Tencent spent over RMB 50 billion in the second quarter on prepayments to secure memory chips, a move management characterized as a "once-in-five-years" strategic procurement, indicating that capital expenditure levels seen in the second quarter may have become the new normal. Meanwhile, AI monetization is accelerating: the gross margin from inference for paying users of Harness has caught up with that of Model-as-a-Service (MaaS), and their token consumption is more than double that of free users, demonstrating strong tiered monetization capabilities.
$TENCENT (00700.HK)$ Large-scale bets on AI infrastructure are coming to light.
According to HSBC's latest research report, James Mitchell, Chief Strategy Officer of Tencent, disclosed during an investor roadshow conference call that the company made prepayments exceeding RMB 50 billion in the second quarter to secure existing and next-generation memory chips. He also highlighted the monetization potential of Harness products, noting that the gross margin from inference for paying users has reached parity with the Model-as-a-Service (MaaS) business, while token consumption by paying users is more than double that of free users.
These remarks were made during the Non-Deal Roadshow (NDR) conference call hosted by HSBC on September 3, 2026. In its subsequently released research note, HSBC maintained its "Buy" rating on Tencent with an unchanged target price of HKD 655, implying approximately 47.9% upside potential compared to the current share price of HKD 442.80.
HSBC pointed out that capital expenditures in the second quarter may have established a new normal operating level, although some fluctuation remains possible. Management indicated that the surge in capital intensity was completed in the second quarter, prepayments may continue into the third quarter, and normalization is expected to begin in the fourth quarter.
The scale of chip prepayments is rare, described by management as a "once-in-five-years" event.
According to HSBC's research report, Tencent's prepayments exceeding RMB 50 billion in the second quarter were primarily used to lock in current and next-generation memory chips at attractive prices. Management characterized this procurement as a partially one-off expenditure, occurring roughly every five years, aimed at addressing supply bottlenecks for memory chips.
Regarding the pace of capital expenditures, management stated that the spending level in the second quarter can be regarded as a new normal baseline, but emphasized that there is still room for upward or downward fluctuations. The company will continue to purchase chips for the remainder of this year and throughout 2027, but the primary jump in capital intensity was completed in the second quarter. HSBC estimates Tencent's full-year capital expenditures for 2026 to be approximately RMB 212.4 billion, a significant increase from RMB 112.7 billion in 2025.
Regarding expenses for new AI products, management noted that the main driver in the second quarter shifted from marketing costs for Yuanbao in the first quarter to operating expenses for the Hunyuan large model (HY) and Harness products such as WorkBuddy and CodeBuddy. In comparison, the operating costs for Xiaowei will be significantly lower than those for HY or WorkBuddy, as it utilizes the lightweight WeLM model, which requires lower computational costs.
The monetization potential of Harness is prominent, with gross margins for paying users catching up to MaaS levels.
In prioritizing AI monetization pathways, management acknowledged that MaaS currently offers the highest returns—driven by GPU shortages and large-scale training demands, with MaaS gross margins currently reaching approximately 40%. However, Tencent has chosen to prioritize long-term resource allocation toward Harness products and Hunyuan model training rather than focusing on MaaS, which offers more certain short-term gains.
Management’s rationale is that MaaS gross margins will face pressure in the future as market demand shifts from training to inference, and model vendors gradually build their own computing infrastructure, thereby reducing reliance on cloud service providers’ distribution channels. In contrast, Harness has upside potential for both revenue growth and gross margin, driven by increasing inference demand, improved user stickiness (through enhanced task history memory features), and the continued conversion of free users to paid subscribers.
Citing management comments, an HSBC research report noted that the inference gross margin for Harness’s paid users is now comparable to that of MaaS. Furthermore, paid users consume tokens at a rate more than twice that of free users, demonstrating strong monetization capabilities across user segments.
Behind the Pursuit of SOTA: CodeBuddy’s Shortcomings Accelerate Proprietary Model Development
Regarding AI strategy, management explained Tencent’s strategic motivation for pursuing state-of-the-art (SOTA) models. Management acknowledged that team restructuring had delayed model R&D progress by 6 to 9 months, but the new unified reporting structure has accelerated the model release cadence to once every two months.
At the product level, WorkBuddy can mitigate the risk of third-party model unavailability by directing users to Hunyuan 3.0 (HY3). However, CodeBuddy previously lacked support from a robust proprietary code model. The preview version of Hunyuan 4.0 (HY4 preview) has achieved significant improvements in coding capabilities, allowing it to handle some code requests from CodeBuddy and fill this gap. Management also pointed out that possessing proprietary SOTA models helps improve the profit margin structure in the long term.
Editor/lambor