027, Tencent will be able to use the infrastructure already built in 2028, significantly improving its financial trends.
Lau also noted that equipment prices are relatively high in 2026 and 2027. Even if similar equipment needs to be purchased in 2028, prices are expected to fall substantially. In addition, computing efficiency is multiplying year by year in the course of model training and inference, meaning the same infrastructure will deliver greater training efficiency and higher token output in the future.
Lau said this is an obstacle for all domestic players: whether they rent computing capacity overseas or build it themselves, less advanced computing capacity is available to them. Training will be affected, and Tencent’s next model version may take longer to launch. In the long run, however, the domestic supply chain is expected eventually to mature and catch up. For AI inference, the domestic supply chain is fully capable of providing support because the models themselves are highly efficient. Over the long term, he believes all Chinese players will ultimately rely on domestic supply chains and domestically produced chips.
Regarding the cloud business, Mitchell said Tencent has not yet seen advanced-chip supply constraints negatively affect it. Domestic chips can already run inference workloads efficiently for highly efficient models such as Hy4 flash. Inference efficiency has undergone a dramatic transformation over the past 12 months, with token throughput powered by domestic chips far exceeding what was imagined a year ago. Tencent continues to add computing capacity through its domestic supply chain.
During the question-and-answer session, Tencent’s management also addressed questions about its AI product plans. Lau said WorkBuddy’s annualized gross revenue exceeded US$500 million over the past seven days, more than twice that of Tencent Meeting, which launched seven years ago. WorkBuddy’s third-quarter revenue was more than five times that of the previous quarter.
Tencent’s goal for Xiaowei, WeChat’s AI assistant, is to make it available to all 1.4 billion WeChat users. Tencent’s management disclosed that Xiaowei has reached 30 million daily active users (DAU). The dedicated vision-language model used by Xiaowei has very low inference costs, and even if Xiaowei is made available to all 1.4 billion WeChat users, its quarterly costs are expected to remain below the quarterly costs incurred by Yuanbao earlier this year.
Tencent’s plan for Xiaowei is for it to “become a multiplier,” driving compounding growth within the WeChat ecosystem. WeChat’s overall annual revenue base currently amounts to several hundred billion yuan, and Tencent’s management believes that selectively adding 10% or 20% on top of that would represent a considerable sum. Xiaowei has the potential to become a highly profitable project. Lau said that given around 18 months, Xiaowei’s capabilities are expected to far surpass where they are today, and at some point it will become as powerful as Meta’s Muse.
AI Trends