By NUPIAO Newsroom
On August 26, we at NUPIAO learned that Dowson Tong, Senior Executive Vice President of Tencent Group and CEO of its Cloud and Smart Industries Group (CSIG), recently published a signed article titled “The AI Marathon: Thoughts from the Past Two Years” in Tencent’s internal publication Zhidian. The piece not only addresses the widespread chatter about “Tencent being slow in AI,” but also lays out his thinking on AI’s technological evolution, real-world applications, and organizational transformation.
Tong didn’t shy away from the pressure of industry chatter. He said candidly, “It’s impossible not to feel a bit anxious. The global large-model race is fierce, with fresh news hitting every day—who’s released a new model, how much bigger the parameters got, and benchmark leaderboards making the rounds everywhere. Our To B customers are paying close attention too, and in recent years, competitors have definitely had the upper hand when it comes to AI momentum.”
He admitted that insufficient computing power in the past did slow down the iteration pace of Tencent’s Hunyuan model. But he pushed back against judging an AI company’s fate purely by short-term hype.
“Getting up early might not be the most important thing. The second half of the AI race is just beginning—new opportunities keep emerging, the market landscape is shifting, and neither business models nor value distribution across the industry chain are set in stone,” Tong wrote.
He drew an analogy from the history of mobile internet, pointing out that plenty of early players in consumer electronics didn’t survive the industry shakeouts. Names like Nokia, Motorola, and BlackBerry fell behind in the competitive race and eventually faded into history.
In Tong’s view, what truly matters is never who took the first step, but who can commit to long-term investment with conviction, calmly think about the essence of things through cycles of breakthroughs and iterations, solve problems and pain points in concrete scenarios, pick out real signals from market noise, and preserve both strength and flexibility. That’s how you ride through cycles.
That’s why he prefers to view the large-model race as a marathon—and we might only be one kilometer in. “Getting up early matters, but it’s those who endure who get the chance to seize the new opportunities that keep coming.”
Tencent’s latest Q2 2026 financial report backs up the promise of long-term AI investment. The numbers show second-quarter total revenue of 204.79 billion yuan, up 11% year-over-year; R&D spending hit 27.28 billion yuan, jumping 35%; and capital expenditures reached 52.78 billion yuan, a massive 176% increase, with most of the new spending flowing into AI computing power procurement, large-model training, and inference infrastructure.
Tong also argued in the article that platform companies must stay committed to foundational large-model research, but they can’t fixate solely on the models themselves. “It’s like the mobile era—if you think only smartphone makers got a ticket to the game, that’s way too narrow.” In his assessment, scenarios are Tencent’s strongest hand in AI.
He also highlighted the dialectical relationship in tech deployment: “Algorithms may define the ceiling, but engineering capability determines how fast you reach it.” Enterprise applications that connect to intelligent agents via skills and MCP protocols to build a rich application ecosystem—that’s the more durable moat.
The article also traces the growth story of WorkBuddy, Tencent’s breakout product. Its predecessor came from a money-losing DevOps R&D tool project that nearly got cut during a cost-reduction push. The project survived by landing some revenue through private deployments of CodeBuddy. As large-model capabilities surged, and with the Tencent Docs team folded in to fill the document-editing gap, WorkBuddy finally evolved into a tool for everyday office workers.
In just three months since launch, WorkBuddy has shipped over forty versions. The team ditched the old, lengthy review processes, using AI to rapidly generate prototypes while humans focused on judgment, debugging, and quality control—essentially forming an AI-native agile development model.
Tong distilled the growth logic of AI-native applications from this: Innovation can’t be planned, but capabilities can be accumulated. Products need to give teams room to explore freely, home in on scenario pain points to quickly produce prototypes, and iterate continuously based on real user feedback.
As for the consumer-facing chatbot Yuanbao, there’s a persistent narrative that the Chatbot race is already decided. Tong disagrees. “Information search and Q&A are long-term user essentials. As long as you deliver better answers than the competition, opportunities will always exist.” In his view, the core of a Q&A product isn’t flashy human-like interaction—it’s answer accuracy, authoritative sources, and real-time information.
He shared that the underlying capabilities Yuanbao has built—like evaluation systems and feedback loops—are now being applied back to WorkBuddy. With multiple AI product lines advancing in parallel across the company, the Yuanbao team can offload some short-term growth pressure and focus on polishing foundational capabilities.
The article also reveals that through the Co-Design model with Hunyuan Hy3, Yuanbao’s search experience and user retention have kept improving, while unit operating costs continue to drop.
On the hotly debated Agent track, Tong proposed a dual-architecture view: “Personal Agents lean left, service Agents lean right.” Tencent won’t bet on just one type of intelligent agent. WorkBuddy represents the personal Agent, working on the user’s side to complete tasks. Meanwhile, the many intelligent customer service bots, HR consultants, and knowledge-base queries inside enterprises are service Agents—they need to follow management intent, take responsibility for data output, manage permissions, prevent misuse, and ensure 24/7 stable operation. The two aren’t substitutes; they’re complementary, together forming a complete architecture for enterprise AI applications.
When it comes to reshaping organizational capabilities in the AI era, Tong said AI-native doesn’t mean simply bolting an AI interface onto old products—it’s a full reboot of product design, R&D collaboration, and team structures. As AI takes over much of the standardized work, traditional assembly-line division of labor becomes bloated and inefficient. Cross-disciplinary collaboration and small-team sprints will become the norm. He stressed that models and products can’t be separated, “otherwise you end up with a painful split between soul and body.” The Co-Design approach, where models and products are built together, will be the key path for AI product evolution.
For individuals, Tong’s advice is simple: the faster the world changes, the clearer your head needs to be. “The ones who truly make a difference are rarely the most eager trend-chasers. They’re the long-distance runners who know their strengths and weaknesses, focus their energy on what they love and what they can actually change, and stay committed even through the valleys.”
