Tencent’s AI Strategy Just Did a U-Turn
After Months of Efficiency Talk, China’s Most Profitable Tech Giant Bets Big on Frontier Models.
On December 17, Tencent announced a sweeping reorganization of its AI research structure. The company created three new departments. It appointed Yao Shunyu, a 27-year-old former OpenAI researcher, as Chief AI Scientist reporting directly to President Martin Lau. The move signals urgency at the highest levels.
The timing matters. Just weeks earlier, Tencent had slashed quarterly capital spending to ¥13 billion in Q3 2025, down 24% year-over-year. President Lau told investors in November that “no decisively better model exists in China right now,” suggesting the compute arms race had reached diminishing returns. That narrative lasted barely a month. The December reorganization reveals Tencent’s brief experiment with restraint is over.
The pivot exposes deeper tensions. Tencent reported net profit of ¥194.1 billion ($26.8 billion) in 2024, far exceeding Alibaba’s ¥120.9 billion. Between Q3 2024 and Q2 2025, capital expenditure surged 178% year-over-year to a cumulative ¥102.3 billion. Only three Chinese tech giants spent over ¥100 billion on infrastructure during this period: Tencent, Alibaba, and ByteDance. The money flows. The compute exists. WeChat and QQ serve billions of users globally. Yet Tencent cannot match competitors in large language model capabilities.
According to UC Berkeley’s LMArena benchmarks, Tencent’s Hunyuan T1 ranks 68th globally as of mid-December. For context, an earlier version called Hunyuan TurboS stood at 8th place in May. Tencent excels at 3D and world models for gaming applications, with Hunyuan World-Voyager holding second place on Stanford’s WorldScore leaderboard. But conversational AI dominates commercial value today. The company’s struggles show in consumer products. Yuanbao, Tencent’s ChatGPT-style AI assistant, offers users a choice between its own Hunyuan model and the open-source DeepSeek. That dual option amounts to a public admission of weakness.
This presents a puzzle worth examining. Why would China’s most profitable tech company struggle to build competitive AI models? And what changed between November’s efficiency rhetoric and December’s organizational overhaul?
Profitable Companies Face Different Constraints
Tencent’s previous AI strategy delivered measurable results through integration rather than frontier research. The company embedded AI into existing high-margin businesses. Over 90% of Tencent engineers now use the company’s CodeBuddy AI assistant. AI generates 50% of new code across the organization. Code reviews involve AI assistance 94% of the time. Hunyuan powers over 900 internal applications, from WeChat Meetings to gaming analytics to advertising optimization. This approach delivered efficiency gains without the uncertainty of chasing frontier model capabilities.
The financial strategy aligned perfectly with this approach. Tencent spent ¥101.4 billion on share buybacks in 2024. The company’s major shareholder, Prosus (a Dutch investment firm that inherited Naspers’s Tencent stake), has been steadily reducing its position for years. Buybacks stabilize the stock price. They protect existing investors. Barclays noted in a November 2025 report that Tencent maintained “rational” AI investment focused on returns, contrasting favorably with competitors’ “aggressive” spending. For a public company with quarterly earnings expectations and fiduciary duties, this strategy made perfect sense.
The talent approach reflected similar calculations until recently. Tencent took a defensive posture in AI recruiting. The company did not chase top researchers with premium compensation packages. It relied on brand strength and existing pay structures. Foreign media reports indicate this changed dramatically in September when Yao joined from OpenAI. Since then, Tencent has offered double market salaries to poach researchers from ByteDance and other competitors. Fresh PhD graduates receive 50% premiums over industry standards.
The previous restraint had clear logic. Frontier AI research requires tolerating massive uncertainty. Leading research teams stay surprisingly small, typically 100 to 200 people even at top labs. DeepSeek reportedly operates with around 200 researchers total. Success depends on talent quality and research insights, not just compute and headcount. Throwing money at the problem guarantees nothing. For a company optimizing shareholder returns, the risk-reward calculation favored integration over innovation.
The organizational culture posed another obstacle. People familiar with the matter indicate that Tencent had been “downgrading” managers skilled in product execution and advertising algorithms but lacking frontier AI research backgrounds. This reveals how deeply application-focused thinking had penetrated AI efforts. Product managers drove decisions. Engineering leaders came from optimization and recommendation systems. These skills built WeChat’s advertising business to generate billions annually. They built game monetization engines. They do not necessarily translate to transformer architectures and reinforcement learning from human feedback. The December reorganization addresses this mismatch by elevating research leaders over product managers. But the fact that such changes were necessary shows how far the culture had drifted from frontier AI requirements.
Between Q3 2024 and Q2 2025, capital spending totaled ¥102.3 billion. Then came the pullback. Q3 2025 spending dropped to ¥13 billion, down 24% from the prior year. The November earnings call provided the rationale. President Lau’s statement about model convergence in China suggested Tencent had found a different path to winning. Integration beats innovation when capabilities plateau. This cost-benefit calculation worked when AI remained a feature enhancement rather than infrastructure. Tencent could license external models via APIs for specific use cases. It could offer users choice between Hunyuan and DeepSeek. The capital efficiency looked attractive compared to building proprietary frontier models from scratch.
The December reorganization reveals this calculation no longer holds.