Reporter |
Editor | Wen Shuqi
On June 5th, at the Tencent Cloud AI Industry Application Conference, Tencent officially unveiled its “Efficiency Agent Toolkit,” targeting three core productivity needs: individual, office, and enterprise. The suite rolls out tailored agent solutions across more than 20 vertical scenarios, clearly positioning Tencent to fight for the hottest market right now: enterprise AI agents.
This is Tencent’s flagship AI event of the year. Beyond a sit-down chat with Chief AI Scientist Yao Shunyu, Tang Daosheng, Senior Executive VP of Tencent Group and CEO of Cloud and Smart Industries, broke his usual media silence. He tackled the tough questions swirling around Tencent’s AI push—commercialization timelines, model strategy, compute bottlenecks, and that nagging concern everyone keeps raising: Is Tencent moving too slowly on AI?
Tang’s answers lay out Tencent’s core takeaways on where AI stands right now.
First, agents are still in the investment phase. Tencent isn’t pushing its teams to hit commercialization targets just yet.
When asked repeatedly about ROI and performance metrics, Tang was straightforward: tools like WorkBuddy are still in a strategic investment stage. The company hasn’t set hard commercial targets for the team.
That doesn’t mean there’s no path to monetization. Tang points out that enterprise workflows already have a clear willingness to pay. He compares WorkBuddy to Tencent Meeting from a few years back: it bridges both consumer and business audiences, and the next step is leveraging that C-to-B flywheel. Notably, Tencent Cloud just turned its first profitable full year, with growth largely driven by SaaS and PaaS services.
Second, Hunyuan’s rebuild is showing results, but Tencent is doubling down on an open model strategy.
Tang has high praise for Yao Shunyu, who joined about six months ago. According to Tang, Yao has championed co-designing models directly alongside product development, shifting Hunyuan’s focus away from chasing external benchmark scores and instead locking onto real user experience as the north star. The impact is already visible: roughly 80% of Yuanshao users are now running on Hunyuan 3 (Hy3), with retention rates jumping noticeably.
Still, Tencent isn’t putting all its eggs in one basket. Tang stresses that Tencent remains open to third-party models, handing the choice directly to users—much like last year’s seamless integration of DeepSeek into Yuanshao.
Third, compute capacity remains an unavoidable reality check.
Tang didn’t shy away from the compute crunch. With resources stretched thin, Tencent prioritizes internal workloads—Hunyuan training, WeChat, Tencent Meeting, and Yuanshao all guzzle GPU power. Renting out GPUs to external clients takes a backseat for now. He’s hopeful that domestic chips ramp up later this year, which would significantly ease pressure on inference workloads.
What about building custom chips? Tang gives a firm no. Designing chips won’t magically fix supply shortages. Tencent’s smarter bet is ecosystem collaboration, partnering closely with a wider range of chipmakers.
Fourth, Tencent is actively courting new AI opportunities.
Tang views the AI race as a marathon, not a sprint. Tencent’s diverse portfolio means it can’t lead every single segment, so it’s completely normal for some units to move faster than others. That said, he notes Tencent actually moved fastest on the recent “Lobster” wave, with WorkBuddy quickly becoming the crowd favorite. Looking back at Tencent’s successful products over the past 28 years, they all went through peaks and valleys. The real secret? Lock in on something you believe has value, and ride the cycle.
Below is the full edited transcript of our conversation:
Media:There’s talk in the marketabout Tencent’shorse-racing approach to AI product development.Do you buy into that? If Tencent’s past horse-racing strategy yielded a few breakout hits, does it still hold water in the AI era?
Tang Daosheng: Honestly, I wouldn’t call it horse racing. The scenarios for agent services are incredibly diverse, and user needs vary wildly. Many AI products out there have carved out their own distinct niches.Inside Tencent, multiple teams are hunting for fresh opportunities within specific workflows,all aiming to serve those varied needs effectively.
Media: Hunyuan’s price per token has dropped, and at the same time, DeepSeek and Xiaomi are slashing prices. Withdomestic chips expected to arrive in the second half of this year,can wepredict what the downward curve for large model pricing will look like?How will this impact Tencent’s AI product pricing, and is another round of cuts likely?
Tang Daosheng: It’s not really my place to comment on other companies’ pricing moves, but the broader industry trend is definitely pushing toward lowering token inference costs. That’s exactly what’s needed to accelerate AI adoption.
But here’s another trend worth watching: according to Scaling Laws, bigger parameters generally mean stronger capabilities. So many vendors are now slicing their offerings into different tiers. You’ve got lighter models for budget-conscious use cases where ROI and lower inference costs matter most, but you also have ultra-complex problems that demand heavier models, which naturally cost more. Pricing strategies will definitely diverge based on that.
Media:Tencent Cloud’s DeepSeek V4pricing dropped,but early storage and CPU costs went up, pushing overall cloud prices higher. Is that pricing logic still holding true?
Tang Daosheng: Upstream semiconductor costs have indeed risen across the board. We’re weighing the market environment against customer needs, balancing competitiveness with long-term business health.
Media:After Hunyuan 3 Preview launched, token usage doubled compared to the 2.0 period. When does the official version drop?
Tang Daosheng: Stay tuned. Please don’t put too much pressure on Shunyu—he’s already working nonstop to get it across the finish line.
Media:Tencent’s been laser-focused on high-quality growth lately, and the cloud division booked a full year of profits in 2025. Take WorkBuddy, which is seeing rapid adoption and huge user numbers—will Tencent run profit or ROI checks on AI Agent products, or is this purely a strategic push? Compute and operational costs are climbing fast, aren’t they feeling the heat?
Tang Daosheng: Tencent operates across multiple tracks, each with different products at different stages. For AI agents like WorkBuddy, we’re still in the investment window. No commercial targets have been set. That said, enterprise clients are genuinely excited about WorkBuddy, and corporate workflows already have a clear path to monetization.I see WorkBuddy echoing Tencent Meeting from a few years ago: it serves both consumers and businesses. We’ll keep leaning into that C2B capability to build a sustainable service loop.
Media:Tencent always talks about building “useful AI.”How does that land in WorkBuddy’s ToB workflows?, and how do you actually meet enterprises’ current AI demands?
Tang Daosheng: Great question. I’ve been wrestling with how to define “useful” while building AI products for the past few years. At the end of the day, we’re all users. If a tool clicks, solves your problem, and doesn’t break the bank, you know it. Subjectively, users vote with their keyboards. WorkBuddy’s spread hasn’t come from aggressive marketing; people tried it, gave glowing feedback, and kept using it.
But in execution, a truly AI-native product requires four things to align: product design, model architecture, evaluation frameworks, and data quality. Those four dimensions act as our compass. When we evaluate a product, especially on open-ended tasks, we need clear metrics that tie directly to product goals.It really comes down to evaluation: getting all stakeholders aligned on the data required to train a model that actually meets those standards.
Media:In your current strategy, which drives Tencent Cloud AI growth more right now: the iteration speed of base models, or the engineering rollout of Agents? Second,when it comes to allocating effort and resources between broad ToB办公 scenarios and verticals like industrial or finance, what gets built in-house versus handed off to commercial partners?
Tang Daosheng: If you’ve used WorkBuddy, you’ve probably noticed its auto-mode calls different models depending on the task. Today’s agents tackle tons of office headaches largely because model capabilities have leveled up. Plug a two- or three-year-old model into the same tool, and you won’t get anywhere near today’s results. So I still firmly believe model iteration is the lifeblood of agent development.The auto-mode switching I mentioned highlights why openness matters. Tencent has always taken a flexible approach to constructing agents and packaging AI solutions. We’re eager to partner with different model providers, just like we did last year integrating DeepSeek into Yuanshao. Today, both CodeBuddy and WorkBuddy operate on an open-model strategy. These general-purpose tools need to support wildly different enterprise environments, so we’re handing model selection directly to the user. Of course, Hunyuan keeps iterating, and we have our own capability roadmap. Clients often show strong interest in calling Hunyuan directly through our agents, but we’ll stick to a highly open playbook for growing our AI agent business.
Media:Where does Tencent AI Agent sit right now on the stack? Are we leaning more IaaS, or SaaS/PaaS subscriptions? What’s the core mechanism driving that platform?
Tang Daosheng: Infrastructure compute has always been tight. Given limited resources, we funnel everything inward—Hunyuan training, WeChat scaling, meeting infrastructure, even Yuanshao consumes heavy GPU cycles. So when we actually lease GPU power via the cloud to serve industries,even though we’ve got some standout case studies, we still can’t fully cover every client’s demand. Growth drivers in China’s cloud market differ from overseas, still heavily driven by basic compute and storage services.
But since the Lobster launch, plus WorkBuddy and CodeBuddy rolling out, our token calls have gone exponential. Of course, compute limits still bite. That’s why we strongly believe the massive opportunity sits at the PaaS and token layers. On the IaaS side, Martin mentioned it in the latest earnings call: we’re really banking on domestic compute coming online in H2 to better support our cloud operations, especially inference workloads.
Media:Some say Agents will eat traditional SaaS market share. Within Tencent’s own product matrix, are Agent workflows cannibalizing legacy cloud services?
Tang Daosheng: User habits have shifted constantly over the past 30 years,so our product teams are actively adapting. Take Tencent Docs: recent updates essentially converted years of document-processing power into reusable Skills, turning them into interfaces WorkBuddy can call. Office workers get access to decades of accumulated processing power,but now it runs on a brand-new agent interface—like dropping a request into WeChat, then letting your PC pull doc capabilities to wrap up the task. What I’m seeing is that many of Tencent’s legacy SaaS and enterprise software tools are actively riding the AI agent wave instead of fighting it.
Media:You just mentioned Tencent’s compute shortage. You’re training internally, shipping products like WorkBuddy, and the cloud arm also sells tokens and GPUs externally. How do you balance competing demands to satisfy both internal and customer needs?
Tang Daosheng:We’ve prioritized internal products over the past couple of years. And remember, those internal products ultimately serve external users too, so protecting internal compute priority makes sense for us. But as domestic silicon arrives in H2, we’ll be able to feed internal needs while simultaneously scaling external services. That’s exactly where we’re heading.
Media:AI is fundamentally changing how Tencent builds products. Tencent has always been a product-first company. You mentioned the entire dev pipeline is shifting earlier today. Can you walk us through that using QClaw or WorkBuddy as examples?
Tang Daosheng:AI-native development is way flatter now. We’ve moved far from the old multi-role, rigidly bounded waterfall delivery. Back in the day, a PM would write the PRD, UX designers mapped the journey, visual designers handled the look and feel, front-end and back-end devs built it, architects stepped in, then coding started, followed by QA, load testing, and so on. It was a step-by-step relay.
Today, since AI generates massive amounts of code, having a sharp idea matters more. Roles are blending. Look at WorkBuddy, or the QClaw and similar projects you mentioned: small squads run the show. These folks usually have solid technical chops and can write code, but their job isn’t just typing syntax anymore. They’re defining what they want to build, architecting for efficiency, designing scalable systems, and figuring out how to adapt quickly to different demands.
A lot of the hard-won operational know-how from veteran engineers used to live in those older pipelines. LLM-generated code sometimes misses that nuance, so seasoned engineers still spend serious time on edge cases and refinement. But it’s not surprising that soon, R&D will be driven entirely by “what outcome do I want.” That shift is what makes the trendy “one-person company” concept actually viable. If you’ve got vision and creativity, you can collaborate with AI to ship a full product. I genuinely expect future teams to shrink further, with lean feature squads of three to five tackling specific builds.
Media:Does this throw major challenges at internal product management or workflow governance?
Tang Daosheng: Absolutely, and we’re adapting in real time. That’s why pilot products usually run the full lifecycle using new methods. Like I mentioned earlier, quality gates, evaluation, and testing are shifting left. Instead of waiting until the end, you’re already mapping out how to test and assess whether generated code actually hits specs while you’re writing the requirement description. All those backend validation steps that used to happen last are now happening first.
Media:Tencent Cloud previously concluded that billing purely by token and API isn’t a sustainable long-term model. If you pivot from resource sales to charging by task outcomes and business value, what’s the roadmap? How do you reconcile today’s token-based revenue with long-term value-based pricing?
Tang Daosheng:I think both models will coexist—they’re just different billing lenses. Token API pricing doesn’t have to be below or above cost. Charging above cost is a perfectly healthy, sustainable model. I’m confident certain industries and workflows will absolutely shift to outcome-based pricing. But hitting a specific result on a given task involves too many moving parts. If a product vendor only contributes part of the solution, it’s tough to charge purely on end results without sparking disputes. You can’t dismiss partial value, but pure outcome billing is messy and controversial in complex handoffs.
Media:How can Tencent Cloud help AI application builders avoid the margin trap where higher user activity just burns through inference costs?
Tang Daosheng:Mobile internet had rock-bottom marginal service costs, so ad models, eyeball economy, or transaction fees could easily cover expenses and leave room for profit. AI-native services, with today’s heavy operational and inference costs, can’t realistically rely on pure ad-funded usage. Especially when user consumption spikes differently based on the question or task, you lose predictable returns. Guaranteeing profitability while advertisers cover unpredictable compute costs becomes nearly impossible.
So, given that token costs tightly correlate with task complexity, ToC monetization—whether subscription, input-token pricing, or granular splits for output tokens, caching states, and so on—involves such complex transaction math that the cost delta is massive. You can’t flexibly monetize that with a simple eyeball model.
Bottom line: if AI inference costs stay elevated, these tools will naturally find their footing in high-commercial-value scenarios where the math works out, or where the new productivity they unlock offsets costs you’d otherwise burn trying to do it manually. That’s where you build a viable new business model.
Media:First half of the year was all about token anxiety, but Tencent leaned hard into product-scenario落地. Today you launched the agent toolkit covering 20+ verticals. What drove that decision, and why prioritize productization over raw scale?
Tang Daosheng: Tencent has always obsessively focused on product experience, solving real user pain points, and delivering tangible value. Those goals require product vehicles to actually reach users,and everyone knows Tencent is a product company,it’s baked into our DNA. I don’t see that changing in the AI era. The only difference is that with varying operational and marginal costs, we need fresh business models to carry those services. I’m convinced that as long as you create enough quantifiable value, customers will absolutely pay for it.
Media:Everyone’s talking token economics now. Will Tencent or Tencent Cloud track specific commercial KPIs like total token calls or industry penetration rates?
Tang Daosheng: Token volume isn’t a commercial metric; it’s an engagement metric. Right now, commercialization isn’t our primary focus. We’re polishing the product, expanding our user base, and proving this thing actually creates value and boosts productivity. We absolutely have a monetization framework—it acts as a filter. Compute is finite, so we need smart mechanisms to identify which users truly depend on the product and recognize its value enough to fund their compute needs. Figuring out that calibration is a key piece of the puzzle as Agent products mature.
Media:Everyone’s eagerly watching how Yuanshao and Hunyuan Threeco-design. You mentioned alignment came down to trust. What’s the collaboration rhythm looking like, and what are Yuanshao’s core positioning, growth targets, and actual KPIs as a consumer app?
Tang Daosheng: The partnership is getting tighter, and recently both teams moved into the same building for easier syncs. Roughly 80% of Yuanshao users are now onHY3 (Hunyuan 3), and retention jumped noticeably. A lot of the diverse services inside Yuanshao are now powered byHY3.
Media:What’s the actual KPI for Yuanshao?
Tang Daosheng: Growth is obvious—we want to win more users and steadily lift retention. Personally, I use Yuanshao daily, and what I want most from the team is to keep sharpening the search service, tap into richer data sources,answer everyday questions more accurately,and even hook into live data streams so the service isn’t stuck relying solely on what the model knew at training cutoff.
Media: Yuanshao burned through a lot of budget on growth earlier. What’s the next-phase growth playbook? Any aggressive ad spend planned?
Tang Daosheng: We’ve got a whole family of AI agents, but my biggest personal bet remains on Yuanshao. We still believe chatbots address a universal need and represent a critical track. But with more specialized agents added to our toolkit, we’re allocating some budget to fresh formats like CodeBuddy and WorkBuddy. I love the百花齐放 vibe in AI. Tencent ships different products for different audiences, and Yuanshao is undeniably a pillar.
Media:TencentTokenHub is scaling fast. Does that mean MaaS (Model-as-a-Service) gets heavier weight and bigger backing in cloud ops? Any revenue forecasts or targets for MaaS this year?
Tang Daosheng: Heavily weighted. It’s an industry-wide consensus. We’re continuously provisioning more compute to power token services, and TokenHub isn’t just fueling internal apps—enterprise clients on the cloud are calling it at blistering speeds too.
Media: On MaaS, rivals are dropping ambitious revenue targets. How do you view competition with other cloud providers here, and what’s Tencent’s own target?
Tang Daosheng: Every company develops at its own pace and rhythm. Tencent prefers letting product metrics and data speak for themselves.
Media:Tencent’s playing across three lanes: Hunyuan at the model layer, then CodeBuddy, WorkBuddy, and QClaw for Agents. Which lane gets the heavier bet to drive real breakthrough traction?
Tang Daosheng: It’s risky to call a final winner a decade out. Maybe there won’t even be a single endpoint; it’ll just be shifting milestones. I’ve spent enough years in this space to know that hype at the start rarely matches reality ten years down the line. Unexpected product pockets often surprise everyone. So we allocate solid resources to iterate across different tracks, watching market signals closely. If a product gains serious traction, Tencent pivots fast and doubles down.
CodeBuddy is a perfect example. That team existed three years ago, but originally shipped a developer-only tool. As AI capabilities matured and internal coding-agent demand spiked, we rolled CodeBuddy deeper into the ecosystem. Early on, it was just for programmers, but this year, as models pushed harder, we saw massive untapped potential. CodeBuddy evolved into today’s WorkBuddy—the kernel and harness design stayed identical. Suddenly a product shape unlocked a huge opening, extending well beyond coders. Non-dev corporate staff realized WorkBuddy massively cuts their friction, something we couldn’t have predicted a year ago. WorkBuddy’s explosive adoption wasn’t plotted two or three years ago. Reacting swiftly to market shifts is the real superpower.
Media: WorkBuddy rode the Lobster wave to fame. Now the buzz is around Claude Code or Codex. Have you considered pushing WorkBuddy global, similar to Codex?
Tang Daosheng: I’m counting on it. Both WorkBuddy and CodeBuddy will start testing international waters this year, serving overseas clients. We’ll actively harvest global user feedback to refine the product.
Media:Prior to this, Tencent consolidated its enterprise collaboration suite around Enterprise WeChat—docs, meetings, and storage all funneled through One ID. WorkBuddy Enterprise just launched the Agent Suite. How does this integration philosophy differ from before, and what’s the next wave of consolidation looking like?
Media:In the mobile era, Chat Apps were sticky powerhouses. Linking Enterprise WeChat with WeChat extended our reach in the Chat App track, letting users grab info and handle tasks instantly.
In the AI era, we’ve noticed capability and entry points can decouple slightly. During the Lobster rollout, lots of users treated Chatbots—like the Bot inside WeChat—as their main portal. But the actual agent logic often runs elsewhere, on your PC or cloud Lightboxes. To unlock years of accumulated value from past modules, we had to convert features into callable skills. This year, Enterprise WeChat opened up historical data capabilities via APIs and Skills, letting other agents tap in. That trend is accelerating fast.
In the AI age, whether it’s Yuanshao or WorkBuddy, the “chat” interface isn’t about social connection anymore. It’s a completely new interaction paradigm. We’ve seen that in professional settings, if you’re not syncing with colleagues or running meetings, tools like CodeBuddy and WorkBuddy actually match how people want to work. When that need exists, we owe it to users to deliver a smoother, native-tailored experience, so we’re continuously expanding coverage.
Enterprise WeChat and WorkBuddy will absolutely coexist on office desktops and in workplace flows. Ent WeCom will keep focusing on internal human-to-human comms, human-to-service touchpoints, or direct OA hooks for approvals. In those legacy workflows, Ent WeCom remains vital. But we can already picture work modes leaning harder into human-AI collaboration, and that’s exactly where WorkBuddy aims to deliver a deeply natural, AI-native experience.
Media:You’ve hammered home the compute bottleneck several times. WorkBuddy pricing is creeping up. Will Tencent ever consider building its own compute chips? Many cloud and model vendors are pushing full-stack synergy for cost advantages. Where do you stand?
Tang Daosheng:Designing chips in-house doesn’t magically solve manufacturing capacity. Having worked closely with chipmakers and partners, I can tell you nobody currently has enough fabrication headroom to meet market demand. Those are two completely separate issues. Our current ecosystem strategy lets us collaborate broadly with chip vendors, making Tencent a showcase benchmark for compute deployment—which partners genuinely welcome.
Media:The government’s been pushing hard on the compute network as one of its “six major networks.” From Tencent Cloud’s vantage point, does unified compute operations and scheduling just got more critical? You’ve got deep tech reserves for heterogeneous scheduling. How vital is the compute grid?
Tang Daosheng: Compute scheduling becomes non-negotiable when resources are scarce—you squeeze every bit of efficiency out of what you’ve got. But hitting that ceiling requires heavy engineering. Different workflows demand compute at wildly different pipeline stages, alternating between raw compute and flash storage. It’s a brutally complex optimization puzzle. Even sitting in the trenches, you need top-tier architecture chops. The more heterogeneous the fleet, the harder absolute matching gets. Juggling multiple chip architectures and model types still leaves gaps.
Data locality throws another wrench in. Chips demand rapid data fetches. If your compute cluster sits in the north but your dataset lives in the south, latency kills performance and wastes cycles. It’s a massive systems challenge. We’re grinding at every level—national initiatives aim for massive data clusters, while Tencent’s internal product teams juggle compute hunger, inference bursts, labeling workloads, multimodal training, and AV processing. Everyone’s workflow is intricate, so maximizing resource yield remains a heavy lift.
Media:Have you cracked any particularly solid patterns yet?
Tang Daosheng: We’re just relentlessly squeezing more yield out of what we have, precisely because supply is tight. Optimization is our best immediate workaround.
Media:Yao Shunyu’s been at Tencent for about six months. Curious what triggered the hire, and what concrete shifts has he brought to Tencent AI?
Tang Daosheng:Picking Shunyu felt inevitable. He’s a heavyweight voice in the field, and pre-hire discussions made his expertise undeniable. His grasp of AI-native fundamentals genuinely breaks from how we operated before.
His arrival sparked massive upgrades for Yuanshao. He aggressively pushed model-and-product co-design. He’s openly noted shifting Hunyuan’s North Star away from chasing external benchmark scores (baseline tests) straight to real user experience. We had plenty of data, but quality lagged. So leading up to Hunyuan 3, his big push was data purification—cutting volumes that looked useful for scale but actually muddy training, flagging harmful noise, and retiring it.
Spot-on intuition about AIGC model trajectories sharpens every decision. Without valuing data quality, you just chase petabyte-scale token dumps and never make the hard calls to prune datasets. Chasing hyper-complex architectures packed with tricks (hacks) makes scaling (expansion) impossible; or if you force scale (expansion), you strip architectures down to basics, guarantee enough compute and parameters, and let pristine data prove the model’s latent power. He’s mastered simplifying complexity. Hunyuan 3 Preview might not look gigantic on paper, but compared to previous generations, it’s a massive leap, and his fingerprints are all over it. Yuanshao and Hunyuan’s progress over the past six months outpaces what took longer stretches before.
Media:Rivals recently floated C-end monetization plans for large models. What’s Tencent’s roadmap for commercializing consumer-facing models? How do you view competitors’ explicit B2C plays?
Tang Daosheng:Different paths exist across the industry. Right now, our focus is sharpening product experience, carving out Yuanshao’s distinct positioning, and serving more users. That’s where we’re anchored.
Media:Last September’s ecosystem summit, you talked about Tencent Cloud trimming fat and building muscle. This year, we immediately saw the cloud division post its first full-year profit. Reports credit SaaS and PaaS surges. So is MaaS or TokenHub the headline target for this year, and what revenue contribution expectations do you have for MaaS?
Tang Daosheng:Absolutely crucial. It’s a high-growth engine. After explosive expansion, it’s become the fuel for agent products. I’m fairly certain today’s MaaS token consumption ties directly into WorkBuddy deployments on the enterprise side. I’m really bullish because demand is insatiable, and as I noted, compute supply caps us temporarily. Token services are compute in disguise, so as silicon loosens its grip, I expect this to unlock a massive new growth vector across the entire cloud market.
Media:Compute’s brutally tight, domestic silicon included. You’ve also flagged Tencent’s capex jumping compared to past years, though still trailing some rivals in sheer scale. What hurdles does Tencent face chasing compute, and where do our advantages lie?
Tang Daosheng: Head-to-head comparisons get messy globally. Everyone’s scrambling for silicon. Between training heavier models and powering Hunyuan to sustain our products and agent token demands, we’ll keep aggressively procuring and refreshing compute.
Media:Tencent Cloud’s AI division is still in the investment phase. In this marathon, how do you prove to the market that AI spend yields measurable returns, and when does AI Agent investment finally offset costs?
Tang Daosheng: Hard to pin down exact dates. The AI business is firmly in the strategic investment window.
Media: Anthropic recently published on AI assisting next-gen AI R&D. Any thoughts?
Tang Daosheng: AI already powers a huge chunk of modern model R&D. Following that trajectory, the proportion AI handles for next-gen foundation models will keep climbing.
Media:Tencent emphasizes three pillars: scenario connectivity, engineering, and models. Focusing on scenario connectivity: logically it’s about linking B2B and B2C products with ecosystem integrations.But how does that theoretical stitching actually perform in the wild? And does connectivity take on a heavier role migrating from chatbots to Agents?
Tang Daosheng: For scenario connectivity, we zero in on concrete moments. Tencent Meeting is a prime example—people log in to collaborate. We ensure agents extract exactly the context they need, auto-generate minutes, and maybe surface action items. If your company uses another system for task tracking, we can route those items straight into personal calendars or trigger reminders. That’s stitching workflows together, and office scenarios are packed with these hooks.
I see it in daily life too. My wife browses sites for flights, drops queries into Yuanshao, and seamlessly jumps to a travel site to compare and book. The imagination space is enormous. We’ve got plenty of room to grow, so we’ll keep investing in wiring disparate scenes together via AI—the exact sweet spot where foundational models shine.
Media:Cloud scaled profit beautifully, but Q1 growth faced headwinds. You mentioned 80% of calls now route through HY3 (Hunyuan 3). For upcoming phases, balancing healthy margins with proprietary model usage—any deeper market sizing considerations?
Tang Daosheng: I won’t forecast revenue, but the team absolutely carries ambitious growth targets.
Media:People watch Tencent’s AI timeline closely wondering if we’re lagging.Where do you admit we’re slower, and where do you hold strong confidence? What dimensions or targets define a signature victory in the second half?
Tang Daosheng: Shunyu touched on this onstage today. The term “second half” gets thrown around loosely; it really feels more like a marathon, a much longer contest. It’s been over three years since ChatGPT dropped, and the landscape has morphed constantly. Tencent’s portfolio is wildly diversified, handling dozens of ventures. Guaranteeing industry-leading status across every slice isn’t realistic. Some units speeding ahead while others crawl is totally normal. Flip the coin: when the Lobster wave hit early this year, Tencent led domestic response times, and WorkBuddy quickly became the sector’s darling.
New openings emerge constantly. Even on the same track, multiple product shapes can thrive. Tencent’s teams actively scout every opportunity. Sure, some squads surge ahead temporarily; others hit resource walls or structural headwinds. Stretch the timeline, especially looking at Tencent’s wins across 28 years, and success rarely ran perfectly smooth. Peaks, dips, recoveries—you name it. Tencent’s operating philosophy is simple: once you validate something carries genuine value, we commit fully and ride the cycle through.
