Tencent prepaid over $7B to lock in memory chips for AI

Tencent spent more than 50 billion yuan ($7.4 billion) in prepayments during the second quarter to secure current and next-generation memory chips, a move the company's management described as a "once-in-five-years" strategic procurement, according to an HSBC ↘ 1.08% research note published after an investor roadshow on September 3.news.futunn+2

Chief Strategy Officer James Mitchell made the disclosure during an HSBC-hosted non-deal roadshow call, where he also revealed that paid users of Tencent's AI "Harness" products are consuming tokens at more than double the rate of free users — a signal the company views as early validation of its AI monetization strategy.gate+1

A Quarter Defined by AI Spending

The prepayment forms part of a broader capital expenditure surge that has reshaped Tencent's financial profile. In the June quarter, the company reported capital expenditure of 52.8 billion yuan, up 176% year-on-year, according to its official earnings release. Bloomberg reported the spending triggered a free-cash outflow of 13.8 billion yuan, while Reuters noted that net profit rose just 0.7% to 56 billion yuan, missing analyst estimates.reuters+2

Management indicated that Q2 spending levels may represent a "new normal" baseline for capital expenditure, though with room for fluctuation. HSBC estimates Tencent's full-year 2026 capex at roughly 212.4 billion yuan, nearly doubling from 112.7 billion yuan in 2025. The chip prepayments could extend into the third quarter before normalizing in the fourth quarter, according to the research note.wallstreetcn+2

AI Monetization Takes Shape

Beyond infrastructure, Tencent outlined a deliberate shift in AI spending priorities. The company acknowledged that its Model-as-a-Service business currently delivers the highest margins — around 40% gross margin driven by GPU scarcity and training demand — but said it is channeling long-term resources toward its Harness product suite and its proprietary Hunyuan model instead.gate+1

The rationale, according to management, is that MaaS margins face structural pressure as the market shifts from training to inference and model developers build their own computing capacity. By contrast, Harness paid users already match MaaS-level inference gross margins, and their higher token consumption suggests room for revenue growth as free users convert to paid tiers.wallstreetcn

Catching Up on Models

Mitchell also addressed Tencent's model development timeline, acknowledging that an internal team reorganization had delayed progress by six to nine months. A unified reporting structure has since accelerated the release cadence to roughly every two months. The preview of Hunyuan 4, the company's latest large model, has addressed a gap in coding capability that had been a weakness for its CodeBuddy product.wallstreetcn