Technology Cycles, Traffic, and Artificial Intelligence — 张小珺 x 郑庆生
Summary
郑庆生把二十年中国科技周期归结为“流量方式”的连续重构，AI正在争夺新一代分配入口。 Before 2005, portals completed the first wave of moving information online. From 2005 to 2010, Tencent, Baidu, Shengda and others held the high ground in PC traffic; after 2010, mobile reshuffled the gateways; short video took off from 2015, and by around 2018 the new traffic had largely been divided up. ChatGPT has reopened the To C window. In his analogy, great internet products are “huge towns”: only after traffic converges do new commercial forms emerge.
内容平台的终局并非视频替代另一种媒体，而是短视频直接“挑战文字本身”，AI则进一步把效率产品从流程推向结果。 Text requires years of learning, while images are more immediate. Short video comes close to taking people to the scene for a look, making it a basic mode of cognition; as long as the information density is high enough, not reading books is not inherently shameful. AI is moving in the same direction: “tool products are results, entertainment products are processes.” In any efficiency scenario, users want as little involvement in the process as possible.
AI网络同时带来三项结构性差异：边际成本不再趋近于零、产品直接对结果负责、底层原理仍存在黑箱。 Every additional user creates token costs, turning subscriptions, pricing and user selection into real unit-economics questions. AI solutions may also bypass structured workflows and deliver results directly from raw information. More uncertain still are emergence and the “mysticism” in prompting: technology and products are evolving in sync for the first time, and model advances can create new applications while also changing the advantages of existing products.
软硬一体的核心机会不是再造一块屏幕，而是完成AI驱动的“深层次数字化”，把过去会随风飘散的信息变成可加工资产。 One hundred hours of recordings or tens of thousands of automatically captured photos a month were previously impossible to organize even if they were collected. AI makes dedicated recording, imaging and continuous-sensing devices possible as new data and traffic nodes. The key advantage of dedicated hardware over phones is that it is “always present,” but whether data, habits and usage patterns can become moats—and where the privacy boundary lies when recording in public—remain questions for law and social norms.
大模型如果胜出，可能成为入口，但应用公司的护城河、显性双边网络效应和AI社交形态仍没有答案。 Models may have an implicit flywheel: the more data and interactions they accumulate, the better the feedback they deliver to new users. But most users still talk to a model in a “many-to-one” structure. If a competing model overtakes it in quality, a product without network relationships could be highly vulnerable. OpenAI has accumulated usage habits and data and has strong product-company characteristics. Its ultimate value boundary will depend on how much technological progress offsets nontechnical product advantages.
郑庆生将当前阶段定义为应用“爆发的前端”，认为泡沫正常且目前可以接受，但最终必须以交付检验。 Unlike previous cycles, when companies first burned money to build scale and had little revenue, this cycle can hit pain points, charge customers and generate decent cash flow from the start. Strictly speaking, rapidly changing models can still invalidate existing PMF, while massive infrastructure costs remain mismatched with revenue. He compares 2025 with 2010 in mobile internet and 2026 with 2011, and expects applications to flourish further.
这一轮最重要的地域性投资判断是“华人大航海时代的开始”：AI产品天然适合全球部署，中国与海外华人团队可以从第一天服务世界市场。 Sequoia China did not treat foundation models as a capital-allocation race to cover wholesale, but entered Kimi, MiniMax, 智谱, Manus and other projects one by one. MiniMax was on its radar before ChatGPT took off, while Manus was tracked from the team’s capabilities and products during the Monica period. Since establishing its seed fund in 2018, Sequoia has moved further upstream; projects where it is the first institutional investor now account for more than half of its investments.
真正最大的未知不是AI产品还会增加什么功能，而是使用AI之后的人类会变成什么样。 If users eventually stop caring whether their conversational counterpart has independent consciousness or a “soul,” the default assumptions behind socializing, intimacy, education and consumption will all change. Education may place more weight on a global view, taxonomy, logic and precise expression, because learning also includes learning how to ask questions. For investors, the corresponding epistemic guardrail is to reconstruct the full set of options available at the time of a decision, rather than accept the faith stories written by survivors after the fact.
Deep dive
1. BASIC on a 1984 computer planted product intuition, but career choices still followed the 1990s job-market order
郑庆生 says his path feels like “everything was somehow destined”: his father was a programmer, and in 1984, shortly after starting primary school, he learned BASIC from a book on a Sharp microcomputer, copying a frogger game and writing a keyword-matching “human-machine dialogue” program.
He admits he had no outstanding aptitude for science or engineering and later studied economics at Fudan. But early programming taught him that handing a program to classmates and watching how other people used it was itself “very interesting.” He later came to describe this product-experience instinct as being a “product experience officer.”
In 1996, he attended an internet lecture at Fudan, used the campus BBS and served as moderator of the economics forum. Around 1998, a classmate studying optics recommended Google, which he used for the first time as a more direct alternative to Yahoo’s directory. By the time he graduated in 1999, internet companies were still not a mainstream career destination. He chose auditing at a foreign company, while trying almost every internet product he could access.
2. Audit and consulting trained an intuition-led person to become an evidence-based investor
The biggest change three years of auditing brought 郑庆生 was the requirement that every judgment have evidence, every working paper be cross-referenced, and every number be “accurate to two decimal places.” Without that training, he believes investing would have been much harder later: intuition remained, but it had to be traceable back to logic and evidence.
Compliance inspections in manufacturing after China joined the WTO took him across factories in Zhejiang, Jiangsu, Shandong, Shanghai and the areas around Beijing, covering apparel, toys, electronics, ties and more. He examined every production line and even employee areas. Nearly a decade later, when he first looked at a new-materials project, that brief experience became an intuitive foundation for understanding Chinese factories.
His next three years in process consulting focused mainly on implementing SAP, Oracle, PeopleSoft and Hyperion: first define the client’s processes, then make standard software fit them. When he studied SaaS in 2018–2019, he realized that what he had done early in his career looked much like customization in SaaS implementation. Six years in professional services also became the foundation for his later understanding of To B.
3. Joining Shengda in 2005 put a heavy internet user inside the industry for the first time
The turning point came through a classmate working at Shengda, who told him: “Shengda is China’s largest internet company; its revenue exceeds that of all other internet companies combined.” 郑庆生’s reaction was that he had already been online for six years outside work—so why not work for an internet company?
His background was not suited to a direct product role, so in 2005 he joined Shengda’s newly established investment division. It was the first post-infrastructure internet cycle’s “hundred flowers blooming”: 起点, dynamic-password technology, various games, internet cafés and battle products emerged in succession. But his experience at the time was that an investment institution might do only a few deals a year, which did not support statistical conclusions.
Shengda was also the only operating company where he worked deeply. He saw for the first time how departments collaborated, how new products moved forward, and how business development and cross-regional operations were organized. In 2007, he joined 智信资本 with Shengda’s former CFO and others, formally becoming a financial investor. In hindsight, simply buying listed internet giants at the time might have produced better returns than many VC investments; no one could foresee how large the digital wave would become.
4. Douban and Dianping turned investing from a job into a long-term method for understanding the world
What truly moved 郑庆生 about Web 2.0 was not complex technology but a reversal in the relationship between people and information. Newspapers, webpages and search all required active behavior, while UGC and reviews could return “things you like but didn’t know you liked.” Delicious showed him a large number of niche sites that could not be found through Yahoo’s directory or active search; Douban brought the mechanism to the Chinese internet.
When Douban launched in 2005, he wrote an internal report calling the format highly novel. Later, during his first long conversation with 阿北 at Diaoke Shiguang in Wangjing, Beijing, he felt that the founder and product were “completely one,” with Douban resembling a projection of the founder’s inner world. He also said that later, both at 智信资本 and Sequoia, the funds he worked for had shareholder relationships with Douban and Dianping, though his original remarks did not specify the direction of those holdings.
When he met Dianping founder 张涛 in 2011, 郑庆生 saw a different character: more mature, more institutionalized, but still equipped with sharp product insight. Douban and Dianping became, in his eyes, the culmination of Web 2.0 and confirmed that investing could be used over the long term to “explore patterns in human behavior and understand the world.”
5. Two decades of internet history can be divided into four eras by the emergence of new traffic gateways
Before 2005 was the beginning of moving information online from scratch. Sina, Sohu and NetEase held the first wave of traffic; Tencent captured communication, Baidu captured search, and Shengda captured games, music and 起点. By today’s standards, they were still just unicorns or innovative companies worth several hundred million dollars.
From 2005 to 2010, these companies held the main high ground of the PC era. Community, UGC, e-commerce and gaming innovation continued, but it was difficult for new projects to take the largest gateways from established giants. It was only around 2010, when mobile internet emerged and old gateways gave up market share, that a new reshuffling began.
郑庆生 places particular emphasis on Pinterest. If he remembers its founding date correctly, it was around 2010, and its waterfall feed was extremely well suited to phones—a product-paradigm revolution. He subsequently spent substantial time in China looking for similar formats and met deeply with Xiaohongshu and Mogujie, viewing this lineage as an influence on content platforms including Douyin, TikTok, Xiaohongshu and Kuaishou.
After 2015, short video began to explode. Meituan, Didi, Mobike and other O2O and online-offline businesses ran throughout the period. By around 2018, the new round of To C traffic had largely been divided up. From 2018–2019 until the rise of AI, consumer-internet investing therefore entered a relatively quiet period.
6. Short video won not a video category, but the basic way humans understand the world
郑庆生 calls text “a highly advanced form of knowledge product.” Mastering written language requires years of education, which is why every country has needed mass literacy campaigns. Images are more direct, so mixed text-and-image formats naturally capture some consumption that once belonged to pure text. Reading an entire book carefully has already become a luxury in terms of time.
Long-form video descends from drama, film and television; it merely changes the interface. Early on, he also liked its business model because of its sufficient length and viewed short video as a smaller video category. He later realized the two should never have been compared on the same plane.
His most powerful analogy is looking at the moon. Writing a paragraph or even a poem to describe the moon is sophisticated, but the simplest approach is to take someone to the scene and say, “Look—isn’t this moon huge, isn’t it round?” Short video replicates that sense of presence. “It is not challenging other media formats. It is challenging text itself.”
He therefore does not believe that failing to read a book in a year is necessarily shameful, provided someone receives information of comparable density through other formats and is not simply “killing time.” Short video becoming a basic cognitive tool is not abnormal in his view.
7. Products are jointly invented by developers and users; AI is simply pushing the path toward results
When 郑庆生 first used Twitter in 2005 or 2006, he did not understand why posts were limited to 140 characters. He could only record what books he had read or where he had gone and wondered, “Who would look at this?” Only after users began writing opinions and repeatedly @-mentioning others did the platform develop its real use through interaction. He later used 饭否 and Weibo as well.
When he started using Musical.ly in 2015, he simply filmed scenery with music and never imagined that dancing would make the product explode. The lesson was not a predictive formula, but that developers provide the structure while creators and users redefine it through use; even the founder cannot fully foresee the product’s boundaries.
Content consumption has moved gradually from active search toward recommendation and become more fragmented. AI is pushing efficiency scenarios toward a more natural endpoint. 郑庆生’s distinction is “kill time and save time”: “tool products are results, entertainment products are processes.” When users want to pass time, they want the process; when they want efficiency, they want the result as quickly as possible.
8. The difference between Douyin, Xiaohongshu and Bilibili lies in how content is organized, not in media format
郑庆生 stresses that these are personal judgments or user-level impressions. On Douyin, the content itself matters most. When a KOL with millions of followers posts something new, followers usually form the basic click-through pool. Xiaohongshu relies more on the quality of an individual post for redistribution; millions of followers do not automatically translate into equivalent exposure, and comments are also an important part of content consumption.
He sees Xiaohongshu as the “most open product structure” of the mobile-internet era. Images, short essays, guides, questions and comments can all fit inside the same posting framework, absorbing most of the UGC and social-media formats that came before. This allows it to co-create its tone continuously with users, expand from women to men and extend offline into activities such as City Walks.
Bilibili followed a different route: it first entered an influential niche community, then gradually broadened to the general population and high-quality content such as documentaries. Unlike Douyin and Xiaohongshu, it did not adopt a more open structure from the outset. 郑庆生 also believes Bilibili’s founder and product share similar characteristics.
ByteDance already had Toutiao and a recommendation engine, so entering short video was a rational strategic choice made from a higher level, with the company completing its positioning through moves including the merger with Musical.ly. 郑庆生 does not believe the different paths have an inherent ranking. They show different product histories: “the times make the hero,” and success is difficult to reduce to a single formula.
9. Investors cannot predict the next behavior, but they can recognize a work once it has appeared
“Perhaps the summary is that there is no summary.” 郑庆生 believes new patterns of human behavior are broadly unpredictable. Standing behind the outcome today, it is easy to explain why a platform succeeded; when short video first appeared, almost no one knew it would ultimately challenge text.
Founders often cannot provide a complete rational forecast or supporting rationale either, but they may possess intuition and elevate it into belief. The problem is that history preserves the faith stories told by winners, while the same beliefs held by failures disappear from the narrative. Retrospective summaries therefore carry an inherent survivor bias.
He compares his role to that of a literary critic or publisher: “I don’t write novels,” but once a novel appears, intuition may suggest it has the potential to become a classic. AI investing likewise means not declaring the answer in advance, but continually engaging with products at the frontier that are most likely to challenge the largest traffic forms, then deciding which deserve deeper work.
10. Didi and Pinduoduo forced him to correct a one-track path of “advanced users go first”
郑庆生 once summarized his preference as “investing in the advanced lifestyles of advanced users.” Dianping, Douban, Xiaohongshu and Musical.ly often first reached users with content-creation ability, a knowledge orientation and a presence in first-tier cities, then spread to the broader population.
Missing investments such as Didi and Pinduoduo made him realize that he did not sufficiently understand China’s most basic and broadest population. Some models do not need to pass through “advanced users” first; they can spread directly from grassroots demand. This became the most important correction to his existing product taste.
Bike sharing and the sharing economy reflected that correction, but also showed him that the logic could not be extrapolated indefinitely. The sharing economy later expanded into power banks, space and other forms. He had initially expected more, then acknowledged that reality had a ceiling. The same logic can deeply change grassroots life without necessarily covering every asset category.
His online-offline integration example is Klook: online, it aggregates traffic through purchase demand, UGC and reviews, then uses procurement scale to influence the supply chain for Asian attraction tickets, subway tickets and distinctive activities. Operations, supply control and organizational design are themselves products; the thin layer on the screen is not the only thing that deserves to be called a product.
11. Podcasts can grow in the short-video era because audio can still run in parallel with life
Sound is not suited to receiving information at the highest density, but it may be the only sense that supports multitasking: while listening, people can still look at things or handle other tasks. From the voice Q&A products 在行 and 分答 to 罗辑思维’s audiobooks and Ximalaya’s growth, 郑庆生 sensed early on that listening had potential, though the product form took a long time to converge on clear PMF.
Podcasts eventually captured the behavior through companionship, allowing content to run for 2, 3 or even 4 hours against intuition. One explanation he offers for their recent growth is that vision has been filled too densely by text-image feeds and short video. New information demand will “spill out like water,” seeking an audio gateway that is not yet saturated.
12. The To C lull was not a blank period, but the stage when cloud, SaaS and AI accumulated strength
From 2018–2019 until ChatGPT, the consumer side lacked major innovation. But today’s AI applications all run on the cloud, many use SaaS-style delivery commercially, and core technologies were developed successively during those years. 郑庆生 therefore prefers to call the period a “dormant” or strength-gathering phase rather than the end of a cycle.
He extends the timeline further back: before PCs came television, and before that radio; after electricity came power grids, while earlier information traveled along roads, railways and canals. This yields a general principle: technological revolutions change how people connect or move, and only after traffic converges do cities, attention economies and commercial forms emerge.
Before AI appeared, he guessed that the next To C gateway might be some kind of glasses. Even capturing only 10%–20% of phone traffic would be enough to make them important. The market therefore surged when Google first demonstrated AR glasses. History ultimately answered from an unexpected direction: AI, not glasses, reached the new traffic gateway before it replaced the phone.
13. Cities were the earliest social products, and even steel can be understood as a traffic revolution
郑庆生’s analogy is: “For a very long time in human history, the biggest social product was the town.” A city’s population roughly equals its DAU, which is why towns formed at intersections of traffic such as roads, canals and ports.
Steel may seem unrelated to information traffic, but beyond laying railways, it enabled reinforced concrete and vertical cities. Earlier cities were mostly flat and could build only a few floors. Once buildings grew upward, urban DAU increased sharply, making supernodes capable of holding tens of millions of people possible.
The first type of node in the digital era came from moving offline behavior, information or transactions online ahead of others. The second came from two-sided network effects. The latter requires delicate design and gradual construction; once it crosses a scale threshold, even established giants cannot quickly replicate it. Great To C products are therefore like “huge towns,” even “markets capable of rivaling nations.”
14. ChatGPT reopened the To C window and exposed 3 structural differences in AI networks
Sequoia China began intensive AI research in mid-2022, but 郑庆生 says ChatGPT was what made him clearly conclude that a major wave of consumer applications was coming. To C performance was already good enough to rapidly propel a group of new applications. If model companies win, they could also become the next generation of gateways.
The first difference is marginal cost. Once grids and traditional internet infrastructure complete massive upfront investment, the cost of adding another user approaches zero. AI consumes tokens with every additional user, so subscription fees, free strategies, pricing and target users must all enter ongoing commercial calculations.
The second difference is responsibility for outcomes. Conventional apps generally provide a process; whether users deploy it well remains partly their problem. Efficiency-oriented AI goes “straight to the result,” requiring providers to prove that their technology can deliver a solution. Existing SaaS has already validated PMF. The next unknown is how AI solutions can bypass structured processes and turn raw information directly into To B outcomes.
The third difference is the black box. Steam engines and electricity had no comparable black box at the level of underlying principles. With AI, it is unclear how emergence occurs or what capabilities will appear next. Prompting guides even recommend adding certain trigger words to improve results, giving the process an element of “art” and “mysticism.” These uncertainties may themselves become part of AI-native products.
15. Deep digitization will turn dedicated hardware into new information gateways
Earlier digitization mainly moved content that was already structured and easy to count on paper onto the internet. AI can process unstructured materials that were previously unusable. One hundred hours of personal recordings could not be organized, and a Microsoft Research-style automatic camera could shoot tens of thousands of photos a month without anyone being able to process them. Collecting them therefore had little practical meaning.
AI can now turn those materials into meaningful online content. 郑庆生 therefore believes AI is driving another, deeper round of digitization. The amount of personal information people possess could increase by hundreds, thousands or even tens of thousands of times, while devices that continuously capture offline increments could become new data nodes and traffic nodes.
Phones can already record and take photos; the issue is that they are not designed for continuous sensing. The key to dedicated devices is that they are “always present.” Users do not need to keep checking whether recording is still running. A device worn on the body or operating continuously in another form is more likely to create stable habits and trust.
If large amounts of data, habits and usage patterns accumulate in a single device, they may form a moat. But privacy has no ready-made answer. Whether a public speech is implicitly open to recording, and how to define a sentence spoken in public that everyone nearby can hear, will still require adjustment through law, social habits and public attitudes.
16. Foundation models have an implicit flywheel, but AI social products have not formed explicit two-sided networks
郑庆生 believes foundation models contain a kind of implicit network effect. After using large amounts of data and speaking with more people, a model may serve new users better. Unlike a traditional platform that matches buyers and sellers one by one, this can still form a data flywheel.
The explicit difficulty is that current interaction is mainly every user facing the same model, a “many-to-one” structure that does not require one person to publish and another to receive. The first wave of socially oriented AI companies in the United States has not shown strong network effects. If a model deteriorates or a competing model improves, its commercial position can deteriorate rapidly.
On whether OpenAI’s group chats could challenge Meta, 郑庆生 sees a remaining gap in the logic: people today know that the other party may simply be a model-generated computation, so the psychological experience is different. But users may eventually stop caring whether the other party has independent consciousness or a soul. At that point, the boundary between human-human and human-AI networks could be rewritten.
The unresolved question is whether AI should be inserted into traditional two-sided networks, or whether the model’s implicit effect should simply be made strong enough. If an explicit network never forms, “the world as a whole may become the world of models.” He explicitly leaves this in the realm of the unpredictable.
17. This commercialization cycle started from a healthier position, but revenue still does not cover infrastructure intensity
郑庆生 believes current AI products already have data flywheels and that commercialization is more mature than in the early internet and mobile-internet cycles. Years of SaaS education have made the market understand subscriptions, and many products have revenue from the start rather than first expanding users indefinitely.
The constraint is that technology is still iterating rapidly, infrastructure investment is enormous, and revenue remains mismatched with costs. A sustainable commercial model can become more stable only once humanity reaches some satisfactory boundary. He remains optimistic, but does not equate current revenue with final unit economics already being proven.
“Cash flow is very good” at early projects is a major reason he accepts the current valuation heat. Compared with burning money first, capturing mindshare and waiting years for monetization, AI can hit pain points and charge from the beginning, which is more comfortable for investors. Whether model changes will destroy existing PMF still requires continuous validation.
18. The boundary between model companies and application companies is blurring; the real dividing line is how much product advantage technology can erase
The core metric for a model company remains how its technology evolves. Application companies face more variables because they may build a good product on the current model, only for the next round of underlying technology to rewrite the capability boundary. 郑庆生 sees this as additional uncertainty created by “technology and products evolving simultaneously.”
Model companies moving downward into applications and application companies training their own models are both entirely possible in his view. The ultimate competition will not be decided by labeling a company “model” or “application,” but by answering: “To what extent will technological progress offset nontechnical product advantages?”
OpenAI already has strong product-company characteristics: large numbers of users, data and accumulated habits of opening and talking to the product. At some point, continued technological progress may no longer solve the problem. That may be the boundary a product company can defend, but the boundary is not yet visible.
19. New giants will still emerge, but entrepreneurship has turned from a binary equation into a five- or six-variable equation
郑庆生’s core belief in early-stage investing is that every cycle adds new names to the list of giants. The uncertainty is whether there will be more than 5, fewer than 5 or fewer than 3. Even when old giants appear to control everything, new gateways still push some companies off the list and bring others in.
Compared with 10 years ago, teams must not only build product technology but also understand models, decide whether to train one themselves, and handle online-offline integration and software-hardware integration. What used to resemble solving a two- or three-variable linear equation may now be a “five-variable or six-variable equation,” with greater startup difficulty and more variables in investment judgment.
His intuition is that moving from relatively simple digitization to intelligentization after full digitization may reduce the number of people capable of starting companies. The counterforce is the existence of foundation models: many companies serving small, highly vertical markets may become profitable from day one, making the market more like a field of many competing flowers.
Early investors will not therefore demand profitability immediately. Long-term scenarios remain more important. Whether value sits in the model or the application, and whether a platform is replaced or continues to grow, will have no universal answer. Tencent retained its position by moving from PC to mobile, while other old platforms were replaced by new ecosystems.
20. Sequoia’s foundation-model strategy came from long-term, case-by-case tracking rather than one-time sector coverage
郑庆生 clarifies that Sequoia China has valued early-stage investing since 2005 and did not only recently move from late-stage to early-stage. The characteristic became more pronounced after it established a seed fund in 2018. Projects where Sequoia is the first institutional investor now account for more than half of its total projects, and the firm continues to track the metric internally.
The principle in the AI era remains “invest small, invest early, invest in technology,” finding team and project strengths one case at a time. For technology-oriented companies, founders’ papers, citations and other background matter more than in the previous 2 cycles, but the underlying aesthetic has not fundamentally changed. “Invest in people” remains central to early-stage judgment.
Entering Kimi was partly related to 循环智能, another portfolio company he had previously handled, and developed step by step from there. MiniMax was on Sequoia’s radar before the ChatGPT and GPT-3.5 wave. 智谱 was also contacted early. These projects were not cases of filling the entire sector with capital after spotting a hot trend.
Manus was likewise tracked from the earlier Monica period. Sequoia already liked the team, product idea and subsequent evolution path, after which the product evolved into Manus. 郑庆生 stresses that what they saw first was the product capability displayed by the team and product, not a slot in a particular sector.
21. AI products are naturally global, opening the core bet of the “Age of Chinese Global Exploration”
Traditional apps must solve language, habits, local operations and implementation market by market. This wave of AI products is naturally suited to cross-market expansion. Whether efficiency-oriented or entertainment-oriented To C products, Chinese and overseas Chinese founders can launch globally from the moment the product is born.
郑庆生 summarizes the key bet today as “the beginning of the Age of Chinese Global Exploration.” Chinese companies and Chinese founders around the world can use AI to enter global markets, comparable to the worldwide expansion of European explorers and innovators or the process by which multinational companies built global businesses.
The AI ecosystem may ultimately have as many layers as the app ecosystem. Some old platforms will be replaced, while others will become larger through new technology. The game for early-stage investors remains the reshuffling of ecosystem positions; this time, competition has expanded from local applications to a naturally global market.
This also changes the geography of talent. Overseas Chinese once often returned to China to start companies serving the domestic market. Now, wherever they are, they can build global products directly. 郑庆生 sees this as a major structural difference between AI entrepreneurship and the earlier PC and mobile-internet cycles.
22. Investment has moved from models and infrastructure toward Agents, hardware and a more dispersed startup map
Looking at the sequence after ChatGPT emerged, 郑庆生 first systematically scanned foundation models, infrastructure and vertical-model opportunities. As the base layer gradually matured, he began searching earlier for Agent applications. While scanning applications, he also saw AI hardware developing rapidly and raised the priority of software-hardware integration projects in parallel.
His cycle analogy is that “2025 may be a bit like 2010, and 2026 like 2011.” Once a mobile gateway formed, applications appeared densely over the following years. AI’s 2026 will likewise deepen 2025, with products and the ecosystem becoming “more prosperous.”
Startup geography is more dispersed than in the PC and mobile eras. Beijing, Shanghai and Hangzhou remain more software-oriented, while Shenzhen is rising rapidly as a hardware base. Asked for his top 5 travel destinations, he mainly named Beijing, Shanghai, Shenzhen and Hangzhou, emphasizing Shenzhen’s acceleration in consumer electronics and AI hardware.
Total workload may not increase dramatically, but finding projects now requires visiting more places. Add overseas Chinese building global companies, and the work feels substantially busier. The character of entrepreneurs in each city is shaped mainly by industrial structure, though he believes truly excellent founders remain broadly similar.
23. Bubbles supply liquidity to technology cycles; AI is now at the “front end of the explosion”
郑庆生 treats “whether there is a bubble” as a talking point rather than the endpoint of decision-making. Every cycle has bubbles at the beginning. Bubbles provide enough liquidity for early founders’ grand ambitions to receive capital and talent support. “It is like foam in the ocean.” The key question is whether the company can deliver later.
He believes AI is not a tulip-style pseudo-concept but a huge technological revolution in human history. This cycle has produced PMF, revenue and decent cash flow from an early stage, so the current bubble level is acceptable. Strictly speaking, continuing model changes can still move the so-called PMF; early charging cannot be treated as permanent validation.
The cycle is not on the “eve of an explosion,” but at the “front end of the explosion.” Competition has begun across the main battlefields, but applications have not yet exploded at scale. His response is excitement rather than anxiety: entering the industry in 2005 gave him the chance to experience 3 consecutive technology waves—PC, mobile internet and AI—from the front line.
24. Whether a product manager can become a large-company CEO depends on whether they can personify the entire organization
From 1 to 10, 郑庆生 believes a founder must first be sensitive to products and user needs. Once the company reaches dozens or more, the leader becomes more like a general leading troops and must become “the personification of the organization and its institutions,” sustaining morale, discipline and the organization’s belief in a common mission.
This does not require such a nature by birth, but once in the position, the person must “play that CEO.” Ten thousand or 200,000 people can still believe in the desert that they are one army, but they need a personified representative. The larger the company, the more the founder must become the embodiment of that abstract living organism. Otherwise, even a brilliant product manager may fail to make the transition to CEO.
Possessing both talents is rare: product sensitivity and the ability to represent an organization, establish plans and build confidence. MBTI is only an auxiliary language for understanding people. N can help with abstraction; NF may suit product empathy; NTJ combines intuition, logic and planning; P may be more romantic; J may be better suited to organization. E or I may not determine success if the person has the right deputy.
He sometimes asks founders about their favorite books, how many apps they have on their phones or their MBTI type, but explicitly says MBTI will not become an investment criterion. He considers himself open-minded about people and does not want to attach a permanent label to an entrepreneur after a single meeting.
25. The real decision method is to reconstruct the choice set at the time and leave room for what humans become after AI
The first-principles knowledge 郑庆生 values most is survivor bias. History is written by winners or other survivors; survival is a necessary condition for a text to exist, but it “provides no predictive power.” Later narratives also dramatize the process, causing people to mistake the traits of successful people for sufficient conditions.
Investing requires reconstructing the decision scene. At the time there may have been candidates A, B, C, D and E, each backed by extensive information. Four months later, the answer may look obvious, but it was not known then. Founders, peers and historical cases all provide biased evidence of necessity. The real difficulty is finding the basis that was sufficient to support the choice at the time. He chose not to elaborate on the details of specific model-investment decisions.
The 2 books he recommends most serve this kind of long-cycle reconstruction. 《美国增长的起落》 describes America’s technological revolutions before the internet, and he believes it is worth spending a year reading. 《从中国出发的全球史》, written by Fudan’s history department, reconstructs global history in a new way and is worth 2 years; it also expresses many ideas he had originally hoped to develop in future articles.
The biggest unknown remains “what humans will become after these AI products arrive.” Education may place more weight on a global view, taxonomy, logic and precise expression, because learning also includes learning how to ask questions; detailed knowledge may matter relatively less. If natural language becomes part of coding, precise expression will also appreciate in value. Today’s products are designed for humans before intelligentization. Once education, consumption, intimacy and concepts of consciousness change, the products themselves will have to be redesigned.