Foreword:

Recently, sitting beside Pony Ma is Yao Shunyu, the 28-year-old Chief AI Scientist.

At almost the same time, young faces are concentratedly moving towards the center of technological power, coinciding with a new round of intensive iterations in domestic large models.

Author | Fang Wensan

Image Source | Internet

Answers written in one summer, from ringing the bell, open-sourcing, to pricing

In January this year, Zhipu AI was listed on the main board of the Hong Kong Stock Exchange, becoming the world's first "large model stock." The next day, MiniMax followed suit, with both unicorns going public within 24 hours; by June, Zhipu's market value once surpassed the HK$1 trillion mark.

Before the applause in the capital market faded, the exam paper immediately turned to the next page: Zhipu's revenue in the first half of the year was 954 million yuan, nearly quadrupling year-on-year, but net losses also reached 2.072 billion yuan. Ringing the bell only secured the exam admission ticket; the break-even line is the real exam question.

The real heatwave came in summer when Moonshot AI released Kimi K3 in July, featuring 2.8 trillion parameters, full weight open-sourcing, and 1 million token context. Caixin's cover story described it as the heaviest chip in the domestic models' "programming arms race" against Claude, while QianWen and DeepSeek's new open-source models directly impacted the pricing of overseas closed-source models.

Other players were also busy: in March, Alibaba open-sourced four Qwen3.5 small-sized models, earning a public thumbs-up from Elon Musk; Tencent's earnings call revealed that Yao Shunyu's Hunyuan 3.0 had entered internal business testing; Doubao released its team agent, "Doubao Work Partner," in September via Feishu 8.0.

Behind the Intensive submission (intensive submissions) lies an industry's unspoken consensus: the freshness period of model capabilities is measured in weeks—whoever stops writing is out.

The dealer has changed hands; a group of post-90s have sat at the center of the table

At the Hangzhou Yunqi Conference this September, Liu Dayiheng stood on the main forum, speaking right after Alibaba CEO Wu Yongming.

His identity is the project leader of Alibaba's ATH Business Group Token Foundry Qwen LLM. Born in 1993 and once selected for Huawei's "Genius Youngster" program, this young man has taken on the important responsibility of developing the QianWen foundational model.

At the same conference, Chen Yusen, born in 1992, made his debut as the CEO of QianWen Office. Previously, he founded the cybersecurity company Changting Tech before joining Alibaba to participate in AI Agent product entrepreneurship. In June 2026, he succeeded as CEO of DingTalk and later took over the integrated QianWen Office business.

The positioning of these two roles reveals an interesting division of labor in Alibaba's AI strategy.

Liu Dayiheng is responsible for pushing the upper limits of model intelligence, while Chen Yusen focuses on integrating intelligence into enterprise workflows. One faces technological supply, the other application demand.

The model team needs to consider how to enhance complex reasoning, code generation, tool usage, and long-term task execution capabilities; QianWen Office must address another set of questions: where are enterprise data scattered across systems, what permissions can agents obtain, how to complete work across applications, and how much efficiency improvement customers are willing to pay for.

Tencent's table has been handed over to even younger players. Yao Shunyu, born in 1998, a Tsinghua Yaoclass alumnus and Princeton Ph.D., proposer of the ReAct and Tree-of-Thought frameworks, left OpenAI at the end of 2025 to return to China as Tencent's Chief AI Scientist, overseeing large language models and AI Infra.

Less than a year into his tenure, Tencent merged the Hunyuan Large Language Model Department and Multimodal Model Department into a Foundational Model Department in July 2026, placing all text, image, video, voice, and 3D R&D under his purview.

Yao Shunyu pursued his Ph.D. at Princeton University, participating in proposing intelligent agent research methods like ReAct before joining OpenAI. His research background resonates with Tencent's current emphasis on Agent capabilities.

Tencent has placed foundational model R&D, training, and inference infrastructure under a tighter management system with clear engineering considerations.

The longer the R&D decision-making chain, the more likely technological advantages are constrained in productization speed. Having core leaders participate in both model and infrastructure decisions reduces some coordination costs.

Xiaomi has adopted a different approach. According to multiple media reports, Luo Fuli has been promoted to Level 22 within Xiaomi, a division of labor closely tied to Xiaomi's terminal business. Smartphones, automobiles, and smart homes have distinct AI demands; even if cloud models possess excellent reasoning capabilities, device computing power, latency, power consumption, and connection stability must be considered.

ByteDance is doubling down in another direction. After Zhou Chang joined ByteSeed, his management scope gradually expanded to cover Seedream, Seedance, world models, and related embodied intelligence research.

For ByteDance, with its short-video, content creation, and commercial advertising scenarios, multimodal generation capabilities inherently have shorter product validation paths.

The responsibilities shouldered by these young leaders vary and cannot be simply equated with Overall management (comprehensive control) over each company's AI strategy. Computing budgets, capital investments, and ultimate operational responsibilities are still jointly determined by larger organizational systems.

But one change is clear enough: individuals who can straddle research, engineering, and product boundaries are gaining greater decision-making space than traditional R&D roles.

This organizational evolution has its technological context. The Transformer architecture emerged in 2017, and the new industrial wave driven by large language models rapidly unfolded at the end of 2022. The pace of technological paradigm shifts is so fast that many experience-based advantages built on long-term accumulation face reevaluation.

A researcher's participation in several rounds of model training and understanding of the relationships between reinforcement learning, data engineering, and inference deployment may hold more direct value than years of managing general software R&D teams.

Age itself has not driven this change; the new technological cycle has altered how companies identify and empower core talent.

The control group illustrates the point: Liang Wenfeng of DeepSeek, born in 1985—the very person who ignited this wave—has already become an "elder statesman" at the table. A generational time scale has been compressed to eight years in the AI industry.

Scorecard: The ruthless stratification of disruptions, growth rates, and retention

Submitted answers must eventually be graded. QuestMobile's H1 2026 report offers a set of stark scores: as of May, AI-native apps collectively had 499 million monthly active users (MAUs), up 85.4% year-on-year, with per-user monthly usage reaching 183 minutes, officially crossing the threshold of daily necessities.

But the distribution of scores is highly uneven: Doubao leads with 382 million MAUs, followed by QianWen at 167 million and DeepSeek at 130 million, with these three accounting for nearly 90% of the TOP 10's total.

The growth rankings are even more dramatic: QianWen surged 5,792.9% year-on-year, Doubao grew 172.1%, while DeepSeek declined 20.3%.

A glaring gap has opened between technological prestige and user scale. Moonshot AI, whose K3 shook Wall Street, had 7.29 million MAUs, down more than 40% year-on-year, ranking ninth.

DeepSeek had the highest proportion (30%) of users engaging for over 10 minutes per session but continued to lose users overall.

Model supremacy and consumer-side losses can coincide—this is the most counterintuitive yet real lesson of 2026: the C-end war hinges on ecosystems, entry points, and subsidies; paper citation counts don't translate to daily actives.

Commercialization scores are equally mixed. Doubao launched a tiered subscription professional version in June, hitting a record 178 million DAU the next day, with business office penetration soaring from 19.2% to 49%.

Moonshot AI's valuation surged sevenfold to $50 billion in six months, with ARR reaching $300 million; DeepSeek's revenue hit 475 million yuan in the first seven months—about ten times the previous year's total—but still posted a net loss of 715 million yuan.

High scores and deficits coexist on the same report card; no candidate is spared.

The louder the applause, the clearer the need to examine the back page of the exam.

The reality of burning cash hasn't disappeared: Zhipu posted an 820 million yuan net loss in H1 2025, with R&D investment exceeding 50% of revenue. MiniMax lost 650 million yuan in the same period, with Amazing income growth rate (astonishing revenue growth) but no scale effects yet.

Primary market liquidity is receding. CVSource data shows AI sector financing fell 42% year-on-year in H1 2025. After going public, markets demand profitability timelines every quarter; "long-term investment" is no longer a free pass.

Organizational exams are equally thorny. MiniMax experienced directional swings in 2024, with Intensive flow ( Intensive personnel flow ) among mid-level managers; half of its initial 100 employees had left.

Moonshot AI's valuation soared to $20 billion, but Yang Zhilin remains nearly the sole pillar of the company's technical narrative, with unresolved issues around bench depth.

Across the table, ByteDance and Alibaba launch saturation attacks using in-house models and ecosystem discounts, leaving startups with no safe flank.

The first half tested model capabilities; the second half tests organizational strength and operational quality. The post-90s generation's required reading has shifted from papers to financial reports, from single-point breakthroughs to long-term operations. Youth is a dividend but also a deadline.

The market's patience for young examinees is often shorter—and more expensive—than for their elders.

Epilogue:

The most captivating aspect of the large model industry is this: a paper written at 25 can reshape the organizational structure of a $3 trillion company by 28 and force Wall Street to check the markets by 33.

The post-90s generation isn't inheriting power—they're inheriting time. When technological iteration cycles are shorter than a generation's growth cycle, youth itself becomes a hard currency. The game is long, but the dealership has changed hands.

Partial references: Investment Interface: "Investors Start Staking Claims on Yao Shunyu," AI Brain Era: "2026 AI IPO Wave: Zhipu, MiniMax, Moonshot AI, DeepSeek's Capitalization Paths and Commercialization Tests"