The AI Battlefield Still Has One Missing Piece: A Tencent Powerhouse

“Tencent AI has finally found its Yao–Shun–Yu.”
That’s the tongue-in-cheek line making the rounds online after former OpenAI scientist Yao Shunyu joined Tencent and the company reshuffled its AI organization. But the joke points to something serious: Tencent is no longer content to play defense in the foundation-model race—it’s starting to push forward, and it’s betting that elite talent and tighter infrastructure execution can open the breakthrough.

On December 17, Tencent issued an internal announcement: within TEG (Technology Engineering Group), it created a new AI Infra Department and a Data Computing Platform Department, renamed the former Data Platform Department into AI Data, and dissolved the original Machine Learning Platform Department. Yao Shunyu was appointed Head of AI Infra and Chief AI Scientist.

Yao’s arrival stands out in two unusual ways.

First, it’s a fast-track promotion with real authority. As a post-95s leader, he’s taking charge of core model-training infrastructure at Tencent. As Chief AI Scientist in the CEO & President’s Office, he reports directly to Tencent President Martin Lau. As the Head of the LLM unit under AI Infra within TEG, he also reports directly to TEG President Luo Shan. That dual reporting line is rare inside Tencent’s technical system—and it signals that “AI Infra” is being treated as a top-tier strategic priority.

Second, Tencent handled the news with notable restraint. Back on September 12, Tencent’s official account “Goose Factory Blackboard” publicly denied a rumor that Yao had joined Tencent on a “hundreds-of-millions” compensation package, calling it untrue. Later, industry insiders said Yao had already started internally and attended meetings, suggesting Tencent’s denial was more about limiting public impact than refuting the underlying move.

Three months later, the final confirmation arrived—underscoring Tencent’s preference to keep its real AI progress out of the spotlight.

And Yao isn’t the only signal. Reports indicate Feng Jiashi has also joined Tencent as the head of the multimodal team at its AGI Research Center. He previously led ByteDance Seed’s vision team and reportedly briefed Zhang Yiming on frontier AI topics. In other words: Tencent’s recruiting engine is accelerating, and it’s targeting the exact capability gaps that matter most in the next phase.

Ever since OpenAI pushed the world into the foundation-model era, Baidu, Alibaba, and ByteDance moved quickly. Tencent, by contrast, looked cautious. According to LatePost, after ChatGPT launched, TEG President Luo Shan told an internal strategy meeting that GPT took three years to productize and Tencent’s model journey would involve many pitfalls—so there was no need to rush.

Many AI practitioners describe Tencent’s historical pattern the same way: less aggressive, slower to pivot, more inclined to ensure it doesn’t fall behind on base capability—while looking for breakthroughs at the application layer.

But reality has gotten sharper. ByteDance’s Doubao and Alibaba’s Qwen are integrating faster across their internal ecosystems—Doubao connecting into Douyin commerce flows, Qwen plugging into Amap and enabling map and street-view style tool calls. Meanwhile, Tencent’s AI “moment” still feels scattered. For many users, the most visible changes remain limited to WeChat’s ecosystem—public account features, video channel comment tools, and lightweight AI assists.

So even if Tencent can afford patience at the base-model layer, it can’t afford invisibility at the application layer. Without more real scenarios—and more frequent experimentation—Tencent risks losing mindshare and momentum where users actually feel AI.

01. Yuanbao Needs to Penetrate the Tencent “Extended Universe”

Earlier this year, major internet platforms rushed to integrate DeepSeek. Now the next phase is even more intense: top models are fighting for commerce guidance, local services, healthcare, and other “high-frequency, high-intent” user scenarios—trying to lock in the habit loop from Q&A to transaction conversion.

Alibaba and ByteDance are clearly moving fast on the application front.

Alibaba recently reorganized under VP Wu Jia, merging the Intelligent Information Group and Intelligent Interconnection Group into a Qwen Consumer Group, covering UC, Quark, Shuqi, and AI hardware. The goal is blunt: Qwen becomes the primary AI-to-C entry point, effectively replacing Quark as the flagship gateway. Alibaba’s positioning is “ALL in ONE”—one app spanning foundation models and a full stack of daily-life services.

Alibaba’s plan is ambitious: Qwen aims to surpass Taobao and Amap as the super entry point for the AI era, using natural-language conversation to help users design travel routes, book tickets and hotels, compare products across platforms, and complete tasks end-to-end. On the hardware side, Qwen is expected to span PC, phone, car systems, and smart glasses—turning into a cross-device personal super assistant that can chat and actually get things done.

Ant Group is also pushing hard, upgrading its AI health product into Ant A-Fu, offering health data monitoring, habit tracking, medical Q&A, assisted registration and drug purchasing, and insurance payment support.

ByteDance’s Doubao tells a similar story—only more aggressively tied into consumption flows.

Doubao has already linked into ByteDance’s key products: Douyin, CapCut, “Jimeng,” and Feishu, and it actively routes users into commerce scenarios. When users ask shopping questions, Doubao recommends matching products and provides direct Douyin Mall links. For local consumption needs, it can send users straight to group-buying pages for nearby merchants.

Against that backdrop, Tencent’s biggest strength—social—becomes a strategic tension. It has the largest social graph, but it lacks the kind of “AI-first application matrix” that Alibaba and ByteDance can mobilize through their owned platforms and tightly integrated scenes.

Inside WeChat, Yuanbao is already becoming a useful add-on for many users. Features like AI search, comment summarization, and direct chat from friend lists have earned positive feedback. Yet up to now, Yuanbao still lacks the kind of strategic depth Doubao and Qwen enjoy. It appears more often as a supporting feature embedded into Tencent Docs or an input method—rather than a unified, scenario-owning AI entry point.

One of the few visible “ecosystem link-ups” happened in February, when Yuanbao appeared in WeChat’s nine-grid alongside entrances like JD.com, Meituan, and Pinduoduo—nudging users to download and try it.

Under external pressure, Tencent may need something bigger: an AI “alliance army.”
Use WeChat as the strategic entry, Yuanbao as the connective layer, and treat WeChat’s nine-grid “friend circle” as strategic partners—pulling in Meituan, JD, Pinduoduo, Tongcheng, Didi, Vipshop, Zhuanzhuan, Maoyan, and others—to proactively cover e-commerce, delivery, group-buying, travel, and ticketing scenarios, building an ecosystem counterweight to Alibaba and ByteDance.

And the window for integration may be narrower than it looks.

In September, Meituan began public testing its first AI Agent product, Xiaomei, designed for one-sentence restaurant recommendations, travel route planning, delivery ordering, and hotel booking—powered by Meituan’s own “Longmao” model. JD is pushing shopping and life services via its assistant Jingxi. Tongcheng Travel has its Chengxin model for travel planning and booking. These products are early-stage in both model strength and user scale, but the direction is unmistakable.

At the same time, for smaller companies without strong in-house models, integrating independent models like DeepSeek is increasingly the rational choice for cost and efficiency. For second-tier internet platforms, though, AI is being treated as a next-generation “core capability,” not a component to outsource or sell. Once AI becomes the platform’s “soul,” Tencent’s historically investment-heavy, non-controlling partner ecosystem looks structurally weaker than the fully owned, multi-scenario application networks of Alibaba and ByteDance.

That’s the real pressure on Tencent: without integrating model + product + scenario, it’s hard to hold a heavyweight position in the AI competition.

02. Is Being a Latecomer to Foundation Models Really “Not Urgent”?

On Tencent’s Q3 earnings call, Martin Lau’s remarks effectively summarized Tencent’s official stance on the model race:

In China, Tencent does not believe there is a single absolutely leading model. Competition is tight; different models have different advantages in different scenarios. Tencent doesn’t view itself as behind, and as Yuanbao continues improving, user activity is steadily rising—suggesting both model capability and AI products are progressing.

Pony Ma has echoed similar thinking internally: stay clear-eyed, don’t overestimate capability, and wait for the inflection point.

But Tencent’s actions suggest a quieter anxiety underneath.

The most visible example was Yuanbao’s massive paid acquisition push in February. As DeepSeek sparked a wave of AI excitement, Tencent quickly pushed Yuanbao to the foreground, flooding nearly all its platforms with ads. According to AppGrowing data, Yuanbao’s ad spend in Q1 reached about RMB 1.4 billion, and monthly active users briefly approached 40 million, topping the China iOS free app chart—surpassing DeepSeek and Doubao.

Yet the spend didn’t lock Yuanbao into the top tier long-term. During that campaign, the core message was “full-powered DeepSeek,” not Tencent’s own Hunyuan model. After the short burst of heat, Yuanbao was overtaken. QuestMobile data shows that for December 8–14, Doubao’s weekly active users hit 155 million, while DeepSeek and Yuanbao were 81.56 million and 20.84 million respectively. Newer entrants like Ant A-Fu and Alibaba Qwen posted 10.25 million and 8.72 million weekly active users.

Beyond paid growth, Tencent has made several strategic adjustments this year.

First: centralization. Tencent consolidated AI apps and products previously scattered across business groups. In February, it moved TEG’s Yuanbao, PCG’s QQ Browser, Sogou Input Method, and ima into CSIG to explore growth and monetization paths.

Second: product intelligence upgrades. After DeepSeek-R1 went viral, Tencent rapidly embedded relevant capabilities into WeChat, QQ Browser, Tencent Docs, Tencent Maps, QQ Music, and other flagship products—ensuring its core experiences wouldn’t fall behind competitors.

Third: organizational reshuffling. Tencent reorganized TEG’s foundation-model structure, establishing dedicated teams for LLMs, multimodal models, intelligent computing architecture, and collaborative innovation. Now it has further created AI Infra and a Data Computing Platform unit—trying to solve fragmentation and push breakthroughs through stronger foundational engineering.

From the outside, Tencent seems to be pursuing a “latecomer counterattack,” keeping Hunyuan close in launch timing to Doubao, but with weaker market presence. Many users interviewed report that Hunyuan’s experience feels less compelling than Doubao’s.

One key reason: Tencent’s Hunyuan tends to optimize inward, while Doubao pushes outward.

On December 18, Volcano Engine released Doubao 1.8, clearly marketed around “helping users do work”—stronger Agent capability that can execute complex tasks autonomously, such as planning a trip and buying tickets, organizing industry materials into documents, or generating a complete short video from ideation to final production.

Hunyuan, meanwhile, has emphasized internal efficiency in WeChat and Tencent Meeting. One representative showcase is 3D model generation—creating navigable 3D worlds from text and images. That may matter a lot for Tencent’s film and game businesses, but it’s less obviously valuable to everyday users.

This is Tencent’s contradiction in motion: it publicly stresses strategic patience and rejects “quick-win fantasies,” yet it also reorganizes teams, empowers star scientists, and aggressively hires from OpenAI and Doubao—trying to regain startup-like urgency and close the gap fast.

03. Tencent Cloud Is Being Squeezed by the AI Wave

Among Tencent’s internal “allies” hoping the base model story strengthens, CSIG is at the top of the list.

Tencent Cloud is currently leaning into a defensive posture, prioritizing stable growth. At its September ecosystem conference, CSIG CEO Tang Daosheng reiterated a familiar line: Tencent Cloud would rather have “150 jin of muscle than 200 jin of bloated weight.”

But pressure is building from every direction. Omdia data shows that in Q2, Alibaba Cloud, Huawei Cloud, and Tencent Cloud held market shares of 34%, 17%, and 10% respectively.

This may connect to Tencent’s emphasis on profitability in IaaS and SaaS—rejecting irrational revenue growth—and its tendency to prioritize internal needs. On the Q3 earnings call, Martin Lau noted that when AI chips are scarce, Tencent prioritizes internal usage rather than external leasing. In other words, if chips weren’t constrained, Tencent Cloud revenue could likely grow faster.

The tougher test is at the MaaS (Model as a Service) layer, where Tencent Cloud faces Volcano Engine head-on. Public data indicates that by December, Doubao’s daily token usage exceeded 50 trillion, ranking first in China, with annual total token consumption surpassing 90 quadrillion.

IDC’s first-half report put Volcano Engine at roughly 264 quadrillion token calls and 49.2% MaaS market share, ranking first. Alibaba Cloud followed at 27%, Baidu was third, and Tencent Cloud was grouped into “Others,” with a combined share of 6.8%.

Beyond Volcano Engine’s aggressive price competition, Tencent Cloud also hasn’t shown strong momentum in multimodal models.

In the 1.8 release, Doubao not only improved Agent tasks, but also doubled its single-pass video understanding frame count from 640 to 1280, enabling low-frame-rate comprehension of very long videos and tool-based high-frame-rate reasoning on key segments. That creates clear opportunity in education, surveillance, manufacturing inspection, and more.

ByteDance’s latest Seedance 1.5 Pro video model reportedly supports millisecond-level audio-video sync output. On Alibaba’s side, Tongyi released its visual reasoning model QVQ-Max as early as March, and more recently integrated its Wanxiang 2.6 video generation model into the Qwen app, offering AI “mini-theater” features to users.

Compared to that, Hunyuan does have a video generation model, but its market visibility and public use cases lag behind the leading pack. That may explain Tencent’s intensified hiring from OpenAI and Doubao: to accelerate multimodal breakthroughs and integrate them into WeChat’s massive ecosystem—finally solving the “last mile” from capability to user-scale impact.

This is where Yao Shunyu’s arrival could matter most. His hire may mark Tencent Cloud’s counteroffensive signal: on one side, Tencent is aligning its foundation-model structure more closely with Doubao—empowering key scientists, strengthening core research, and tightening infra execution. On the other side, WeChat Agent is emerging as a potential next breakout point. As a national-level super app, WeChat’s Agent capabilities could deliver a “dimensionality reduction” advantage over smaller products—revitalizing Tencent’s broader AI playbook from the inside out.