Tencent Prepaid Over 50 Billion Yuan in Q2 to Secure Memory Chips; Harness Paying-User Gross Margin Catches Up with MaaS
nashnova research
Tencent prepaid over RMB 50 billion in Q2 to secure memory chips in what management called a "once-in-five-years" spend; meanwhile, Harness paid-user inference margins have caught up with MaaS — signaling a monetization shift from selling compute to selling products.
RMB 50 billion prepayment — what did Tencent buy, and why now?
Chief Strategy Officer James Mitchell disclosed at an HSBC investor roadshow on September 3, 2026 that Tencent prepaid over RMB 50 billion in Q2 to lock in current- and next-generation memory chips — the storage components AI servers consume in bulk.
Management characterized this as a "once-in-five-years" partial one-off spend to resolve a memory-chip supply bottleneck. This means → Tencent sees the current supply window as unusually tight and chose to lock in early rather than wait.
In plain terms = this is not a recurring quarterly expense — it is a strategic stockpile purchase made while supply is constrained, similar to bulk commodity pre-buying.
Capex nearly doubled — will the burn rate continue?
HSBC estimates Tencent's full-year 2026 capex at roughly RMB 212.4 billion, nearly double the RMB 112.7 billion spent in 2025.
Management said Q2 spending can be treated as a new normalized baseline, though it may fluctuate. This means → the major step-up in capital intensity is already behind; RMB 50-billion-class prepayments will not recur every quarter.
Prepayments may extend into Q3 but are expected to normalize from Q4 onward.
Where is the AI spending going now?
The main AI cost driver shifted from Yuanbao marketing spend in Q1 to operating expenses for the Hunyuan large model and Harness products (WorkBuddy, CodeBuddy, etc.) in Q2.
This reflects a phase change in Tencent's AI investment — from customer acquisition to model training plus product operations, i.e. from burning marketing dollars to burning R&D and ops dollars.
Xiaowei uses the lightweight WeLM model, which requires less compute; its running cost will be significantly lower than Hunyuan or WorkBuddy.
MaaS is more profitable today — why isn't Tencent prioritizing it?
Management acknowledged that MaaS — model-as-a-service, selling large-model capabilities to other enterprises via the cloud — currently delivers the highest returns, with gross margins of roughly 40%, driven by GPU scarcity and heavy training demand.
Yet Tencent chose to allocate long-term resources to Harness products and Hunyuan model training over the more immediately profitable MaaS. This means → management is betting on "selling products" rather than "selling compute" as the durable path.
The logic: as demand shifts from training to inference and model developers build their own compute, MaaS margins face downward pressure; Harness benefits from rising inference demand, stronger user stickiness, and ongoing free-to-paid conversion — giving both revenue growth and margins room to expand.
How strong is Harness paid-user monetization?
Per management comments cited in HSBC's report, Harness paid-user inference margins already match MaaS levels.
Paid users consume tokens — the volume and depth of model calls — at a rate more than twice that of free users. In plain terms = users who pay not only pay, they also use the product far more intensively, making each paid user worth multiples of a free one.
This reflects a user-tiering monetization model that is working: the free tier drives volume; the paid tier drives profit.
How is Tencent fixing CodeBuddy's weakness? What does Hunyuan 4 solve?
CodeBuddy — Tencent's AI coding assistant — previously lacked a strong in-house code model, which constrained the product's capability.
The Hunyuan 4 preview delivered a major improvement in coding ability and can now handle a portion of CodeBuddy's code requests, filling that gap. This means → Tencent no longer needs to rely entirely on third-party models to power a core product.
A team restructuring had delayed model development by six to nine months, but a new unified reporting structure has accelerated the release cadence to once every two months. HSBC maintained its Buy rating with a target price of HK$655 unchanged.
市场有风险，内容仅供研究参考，不构成投资建议。