Tencent Holdings (Trans): Capex up 190%, AI drag RMB 10.5bn

Dolphin Research compiled the Trans of $ TENCENT.HK FY26Q2 earnings call.

I. Core takeaways

1. Shareholder returns and capital allocation

a. Paid the 2025 dividend of RMB 41.6bn this quarter; stepped up buybacks since May. Management will dynamically reallocate capital — if returns look better from scaling compute and deploying it to in-house models/WorkBuddy tokens/MaaS, more cash will shift to capex with buybacks reduced accordingly.b. AI-native capex is a lump-sum this year and next, not a recurring annual spend. Model training is closer to a fixed cost, while inference capacity will be added only when returns are attractive.c. Funding this lump-sum should be viewed against operating cash flow plus cash on hand, the investment portfolio, and prudent leverage capacity, not op. cash flow alone.

2. Group totals and profit

a. Revenue RMB 204.8bn, +11% YoY; GP RMB 118.4bn, +13% YoY; IFRS OP RMB 67.3bn, +12% YoY.b. Non-IFRS OP RMB 75.6bn, +9% YoY; non-IFRS NP (to shareholders) RMB 68.4bn, +9% YoY; diluted EPS RMB 7.433, +9% YoY.c. Financials: interest income RMB 4.2bn (+2% YoY); finance costs RMB 3.0bn vs. RMB 3.9bn, reflecting favorable FX and lower avg. rates; income tax expense RMB 11.7bn (+3% YoY).

d. Associates/JVs: IFRS loss share of RMB 10.0bn, mainly due to FV adjustments on convertible redeemable prefs at a non-listed investee whose own valuation rose; this is excluded under non-IFRS. Non-IFRS profit share RMB 6.4bn vs. RMB 6.3bn YoY.

3. Segment revenue and GPM

a. Mix: Social networks 16%, domestic games 23%, Intl games 9%, online ads 21%, fintech & biz. services 30%.b. VAS revenue RMB 98.0bn (+8% YoY), with social networks RMB 32.0bn (+1% YoY). In-app game item sales grew, partly offset by a 6% YoY decline in long-form video subs; music subs rose 8% YoY.c. Domestic games +17% YoY; Intl games -1% YoY, +4% at constant FX.

d. Online ads RMB 44.0bn (+22% YoY); fintech & biz. services RMB 60.0bn (+9% YoY).e. GPM: total 58% (+100bps YoY); VAS 64% (+400bps YoY) on higher mix of self-developed high-margin titles; ads 57% (-30bps YoY) as depreciation and opex from AI infra expansion were largely offset by revenue growth; fintech & biz. services 52%, broadly stable. GPs grew YoY by VAS +14%, ads +21%, fintech & biz. services +9%.

4. AI new products: investment and profit view

a. Ex-AI new products, non-IFRS OP was RMB 86.1bn (+19% YoY). AI new products dragged by about RMB 10.5bn vs. ~RMB 8.8bn in Q1.b. Non-IFRS OPM was 36.9% (-60bps YoY); ex-AI new products 42.0% (+280bps YoY).c. Opex: S&M RMB 11.9bn (+26% YoY) supporting games and AI-native product launches; R&D RMB 27.2bn (+35% YoY) on Hunyuan upgrades, WeChat AI, and AI features across products; G&A ex-R&D RMB 11.5bn (-1% YoY). Headcount ~116k, +4% YoY and +1% QoQ, mainly in games and tech platforms (incl. AI).

5. Capex and cash flow

a. Op. capex RMB 51.8bn (+190% YoY, +66% QoQ); non-op. capex RMB 1.0bn.b. FCF was -RMB 13.8bn, reflecting heavy AI infra capex, AI-related prepayments, and seasonal softness in game receipts. Ex-advance payments for compute, FCF was +RMB 37.6bn.c. Net cash RMB 58.2bn vs. RMB 146.9bn at Mar-end, after capex RMB 59.3bn and the 2025 dividend of RMB 41.6bn during the quarter.

II. Earnings call details

2.1 Management highlights

1. AI strategy and downside protection

a. Moats in existing businesses come from network effects, deep and value-added supply chains, IP, low take-rates, regulatory barriers, and proprietary data. AI is being leveraged across WeChat, games, and ads to lift returns and fund new AI businesses.b. Another reason to invest big in AI is clear downside protection: most spend is on infra, which in the worst case (which management sees as unlikely) can be rented out via Tencent Cloud at cost or better.c. The company views itself as two parts — existing franchises on a high-quality growth track with operating leverage, and new AI-native businesses (in-house models, new apps, and related compute infra). The latter is disclosed as a separate operating line.

2. Hunyuan foundation model

a. Hunyuan 3 GA shows clear gains vs. the preview: closing the loop with product teams improved training data quality and diversity, and scaled up RL, lifting task completion, while cutting hallucinations and error rates.b. Biggest gains are in agentic abilities and UX; advances in reasoning, agentic workflows, and long-context tasks deliver advantages in coding, office productivity, financial modeling, and front-end design.c. Adoption is accelerating: daily tokens across channels are ~6x the preview. By token usage, Hunyuan 3 consistently ranks top three globally on OpenRouter.

d. Product embedding: in WorkBuddy, Hunyuan powers complex agent workflows with higher success and shorter completion time; in Yuanbao, it leads in info retrieval, data processing, document flows, and everyday decisioning; in games, it supports AI teammate generation and code review (incl. the flagship 'Game for Peace'). It also powers WeChat Official Accounts AI assistants and Mini Program dev tools.e. Roadmap: a rapid iteration system validated by Hunyuan 3 is in place; stronger RL post-pretraining and multimodal upgrades are underway. The larger Hunyuan 4 is expected within the year, followed by Hunyuan 5, with ambitions to reach SOTA capability.f. Building in-house models enables co-design across apps–model–compute for better unit economics and faster innovation, and captures the value of 'intelligence' earlier in the AI diffusion curve.

3. AI apps: WorkBuddy and CodeBuddy

a. Both saw capability and user growth breakouts; by monthly interactions, WorkBuddy is the No.1 AI productivity service in China.b. Positioned as a one-stop workspace, WorkBuddy orchestrates multiple agents end-to-end for complex jobs. Users can control it via WeChat, WeCom, and PC, tapping 70k+ skills in Tencent Cloud's skill center.c. Rapid adoption comes with strong retention and willingness to pay given direct productivity lift. Embedding paid skills and a call-revenue split into task flows monetizes developer skills, energizing the dev ecosystem.

d. Focus now is market education and extending the lead; long-term unit economics should be compelling. Enhanced membership accelerates paid user growth, while agent, inference, and model efficiency drive token cost down.e. With WeChat, WeCom, and Tencent Meeting already ubiquitous in enterprises, WorkBuddy opens a new path to monetize enterprise relationships.

4. WeChat Xiaowei and Yuanbao

a. Xiaowei's prototype recently launched, offering embedded, context-aware agentic experiences inside WeChat, with access to social graphs, knowledge graphs, merchant resources, and payments.b. Powered by a WeChat-custom VLM focused on privacy, WeChat-specific scenarios, and cost efficiency. It helps users navigate WeChat's content universe and extract insights, discover products via Mini Programs, make decisions, and transact, laying the groundwork for agent-to-agent commerce.c. For safety, despite technical ability to handle advanced agentic workflows, the prototype requires user intervention and multi-step confirmations for now. It will open to broader users in phases.

d. Pre-wider rollout priorities include upgrading dialog, memory, and recommendations; expanding service and content access; scaling AI infra; and upgrading harness to support far larger user bases.e. Yuanbao focuses on search, ASR, and TTS upgrades, and covering a wider long-tail of AI needs (incl. multimodal generation). Yuanbao's dialog use cases feed valuable signals to Hunyuan models; hardened features are distilled into atomic capabilities for reuse by WorkBuddy, CodeBuddy, WeChat, and QQ Browser.

5. Social networks and digital content

a. Combined MAU of Weixin and WeChat rose YoY and QoQ to 1.40bn.b. Long-form video subs revenue fell 6% YoY, though the exclusive drama 'The Lead' was the top show across national platforms this quarter.c. Audio subs revenue rose 8% YoY, driven by higher music ARPU and richer content after Ximalaya consolidation. Users can now jump from Channels to QQ Music with one tap to discover new songs.

d. The Ximalaya acquisition closed in May, enhancing Tencent Music's resilience, deepening content ties with Yuewen and Ximalaya, and adding audiobooks and podcasts.

6. Games

a. Domestic momentum came from 'Delta Force', 'VALORANT' (PC/mobile), and 'Roco Kingdom: World'. 'Delta Force' DAU hit a record on the back of the Burst Fest, the first pro esports finals, and global 20v20 events; AI has been embedded in multiple production workflows, including performance analysis via data agents and asset generation via Hunyuan 3D.b. 'VALORANT' PC also hit record avg. DAU, helped by Skirmish Ascension with progressive weapons and the Summit map with breakable walls. Reach expanded via KOL collaborations, offline city events, and promotions across 10k+ internet cafes.c. 'Roco Kingdom: World' ranked No.5 by avg. DAU and No.8 by grossing among new mobile titles in Q2, and is the best-performing new launch YTD (management elsewhere noted it ranked No.1 by both avg. DAU and grossing among new domestic titles). Frequent content updates added 100 new sprites and seven new map regions.

d. 'Runaway Evolution' launched on Jul 9, adapted from the PC survival open-world builder 'Rust', which has ranked top-20 by concurrent users on Steam for eight years with its high-risk/high-reward weekly survival loops. The new title launched on both mobile and PC, adding sandbox safe zones to fit China user preferences.e. Intl games: 'League of Legends' DAU grew YoY, driven by ARAM Mayhem and the nostalgia mode League Classic (classic champions, runes, and the original Summoner's Rift). 'Warframe' DAU grew and revenue hit a record, helped by a wolf-themed Prime warframe and new storyline Jade Shadow's Constellations.

f. Lessmore (Miniclip) launched 'Arrow Puzzle Escape', the No.1 downloaded mobile game globally in Q2; its monetization is via in-app ads, so revenue is booked under ads, not Intl games. Adjusting for Arrow and similar ad-driven games, Intl games revenue growth would be 4ppts higher than reported.g. The -1% YoY in Intl games reflects growth in 'Wuthering Waves' and 'VALORANT' PC offset by declines at two Supercell titles.

7. WeChat ecosystem and online ads

a. Channels total watch time rose over 20% YoY, supported by richer content, better interactivity, and a new multi-variable ranking system. Partnerships with game studios, labels, and TV/film IPs broadened youth appeal and expanded the pool of creators earning directly within Channels.b. WeChat Shops GMV grew, with a notable lift from the centralized e-comm entry page. Merchants got tools like lotteries, while consumers received stronger repeat-purchase rewards to extend LTV.c. Ads growth was driven by both eCPM and impressions, with increased budgets from e-comm, internet services, and local services.

d. AI Marketing+ execution was upgraded end-to-end to better serve Shops and short-drama closed-loop advertisers. Merchants can auto-select SKUs, generate creative linked to products, and acquire traffic via smart bidding.e. Model size in ad recommendations expanded materially to capture user interests at finer granularity, lifting conversion. Channels ad impressions grew rapidly YoY on higher plays and ad load, though ad load is still well below short-video peers; Mini Programs drew rising spend from short-drama and mini-game studios.

8. Fintech and biz. services

a. Fintech revenue growth was driven by commercial payments, wealth, and consumer lending. Commercial payment volumes rose YoY, with a narrower YoY decline in per-transaction value; client AUM in wealth rose on the popularity of automated strategies and thematic index funds.b. Biz. services remain capacity-constrained, but cloud revenue growth accelerated from high-teens in Q1 to low-20s in Q2, supported by AI demand, Intl expansion, and higher usage and pricing in general cloud.c. AI demand translates into multiple revenue lines: GPU rentals, MaaS, and tokens consumed by WorkBuddy/CodeBuddy.

d. Intl cloud is expanding fast; CodeBuddy-built skills speed up client cloud migration vs. before, e.g., for a leading Indonesian telco.

2.2 Q&A

Q: Capex was RMB 53.0bn this quarter, up sharply QoQ and annualizing above RMB 200bn. How should we think about depreciation, coverage by AI-driven incremental revenue, and profit impact over coming quarters? Including R&D, what is the payback lag?

A: Given the surge in compute demand and rental prices, we could act like many neocloud players, rent out compute, recover depreciation almost immediately, and earn solid returns quickly. But we are pursuing a different game and a bigger strategy — allocating a large share of new compute to two goals: pushing our in-house models to SOTA and deploying/landing AI apps to lead the China market.Our view is that stronger intelligence from SOTA models and leading apps will translate into better long-term economics, e.g., via WorkBuddy token sales. Think of Tencent as two businesses: the existing franchise with steady growth and leverage, and a new AI-native business comprising in-house models, new apps, and the related compute infra, disclosed separately; revenue and profit should be mapped accordingly.Capex also splits: one part for existing franchises, still cash-generative and viewed as op. cash flow minus capex equals FCF; the other part for AI-native, essentially a lump-sum upfront to secure training compute, prepare for inference demand, and build AI compute and cloud.

We are investing now because without it, the business cannot scale. The upside is clear — better models and apps, while demand for compute is strong enough that renting via Tencent Cloud would yield significantly higher revenue and attractive capex returns; some prepayments and orders from months ago could be resold today at 30%+ gross margin.We believe allocating in the order of in-house models → own apps → compute rental will build a large, highly profitable, cash-generative AI-native business over time.

Q: Many AI labs will build their own harness apps. How will first-party vs. third-party harness share the market? How will WorkBuddy compete? Is WorkBuddy a peer to Tencent Meeting/Docs, or a new platform business?

A: It is a new platform — a highly flexible foundation for agentic AI. The core goal is to solve end-to-end productivity needs for office workers and entrepreneurs, including one-person shops.There is a harness layer to call different models for agentic tasks. Over time many models will serve users via WorkBuddy, and developers will keep publishing skills; we orchestrate models and skills to solve user problems perfectly and economically.Hunyuan will be one of the models in WorkBuddy; if it solves a large share of problems effectively, it can be a primary model there, but not the only one.

Q: Early feedback and challenges from Xiaowei tests? How do you assess net monetization? Shorter agent-driven journeys might shift existing Mini Program GMV to higher-cost agent flows and reduce high-margin ad inventory.

A: Those risks do not hold. Making the WeChat ecosystem smarter to execute transactions, explore content, and manage daily life will make it far more useful for users.Analogous to QQ in the PC era and WeChat in mobile, AI can make the WeChat ecosystem AI-first over time, delivering great user experiences. The key is delivering that experience and controlling cost.The VLM is designed for privacy, cost efficiency, and meeting all in-WeChat needs; if achieved, WeChat can enter the AI era at controlled cost. Ecosystem expansion will translate into value under existing monetization, and our confidence is rising post-prototype launch.

Q: With token prices commoditizing and China cloud structurally price sensitive, how do AI Cloud margins compare to US IaaS/PaaS? As AI adoption scales, how will margins evolve?

A: Token prices are low domestically, but token production costs are also very low, far below most external estimates. Even at these prices, token businesses can run positive gross margins.For WorkBuddy paid users or MaaS, current GPMs are comparable to Tencent Cloud overall. Total WorkBuddy GPM is lower given subsidized free users for share growth, but paid cohorts are attractive.On pricing, the market has shifted as costs, especially memory, rose; we raised prices across Tencent Cloud in May and cut discounts more sharply. The China pricing environment is no longer as tough as before.

Q: Hunyuan 3 differentiates on cost efficiency and agent abilities. With more 3T-parameter models arriving, where will Hunyuan 4 differentiate?

A: Hunyuan 3 is a small model by today's standards but broadly used, with two traits: it matches or beats much larger models and focuses on real use cases over benchmarks, making it more useful in practice.The same principle applies to Hunyuan 4 — it will be larger, beat even larger peers, and be far more usable. That moves us into the next phase of delivering better intelligence to many users.We also co-design products with the model; products powered by Hunyuan 4 will be stronger and easier to use. Hunyuan 4 is a waypoint toward Hunyuan 5 and SOTA, with a family of models at different cost-efficiency tiers addressing varied needs, and tighter app–model co-design to enrich functionality and speed.

Q: Given US peers' hyperscaler pivots, when will cloud become a capex-priority high-ROI business? How does Tencent Cloud's 'apps-to-cloud' path differ from US peers?

A: Primary capex use in coming months is training bigger, better Hunyuan models. A key secondary use is inference capacity for Hunyuan and for WorkBuddy's reliance on DeepSeek and other models.WorkBuddy aims to drive adoption of a strategically critical app with feedback loops to models and the broader ecosystem, while also generating revenue early. From an accounting view, much of WorkBuddy spend is subscription-based, creating a lag between cash collection and revenue recognition; cash is ramping now and will convert into Tencent Cloud revenue through year-end.By late year and into next, we will have ample GPU/ASIC capacity to scale bare-metal rentals or MaaS. Among GPU rentals, MaaS, and token production for WorkBuddy, we see token production as the most durable economic opportunity and prioritize it today.

Q: Please elaborate on 'agent-to-agent commerce'. Does this imply a fully autonomous in-WeChat agent ecosystem long term?

A: We envision users executing commands and transactions via Xiaowei and agents, shifting from direct user–content and user–Mini Program interactions. Over time, many Mini Programs and merchants will also have agents that interact with user agents.Longer term, each user may have an agent that collaborates with others to complete transactions. We are building the architecture to make this possible.

Q: What are the challenges and benefits of on-device inference for Xiaowei? Does on-device inference make an in-house VLM more suitable?

A: On-device inference will happen in stages, with most inference moving to devices only in the long run. It is likely some inference will be on-device and some in-cloud as device compute rises and models get more efficient.This returns to industry norms: in PCs and smartphones, most compute resides on devices with the cloud as a smaller share. Early AI infra requires strong compute and current devices lack sufficient, cheap, power-efficient on-device compute, so most runs in the cloud today.As more GPU-class capability reaches phones and PCs, more inference will move on-device, making software and models more important and improving returns for running models and apps as capex is shared across the ecosystem. We are preparing for this shift. (No direct comment on whether in-house VLM is better for on-device.)

Q: Ads revenue growth accelerated to 22%. Can continued stack upgrades and automation sustain momentum? What incremental gains come from deeper Hunyuan 3D integration?

A: Our ad growth has fluctuated and will likely continue to do so, so we caution against linear extrapolation. First, as the flip side of ad-driven games weighing on Intl games revenue growth, they added ~2ppts to ads growth this quarter; this is a new category for us and, to an extent, globally, so its trajectory is less predictable than traditional ads.Second, China's consumption and ad markets remain choppy, and macro headwinds can affect ads. That said, we have consistently outperformed the China ad market and expect to continue doing so, driven by AI targeting upside, fast user engagement growth in Channels, and early-stage progress toward more closed-loop ads with higher pricing.

Q: Buybacks increased since May while capex accelerated. With AI still early, how should investors think about capital allocation over the next 12–24 months?

A: Capital allocation is dynamic and reflects the opportunity set. If scaling compute and allocating it to in-house models, WorkBuddy token production, or MaaS offers better returns, we will shift more cash to capex and reduce buybacks accordingly.Importantly, AI-native capex resembles a lump-sum over this and next year, not a steady annual spend. Training capacity is closer to fixed cost and does not need constant add-ons; inference capacity must reach sufficient scale but will only be expanded with attractive returns, otherwise held at current levels and tied to business ROI.Funding this should be assessed versus cash on hand, the investment book, op. cash flow, and prudent leverage, not op. cash flow alone.

Q: Hunyuan 3 wins on cost efficiency. If you build a frontier model that is larger and costlier, what incremental commercial value can it deliver — stronger WeChat agents, better ads, or enterprise use cases? What would justify materially higher training spend over the next 12 months?

A: WeChat's model strategy is distinct — WeChat agents do not require Hunyuan at SOTA. WeChat's design centers on privacy, solving all necessary interactions and agentic tasks within WeChat, and cost efficiency.SOTA status enables a large token business and lets WorkBuddy handle more challenging, higher value-add services. We are seeking value-creating use cases that deliver additional returns to users, including helping them earn more in some scenarios, which unlocks many business models.Once at SOTA, we will spin out many task-specific models at different cost-efficiency tiers along the frontier to cover varied needs. Costs differ by tier, but we can maintain positive GM as we control models, inference costs, and compute.

Q: The AI loss drag rose from ~RMB 8.8bn in Q1 to ~RMB 10.5bn. How do you manage these investments — by envelope, return threshold, or strategic importance? What signals would trigger further investment or a shift to harvest?

A: It is very dynamic. We invest prudently until a breakout opportunity emerges, then we may step up — effectively a profit-linked envelope, with flexibility to accelerate when returns justify it.We view this as a long-run business where economics will show over time and turn profitable at some point. Notably, switching to renting out compute would be profitable today, providing a fallback and confidence to invest.We also reprioritize within the envelope. Versus Q1, the RMB 10.5bn in Q2 shifted materially toward WorkBuddy as we saw traction, while de-prioritizing some other AI products. Think of it as a capped budget with dynamic allocation.

Q: With model training and possibly cloud prioritized, where does Xiaowei inference sit post-launch?

A: Xiaowei's cost envelope is below what we invested in Yuanbao over the past year. Costs are controllable, and as experience improves, returns will accrue and soon exceed the spend.

Q: Timing and ROIC visibility for these AI initiatives? Should we expect near-term profit growth to be near zero, and when will reported non-IFRS OP growth (incl. AI) exceed the ex-AI metric?

A: We do not provide such precise guidance. We will keep discipline consistent with Tencent's approach, within a capped envelope, and invest when we see opportunities to build a large profitable business.We also retain the option to shift more compute to rentals for quick revenue, profit, and returns at any time. That fallback provides comfort as we invest.

<End of transcript>

Risk disclosure and statement:Dolphin Research Disclaimer and General Disclosure