Tencent’s RMB 10.5 Billion AI Loss Masks a Surprising Margin Edge
Tencent reported Q2 2026 results on August 12 showing non-IFRS operating profit of RMB75.6 billion, up 9% year-over-year. Exclude its new AI products — Hunyuan, Yuanbao, CodeBuddy, WorkBuddy, and Xiaowei — and that figure reaches RMB86.1 billion, a 19% increase. The RMB10.5 billion gap represents quarterly spending on model development, R&D, inference compute, and user acquisition, up from RMB8.8 billion in Q1.
Capital expenditure hit RMB52.8 billion, up 176%. Free cash flow turned negative at minus RMB13.8 billion, though management said it would have been positive RMB37.6 billion without compute prepayments. Net cash dropped from RMB146.9 billion in March to RMB58.2 billion at June-end.
These are historic spending rates. What complicates a simple cash-burn reading is what management disclosed about the economics underneath.
Cheap tokens, fat margins
The most consequential exchange on the earnings call came when strategy chief James Mitchell responded to UBS on China's low API pricing. Mitchell said Tencent's domestic token-production cost is "extremely low." Paid WorkBuddy subscribers and Model-as-a-Service offerings already produce gross margins comparable to Tencent Cloud overall — a segment running 52% gross margin in Q2. Cloud revenue accelerated to low-twenties percentage growth, up from high-teens in Q1.
Public TokenHub pricing for Hy3 adds external evidence: $0.132 per million input tokens, $0.528 output. Western list prices for comparable tiers run multiples higher — OpenAI's GPT-5.6 Luna at $0.20/$1.20, Google's Gemini 3.1 Flash-Lite at $0.25/$1.50, Anthropic's Claude Haiku 4.5 at $1/$5. Models and workloads differ, but the dollar gap is concrete.
The mechanism is specific: Hy3 is a compact model with strong cost-performance characteristics. Fewer floating-point operations per useful answer means lower inference cost, which means rock-bottom token prices can still clear positive gross margin. Management's strategy is to reach high model quality, then spawn task-specific variants along the cost-efficiency curve.
WorkBuddy prices like a commercial product
On May 15, Tencent raised its enterprise WorkBuddy/CodeBuddy subscription from RMB78 to RMB198 per user per month. The private-cloud tier doubled to RMB316. Tencent is pricing the application commercially; the cheap intelligence sits one layer below, inside its own inference stack.
Aggregate WorkBuddy gross margin remains lower because Tencent subsidizes free users for adoption. A third-party metric ranked WorkBuddy as China's most-used PC-based AI office agent, with healthy retention and willingness to pay — but conversion from free to paid will determine whether unit economics reach product-level margins.
The compute optionality that reprices the risk
This is the disclosure that should change how capital allocators frame Tencent's AI bet.
President Martin Lau told analysts that if Tencent abandoned its AI applications entirely and rented its compute to third parties, it could recover depreciation almost immediately and earn a decent return. Some capacity purchased months earlier could now be sold at prices exceeding purchase cost by more than 30%, because supply has tightened. Tencent Cloud itself raised AI-compute prices 5% in May, citing hardware-cost inflation.
The quarterly capex is funding an asset — raw compute — that retains and potentially appreciates in market value regardless of whether Tencent's AI applications succeed. If WorkBuddy gains traction, compute generates cheap tokens feeding high-margin applications distributed through WeCom and Tencent Meeting. If those applications disappoint, the same hardware redirects into GPU rental and MaaS, where third-party demand is absorbing capacity faster than Tencent can add it.
Lau described external compute rental as the explicit fallback — the reason management can tolerate RMB10.5 billion in quarterly AI losses without existential concern about stranded capital. The hardware is fungible across training, inference, MaaS, and bare-metal rental, and current market conditions favor the seller.
The strongest counterargument follows Jevons-style logic: Tencent's token costs may keep falling, but competitors can pass every efficiency gain to customers, compressing prices in lockstep. Hardware costs are simultaneously rising. And if competitive pressure blocks conversion of free users to paid subscribers, today's encouraging paid-cohort margins stay confined to a thin slice of the base.
Compute optionality changes how to price that risk. A company whose AI capex can be liquidated at a premium carries a different risk profile than one whose entire investment depends on a single application's adoption curve.
not investment advice