D

腾讯控股(纪要):WorkBuddy 起量,资本开支已近峰值

For the earnings review, see ‘Tencent: A Rare Profit Reversal — How Long Will the AI ‘Tuition’ Last?’

Below is Dolphin Research's $腾讯控股(00700.HK) Q3 2026 earnings call Trans

I. Key takeaways from the print

1. Capex and cash flow

a. Capex: Operating capex paid was RMB 63.6bn (+257% YoY), accelerating AI infra investments; non-operating capex paid was RMB 1.5bn.

b. FCF: -RMB 69.1bn, mainly due to large prepayments for compute; excluding prepayments, +RMB 16.9bn.

c. Net debt: Net debt was RMB 13.7bn at end-Sep vs. net cash of RMB 58.2bn at end-Jun, largely due to the above prepayments. This was partly offset by RMB 8.4bn net inflow from net disposals in the investment portfolio.

2. Profit drag from new AI products

a. Operating profit: Non-IFRS OP was RMB 68.1bn (-6% YoY). Ex-AI products it was RMB 83.3bn (+8% YoY), implying a gap of ~RMB 15.2bn.

b. Margin: Non-IFRS OPM was 33% (-5pct YoY). Ex-AI products, OPM was 40%, roughly flat YoY.

c. Net profit: Non-IFRS attributable NP was RMB 66.5bn (-6% YoY).

3. GPM and opex

a. GPM: Group 56%, flat YoY. VAS 64% (+3pct, higher-mix of self-developed, higher-margin titles); Online ads 52% (-4pct, AI-related opex and short-drama content costs); Fintech & business services 50%, broadly flat.

b. R&D: RMB 32.4bn (+42% YoY), focused on the Hunyuan model, new WeChat initiatives, and AI capabilities across products.

c. S&M and G&A: Selling expense RMB 12.1bn (+6% YoY), mainly to promote AI-native products such as WorkBuddy; G&A (ex-R&D) RMB 11.7bn (+2% YoY), non-AI spend remained under control.

4. Investment income and taxes

a. Associates/JVs: Non-IFRS share of profits was RMB 8.8bn vs. RMB 10.3bn a year ago, mainly on lower profits at certain domestic internet associates.

b. Income tax: RMB 9.1bn (-7% YoY), driven by lower taxes at an overseas subsidiary; excluding that, taxes would be up YoY.

c. Investment portfolio: As of end-Sep, listed investments at fair value were RMB 505.4bn; unlisted investments at carrying value were RMB 397.2bn.

II. Earnings call details

2.1 Management highlights

1. Foundation models: Hunyuan Hy4 has reached domestic SOTA levels, and a low-cost, lightweight Hy4 flash was launched to broaden adoption.

a. At launch, Hy4 topped domestic models on the Agent Arena leaderboard and added vision and computer use capabilities.

b. Hy4 flash is priced at 1/20 of Hy4 and ranks No.1 in daily token consumption among models released this year on OpenRouter.

c. Multimodal R&D has been folded into the Hunyuan team; 3D generation leads, and video generation is under development.

2. WorkBuddy: Now the No.1 AI app domestically by token usage, with management raising its view on paid enterprise software TAM in China.

a. DAU exceeds 10mn, with over 100 quadrillion tokens consumed daily; 7-day annualized revenue surpassed US$500mn, more than 2x Tencent Meeting's after seven years of launch.

b. Since the enterprise edition launched in Jun, 60% of China's top 100 enterprises have become customers; FDE support is being offered with partners.

c. Unit economics: Will improve by lifting pay rates, increasing tokens per user, and lowering token costs; currently free for teachers and certain groups on a temporary basis.

3. MiniBuddy: The in-WeChat personal assistant focuses on mass AI adoption breadth, with DAU reaching 30mn.

a. Integrated with Moments, Official Accounts, and Channels; now integrating with Mini Programs, merchants, and WeChat Shops to enable transactions.

b. Even if it scales to all 1.4bn WeChat users, quarterly costs would be below the quarterly cost of Yuanbao at the start of the year.

c. Monetization: Positioned as a revenue 'multiplier' for WeChat, with near-term monetization from ads and transactions; longer term, agent-to-agent matchmaking services could add revenue.

4. Cloud: Cloud revenue growth accelerated to +35% YoY (vs. +21% in Q2, vs. low-teens % in 2025).

a. In Q3, WorkBuddy revenue was over 5x QoQ, MaaS revenue rose over 3x QoQ, and GPU rental revenue increased 30%+ QoQ.

5. Core biz.: Several evergreen games hit record DAU, and ads grew despite industry headwinds.

a. Games: Domestic +11%; Delta Force: Hawk Ops saw peak DAU above 58mn in Oct. Overseas -9% (cc -4%), impacted by Supercell's high base, among others.

b. Ads: Marketing services revenue was RMB 43.0bn (+19% YoY); ad impressions on Channels rose 50%+ YoY, with meaningful contribution from short-drama entry points.

c. Social & fintech: Social network revenue was RMB 32.0bn (-1% YoY); total time spent on Channels rose 20%+ YoY, with per-capita time still less than half of the industry leader; fintech grew low single digits.

2.2 Q&A

Q: Capex and prepayments have been high for two consecutive quarters. How far along are compute purchases?

A: Prepayment outflows will decline, cash capex is at peak and should ease after 2027. Prepayments were abnormally high in Q2–Q3 as we locked in desired long-term contracts; in some quarters, prepayments may reverse if certain contracts are canceled.

When combining capex and prepayments, use a cash basis. This quarter, cash capex was approx. RMB 65.0bn rather than RMB 81.0bn on an accrual basis to avoid double counting; for the rest of 2026 and 2027, it should fluctuate around peak on a quarterly basis.

Also, the 2026–27 spend is akin to building 'factories' that can be reused to train new models and produce tokens. Pricing in 2026–27 is elevated, while similar hardware should be meaningfully cheaper by 2028, and compute efficiency doubles year by year—so 2028 and beyond should look much better.

Q: With tighter U.S. chip restrictions, what is the impact on existing/future compute deals and on cloud acceleration?

A: Cloud revenue should not face material adverse impact, though next-gen model training may be delayed in the near term. Cloud workloads are mainly inference, especially with efficient models like Hy4 flash; inference efficiency has jumped over the last 12 months. A small number of domestic chips can now handle inference efficiently, generating far more tokens than expected a year ago, and inference drives cloud revenue.

Training will be affected: while Hy4 trains efficiently, the next version may take longer to release near term, with domestic supply chains catching up over time. The new restrictions set all domestic players back a step, regardless of when they started renting or building compute overseas; in the long run, everyone will rely on domestic chips, and the rules should further accelerate local supply chain development.

Q: How big can WorkBuddy's user base and revenue get, are users mostly heavy, and is it constrained by chips?

A: It is already China's largest AI app by token usage with diversified token supply. WorkBuddy connects to multiple models, sourcing tokens partly from Tencent and partly from other model providers, and is the largest customer for almost all open-weight models in China.

Successful AI-era products tend to grow 10x; WorkBuddy fits that pattern, with DAU still growing fast and token consumption growth accelerating. The model is freemium with generous tokens for paid tiers; heavy users keep adding tokens, and some users already spend up to RMB 40,000 per month on tokens, unlike traditional domestic SaaS that relied on fixed subscriptions.

Q: What is the strategy, scope, and resource allocation behind the recent organizational changes, and when will the financial impact show?

A: Still very early; aim to run legacy businesses with fewer people and redeploy talent to build new growth engines. AI can lift operating efficiency, but it requires major org redesign and workflow reengineering; AI also empowers small teams to pursue new ideas and businesses.

The model team, the WorkBuddy team, and the efficiency gains at clients validate this, so now is the time to scale the reorg. A case in point: the model team was rebuilt from zero and within a year pushed the model into the domestic top tier; the WorkBuddy team was little known six to seven months ago and is now a real business. No timetable provided for financial impact.

Q: For Hunyuan and WorkBuddy going overseas, where will AI workloads reside, and how flexible is deployment?

A: We will solve infra where there is market opportunity; Hy4 already sees a meaningful share of usage from overseas. Some overseas demand can be served by domestic inference compute, and we aim to promote Hy4 globally, potentially via inference partners. As software, WorkBuddy can connect to different models with diversified token supply, and does not necessarily require proprietary inference infra, creating sizable overseas opportunities.

For cloud, we can build non-GPU capacity overseas at competitive cost; for GPUs, we can no longer provide high-end GPUs and may partner locally in various regions. No split provided for onshore/offshore deployment.

Q: Will abundant low-cost flash models drag down WorkBuddy's blended token price, and how does that affect monetization if volume rises but price falls?

A: Flash models drive acquisition, stronger models drive monetization; they are complementary. Flash lets us offer highly cost-competitive tokens to free users; when free users need more capabilities, they move to subscriptions, and for more complex work they switch to stronger models and higher token spend.

Management said this aligns with WorkBuddy's ability to monetize across the demand curve, and flash models only help. Price and volume elasticities were not quantified.

Q: When will overseas gaming return to positive growth, and will GTA 6 delay the inflection?

A: The slowdown started before GTA 6 buzz; we expect to resume a normal cadence of premium title releases next year. Multiple internal and external factors drove the slowdown and are being addressed; several evergreen games remain early in their growth curves, and new titles could also become evergreen, but there are no premium buy-to-play blockbusters this year.

Global gaming innovation has slowed somewhat. AI is currently more useful for content production and pre-production of new games rather than full production and live ops, and sustained leadership in tech, business models, and gameplay should naturally drive overseas revenue growth.

Q: AI investments are weighing on profits. Any nearer-term cost cuts?

A: No specific near-term cost cuts were provided. The company reiterated the framework of AI-native vs. legacy core businesses, with AI-native going through invest, then revenue, then profit; over the past six months, there have been substantive breakthroughs in models and products.

Legacy businesses will leverage AI and the learnings from building AI-native offerings to improve efficiency, while enabling teams to create more revenue and growth opportunities. No timetable given.

Q: How is testing for MiniBuddy, what are users saying, what are milestones for full rollout, and how do you view the monetization timeline, scale, and path?

A: It targets the mass WeChat user base rather than a small paid cohort; we expect to gain clarity on monetization in 6–12 months. It now has 30+ touchpoints, with rising interactions per user; feedback cites ease of use, though some say it is weaker than Muse by design, with capabilities to be built as it connects deeper across Mini Programs and the ecosystem.

We use small models to control cost; even at full coverage, quarterly costs would be below Yuanbao's at the start of the year. Compute efficiency doubles annually, so capacity should be multiple times better in 12–18 months, eventually reaching Muse levels, and adding 10%–20% to WeChat's existing multi-hundred-billion-RMB annual revenue would be material and sufficient to make MiniBuddy profitable.

Q: Ad GPM narrowed notably. How fast can AI efficiency gains offset costs, and will short-drama content costs be a structural drag?

A: The new variable this quarter was the short-drama surge, which carries upfront content and traffic costs. The ad tech stack keeps upgrading, with costs often incurred early in the quarter and benefits appearing late in the quarter or next quarter; that was a factor again this quarter but not new.

The new element is the breakout of the Green Bubble short-drama entry point, which contributed meaningfully to ads but is still in the early ramp. We need to attract independent studios whose content is not on major competitors, and the costs to drive traffic and revenue are somewhat front-loaded; long term, short dramas should form a new closed-loop scenario in WeChat, similar to mini-games, with the biggest opportunity still in Shops closed-loop e-commerce.

Q: Are frontier models in interactive narrative and world generation a threat or opportunity for Tencent's games?

A: AI cannot yet take over the production and live ops of mainstream games. Ultra-casual single-player mobile games have seen AI clones created via reverse engineering; large model labs have recently acted to block such reverse engineering, and cloning has been curbed.

Building and operating evergreen, multiplayer, cross-platform, multimodal titles is vastly harder than ultra-casual singles, which is where Tencent's gaming revenue concentrates; AI can currently accelerate pre-production. The industry does not lack games; the core remains sustained distribution and retention.

AI will enable a new generation of AI-native game experiences that can stand alone or be integrated into evergreen platforms; AI creation tools will also push evergreen games toward UGC platforms, with Peacekeeper Elite already performing well, and management sees this as a major positive for Tencent.

Q: With a high ads base in 2026, is there risk of a slowdown in 2027?

A: We do not see base effects as a risk and see more upside than downside for 2027. The ad market has been weak for months in 2026, hardly a banner year; ads are cyclical, and when consumer confidence and spending recover, ad budgets will rebound. We aim to continue to outgrow the market.

Q: New AI products are clearly dragging OP. What milestones define ROI, and when will the drag narrow?

A: No timetable for narrowing the profit drag, and 'nobody knows' exactly when it will be profitable. Management first looks at whether investment creates value: after launching Hy4 and Hy4 flash, users consumed massive tokens, validating training spend.

Yuanbao underperformed expectations and shifted from user acquisition to capability building, with AI search launched, and Yuanbao-related spend has been cut substantially; WorkBuddy, launched later, achieved strong success. The path is usage first, then revenue, and eventually profit.

Q: Cloud revenue accelerated to +35%. With constrained compute, how sustainable and high quality is AI cloud demand, and what are the revenue and margin runways?

A: Chip import restrictions will not derail cloud revenue growth, with WorkBuddy and MaaS as key drivers. Both deploy efficient models that can run on domestic chips unaffected by the new rules, and Hy4 flash being 95% cheaper than Hy4 is an example.

In recent months, tokens generated per unit of fixed compute have surged, leading to much better utilization of existing capacity, while we also add compute from domestic supply chains; together these support cloud revenue growth. No margin outlook was provided.

Q: Ex-new AI, OP growth slowed. What are the main cost pressures, and how do you see the next few quarters?

A: Two main pressures: weak consumption plus short-drama ramp costs, and rising prices for memory and other components. Weak consumption directly slows fintech services and drags on marketing services, partially offset by short-drama acceleration, though the latter carries heavy ramp costs.

Component prices spiked late last year—DRAM was up ~400% YoY early this year—and core businesses are now absorbing the pass-through from higher memory and other component costs. We have tightened memory usage, prices are stabilizing, and we expect these two pressures to persist for several quarters and then fade as high-cost inventory is consumed and macro conditions stabilize.

Q: Why re-enter search and launch Yuanbao Search?

A: Yuanbao's search capability is strong enough to open as an external service and should contribute some revenue going forward. AI agent capabilities fundamentally change search, making it a key AI-era capability; Yuanbao has built strong search within its services and can be offered as an atomic capability to other internet and tech companies where there is external demand.

<End of text>

Risk Disclosure & Statements:Dolphin Research Disclaimer & General Disclosure

登录即免费解锁16,517字全文

该内容仅登录用户可见,登录长桥账户后即可阅读全文