Sequoia’s 郑庆生 on Traffic, Human Behavior, and Founders
Summary
郑庆生 uses “traffic” to rewrite economic history: roads, canals, railways, telegraphs, radio, television, PCs, mobile phones and AI all change how people, information and attention move. Towns were once the biggest social products—“a city had as many DAUs as people”; today’s best To C platforms are “markets capable of rivaling nation-states.” The core investment question is what the next technology will turn into the next super-node.
AI is not a simple sequel to the mobile internet because it changes cost structures, delivery models and the explainability of technology at the same time. Every additional user generates Token costs, so the network’s marginal cost no longer approaches zero; efficiency products are shifting from walking users through a process to taking them “straight to the result.” More unusually, emergence remains a black box, while Prompting carries an element of “art and metaphysics”—meaning technology, products and business models must be considered together.
Beyond foundation models, the next generation of traffic gateways may come from integrated hardware-software devices that continuously capture the physical world. In the past, even if a hundred hours of audio or tens of thousands of automatically captured photos were digitized, they could not be organized; AI is making it possible for the “information that used to scatter in the wind” to be processed for the first time. If the amount of personal information owned by each person expands by hundreds, thousands or even tens of thousands of times, hardware may build moats through data, usage habits and new ways to play—a shift 郑庆生 calls “deeper digitization.”
The credibility of precise predictions about AI product forms is very low because humans will co-create behaviors with products that did not previously exist. Short video began as a minor video category but ultimately “challenged text itself”; Twitter’s 140-character limit and Musical.ly’s early usage patterns also exceeded what early users could have extrapolated. Investors can preserve product intuition and, once new forms emerge, identify whether they have the potential to become “classics.”
Value allocation between models and applications has not converged, but AI’s commercialization starting point is better than that of the previous 2 internet cycles. Subscription models have been normalized by SaaS, most products generate revenue relatively early and can benchmark ARPU, and early-stage cash flows are healthier than the “burn for scale first, monetize later” model; the cost is that rapid model upgrades may flatten application advantages. Conversely, user habits, accumulated data and non-technical product capabilities may still draw boundaries that technology cannot penetrate.
The biggest structural change with Agents is that they are “global by birth,” giving Chinese and overseas Chinese founders their first chance to serve the entire world directly. 郑庆生 likens 2025 to 2010 in the mobile internet and 2026 to 2011: after the underlying models, applications, Agents and intelligent hardware are entering a more prosperous competitive phase. His core bet today is “the beginning of the great overseas expansion of Chinese founders,” but whether incumbent platforms will be replaced, enhanced or continue expanding still requires an ecosystem-by-ecosystem judgment.
郑庆生 does not deny that AI has a bubble, but believes the bubble itself provides liquidity for early innovation; the only question is whether products can ultimately be delivered. AI can already hit pain points and charge for them, making it different from speculation built on pseudo-concepts; the current moment looks more like the front end of an application boom than a countdown to a burst. “Bubbles exist in the ocean”—rising and falling tides are acceptable as long as products backed by real money and real revenue remain.
For founders moving from product manager to large-company CEO, the job is to move beyond understanding users and become the organization’s “personified symbol.” From 1 to 10, product sensitivity matters more; beyond 10, leaders must maintain institutions, morale and a shared imagination like generals—“as long as you play that CEO, regardless of what your own nature is.” MBTI can help start a conversation but cannot be an investment criterion; the more reliable discipline is to resist survivorship bias and reconstruct all available options and sufficient conditions at the time decisions were made.
Deep dive
1. 郑庆生’s Career Has Always Tracked China’s Digital Wave
郑庆生 first learned to program in 1984. Just after starting elementary school, he copied BASIC programs from a book onto his father’s Sharp computer and even built a keyword-matching “human-machine conversation” system. It was not truly intelligent, but he already experienced the pleasure of making a machine respond to a person.
In 1996, he attended an internet lecture at Fudan, entered the campus BBS and served as moderator of the economics forum. Around 1998, a classmate recommended Google as “better than Yahoo.” He later placed these milestones against the backdrop of the era before and after China’s accession to the WTO, seeing them as a lasting imprint on his education and career.
At graduation in 1999, internet jobs in China were still scarce, and finance majors generally favored traditional paths such as audit and banking. He chose auditing at a foreign company, but “tried almost every internet product,” and says he may have used more of them than many people working in the internet industry.
The real turning point came in 2005, when a classmate at Shanda told him that Shanda’s revenue exceeded that of all other internet companies combined. He suddenly realized that he had already spent 6 years browsing the internet in his spare time: “Why wasn’t I working for an internet company?”
2. 6 Years in Audit and Consulting Turned Intuition into a Verifiable Loop
3 years of audit training filled in the numerical and evidentiary discipline missing from 郑庆生’s liberal-arts background: numbers had to be accurate to 2 decimal places, evidence had to be cross-indexed, and every cross-reference ultimately had to close the loop. Without that training, he believes investing would have been “much harder.”
The global manufacturing shift triggered by WTO accession brought him into factory compliance inspections that were not yet mainstream. He traveled across Jiangsu and Zhejiang, Shandong, Shanghai and the areas around Beijing, examining workshops, production lines and employee areas at apparel, toy and electronics factories one by one. Nearly a decade later, those memories from the field became useful again when he reviewed materials and contract-manufacturing projects.
He then spent 3 years in management consulting focused on process implementation. Once standard products such as SAP, Oracle, PeopleSoft and budgeting software entered China, consultants had to define client processes and match them to software. When he looked at SaaS in 2018–2019, he realized that this work resembled the “customized implementation” of later years.
张小珺 pointed out the contrast: his résumé was intensely rational and process-driven, while he himself seemed more intuitive. 郑庆生’s explanation was that “2 lines” ran in parallel—on one side, the most demanding professional delivery; on the other, serving as a tester for almost every new internet product.
3. Shanda Turned Investing from a Career Choice into a Gateway to Understanding Companies
By 2005, Shanda was no longer just a games company. It had incubated content businesses such as Qidian and was exploring dynamic passwords, internet-café games and multiplayer products. 郑庆生 sees those years as the “flourishing of a hundred flowers” after the first wave of internet infrastructure was complete: e-commerce, games, QQ and a large number of applications began competing for the traffic left behind by the portal era.
In Shanda’s investment department, he entered a real operating company for the first time. He saw how departments collaborated, how products advanced, how business development was executed and how cross-regional operations connected. After spending much of his career in professional and investment institutions, this became his only internal sample for understanding a large internet organization.
In 2007, he and a group of Shanda colleagues followed Shanda’s former CFO into financial investing. Looking back at 2005–2010, he believes the best investments may actually have been the stocks of the giants that later went public: “Nobody expected the digital wave that followed to be this big.”
4. Web 2.0 Was the First Time the World Returned Things You Liked but Did Not Know to Look For
Delicious showed 郑庆生 that user ratings could surface niche websites that neither Yahoo’s directory nor Google’s active search could find. Search requires users to know in advance what they are looking for; UGC and reviews can feed unknown interests back to users. It was a “change in the form of cognition.”
When Douban launched in 2005, he wrote an internal report at Shanda calling its format highly novel; Dianping also entered his field of vision around the same time. The 2 companies later became investment relationships for his fund and Sequoia respectively, and important milestones in his decision to pursue investing for the long term.
In 2009, he had his first long conversation with 阿北 at Carving Time in Wangjing, Beijing. He felt that “the person and the product were completely one,” with Douban resembling a projection of its founder’s inner world. When he met 张涛 in 2011, Dianping already had a structured team, and 张涛 projected a more mature temperament combining product insight with an entrepreneurial style.
His historical positioning is that Douban and Dianping were the culmination of Web 2.0 around 2009, completing online content innovation after the first digitization of human society. Shared-economy and bike-sharing models later pushed innovation toward large-scale online-offline integration.
5. 20 Years of Internet Investing Can Be Read as 4 Rounds of Traffic Redistribution
Before 2005, Sina, Sohu, NetEase and other portals completed the first migration of information online and captured traffic; Tencent controlled the information gateway, Baidu controlled search, and Shanda controlled games and content. By today’s standards, those eventual giants were then merely “unicorns or innovators worth several hundred million dollars.”
From 2005 to 2010, the main traffic high grounds of the PC internet were gradually captured by the giants. Community and application innovation continued, but it became difficult to take back the main gateways. 郑庆生 emphasizes that a fixed structure does not mean innovation has stopped; it means innovators struggle to capture the main traffic.
Around 2010, smartphones opened entirely new gateways. Old gateways surrendered market share and the mobile internet redistributed traffic; around 2015, short video became the dominant content format. At the same time, Meituan, Didi and Mobike continued connecting online demand with offline fulfillment through O2O.
By around 2018, the new round of mobile traffic had probably been divided up again, and To C investing fell quiet. Only later did 郑庆生 understand that this quiet period was “a time of dormancy or gathering strength,” not the permanent end of consumer innovation.
6. Pinterest Moved Mobile Content from Web Layouts to the Feed Paradigm
郑庆生 remembers Pinterest appearing around 2010, while preserving the time qualifier “if I remember correctly.” His judgment of its significance is much stronger: the waterfall feed suited to mobile was a paradigm revolution, not simply another image website.
He then searched intensively in China for Pinterest-like products and had deep interactions with Xiaohongshu, Mogujie and others. In his personal lineage, this organizational form may have influenced later content platforms including Douyin, TikTok, Xiaohongshu and Kuaishou, and may even have been the “originator” of the format.
Pinterest’s specific form did not rule the market unchanged, but mixed image-and-text layouts did not disappear. When browsing Xiaohongshu today, images and text remain better than pure video for guides, questions and lists. What was inherited was the information-organization paradigm, not a single media format.
7. Short Video’s Ultimate Competition Was Not Long Video but Text
郑庆生 first defines text as a knowledge product with a high barrier: humans need years of learning to master abstract symbols, which is why countries had to pursue mass literacy. Images are more intuitive, so image-and-text formats naturally cover some purely textual use cases; long video extends theater, film and television, content forms humans already know well.
He once favored long video because its length seemed to support more business models, and initially treated short video as a shorter version of long video. He later changed his view: “Short video is fundamentally how humans understand the world,” and “should not be compared with long video.”
His strongest analogy is the moon. Telling someone, “The moon is especially big and round tonight,” in writing—or even poetry—is sophisticated; the simplest approach is to take the person outside and say, “Look.” Shooting a short video and taking someone to the scene offer nearly comparable information richness.
That is why he says short video “challenges text itself.” Not reading for a year is not shameful, provided other media deliver the same information density and volume, rather than merely serving as time-killers.
8. Efficiency Products Need the Result; Entertainment Products Need the Process
Extending the short-video discussion to AI, 郑庆生 offers a concise distinction: “A tool product is the result; an entertainment product is the process”—saving time versus consuming it. Entertainment relies on the process to occupy time, while efficiency tools should minimize the user’s involvement.
This explains why AI is not merely an interface upgrade. Users no longer want to click through a task step by step; they want the result directly. If the process itself has no experiential value, the model will compress it, and product evaluation will shift from “is it easy to use?” to “how good is the result?”
张小珺 summarized this as content platforms moving from active to passive consumption and toward greater fragmentation. 郑庆生 adopted “from active to passive” and “consumption becoming increasingly fragmented” as his own judgments, while warning that these changes were not a script founders wrote in advance. They emerged through interaction among product rules, creators and users.
9. Developers and Users Co-Invent a New Product’s Real Use
When 郑庆生 first used Twitter in 2005 or 2006, he did not understand why it had 140 characters or what he was supposed to write. He could only record books he had read and places he had visited, then concluded: “This is so boring. Who would read this?”
Users later began posting opinions and frequently @-mentioning others, and the product’s form gradually took shape. His conclusion was: “When a product form takes off, it is actually created together by the product’s developers and users.”
When he used Musical.ly around 2015, he only knew how to film scenery with music and had no idea that dancing and similar content would later drive growth. This reinforced his central conclusion: “New human behavior patterns are, in aggregate, unpredictable.”
Founders may not be able to provide a rational forecast of the ceiling for growth, but they may possess intuition and even elevate it into conviction. Winners later narrate their beliefs; losers disappear from history. Any story that a founder “saw everything clearly from the beginning” must therefore be tested against survivorship bias.
10. Douyin, Xiaohongshu and Bilibili Flourished through Different Information Orders
Judging by user experience, Douyin depends more on the content itself. When a KOL with 1M followers posts, those followers probably form the base of clicks and likes. On Xiaohongshu, comments matter alongside the original post; further distribution depends more on the quality of each individual post, and 1M followers do not automatically provide a base.
郑庆生 calls Xiaohongshu the “most open product structure” of the mobile internet era: images, a paragraph, guides, questions and comments accommodate almost every major form previously seen in UGC and social media. It has therefore become “a friend of time,” with users collectively growing it across female, male, online, offline and even City Walk scenarios.
Bilibili took another route: entering through an influential niche community before expanding to broader audiences and high-quality content such as documentaries. It more closely resembles a projection of the founder and core community’s temperament than an open structure designed for everyone from day one.
After Toutiao, ByteDance already had a recommendation engine and strategic altitude. Entering short video was more like “taking the killer weapon of a recommendation engine and finding a better-matched format.” 郑庆生 does not rank this path against founder-as-product models; he only believes success is difficult to reduce to a single formula.
11. The Most Dangerous Investing Mistake Is Treating Your User Circle as the Entire Market
During the Web 2.0 era, 郑庆生 gradually developed product taste and once defined his direction as “investing in advanced lifestyles for advanced people”: reach knowledge-oriented, first-tier-city users first, then spread to the broader population. Douban, Dianping, Xiaohongshu and Musical.ly all fit this path to some extent.
Unfunded opportunities such as Didi and Pinduoduo forced him to reflect. UGC requires users capable of creating content, making it inherently more knowledge-oriented; he lacked sufficient understanding of the offline world and China’s most basic and widespread users. Some basic models can spread directly without evolving from “advanced users.”
Bike-sharing and the shared economy became a deliberate correction. Mobike, shared power banks and shared spaces showed that the shared economy did change life for China’s grassroots population, although the logic encountered real-world ceilings as it developed into different forms.
His “methodology” for AI products is therefore not to predict a specific form, but to maintain enough experience and discrimination: like a literary critic who cannot write a novel but can still recognize that a new work is probably a classic and worth studying further.
12. A Product Is Not a Layer of Interface but the Sum of Technology, Operations and Supply Chain
During the PC and mobile internet eras, foundational technology was relatively mature, so the investment case for application companies naturally came more from the product. In the AI era, technology and applications are advancing together. 郑庆生 therefore rejects the idea that he only favors product companies; different cycles determine which variables deserve the most attention.
He also revised his belief that only online products count as products. How an offline operations team is organized, managed and fulfills its obligations is itself a product. Going online first solves whether something exists; rule innovation then determines how it is used; later, human intervention and heavier operations become necessary.
Klook is an example. It aggregates traffic through online UGC, reviews and purchase demand, then uses large procurement volumes to control the supply of attraction tickets, metro tickets and distinctive activities across Asia. The real moat thickens only when the online product and offline supply chain deepen together.
张小珺 noted that Douyin’s success also depended on operations. 郑庆生 agreed: the deeper digitization goes, the less entrepreneurship is about “building a thin product” and the more it is about combining content, organization, supply and technology into a complete system.
13. Podcasts Reclaimed Time Crowded Out by Visual Interfaces through Companionship
Audio is not suited to the highest-density information intake, but it is almost the only sensory channel that can be used in parallel: people can listen while looking at something or doing their own tasks. Years ago, when investing in Zaihang, Fenda and products such as audiobooks and Ximalaya, 郑庆生 already felt that “listening” was worth studying.
The counterintuitive feature of podcasts is that, after short video trained fragmented habits, content can still run for 2, 3 or even 4 hours. The reason is not that listeners suddenly recovered long attention spans, but that audio can “exist in the form of companionship.”
On why podcasts rose so visibly only in recent years, he offers a qualified judgment: the visual field has been filled too completely, and “all innovation will spill out like water.” Once visual input is saturated, people will still seek additional channels for information.
14. The To C Lull after 2018–2019 Was Building the Foundations for AI
After short video divided up mobile traffic, the consumer internet saw few major innovations from around 2018 or 2019 until the rise of AI. 郑庆生 focused more on To B opportunities such as SaaS, while global investors also poured substantial energy into cloud services and SaaS.
In retrospect, today’s AI runs on the cloud, its business models are highly SaaS-like, and key technologies were developed incrementally during those years. Once old traffic was fixed in place, product and technology innovation entered a latent phase, just as Web 2.0 was already forming inside the portal era.
This experience made him stop treating technology cycles as unrelated stories. If PC, mobile, radio and television represented only terminal changes, they could not explain the support for the next cycle. He therefore pulled the timeline back to telegraphs, railways, canals and cities.
15. “Traffic” Connects Transport, Media and Internet History on One Line
郑庆生 chooses traffic as the central support of economic history: “traffic, movement or connection”—human history can all be understood as the process by which people, information and attention continually change how they move.
Before electricity, information traveled along roads, railways, canals and ports. After electricity, telegraphs, radio, television, PCs and the mobile internet progressively detached information from physical distance. Every technological revolution ultimately changed the way traffic moved, producing new commercial prosperity where it converged.
Attention economics and urban commerce are not fundamentally different within this framework. Whether people live in a place or attention rests there, traffic creates transaction and allocation rights. Investing means finding which position new technology will turn into the next point of convergence.
16. Cities Were the First Social Products; Steel Pushed Their DAUs into the Tens of Millions
郑庆生’s most striking analogy is that “for a very long time in human history, the largest social product was the town.” Cities grew where roads, canals or ports intersected: “A city had as many DAUs as people.”
Steel appears unrelated to traffic, but transformed traffic capacity through railways and reinforced concrete. Without steel, most cities could only spread horizontally; once cities became three-dimensional, tens of millions of DAUs could gather in limited space. The materials revolution ultimately pointed back to node density.
Today’s excellent internet To C products are likewise “massive towns,” even “markets capable of rivaling nation-states.” Any platform with 2-sided network effects is a traffic node, but super-nodes do not all need to display obvious 2-sided network effects.
Nodes commonly form in 2 ways: digitizing an offline behavior, information flow or transaction ahead of others, or gradually building 2-sided network effects through elegant product design. Once a scale threshold is crossed, incumbent giants find it difficult to copy and catch up immediately.
17. ChatGPT Suddenly Turned AI from Technical Research into a New To C Gateway
Sequoia China began intensive AI research around the middle of 2022, but 郑庆生 admits that only after ChatGPT appeared did he clearly feel a new application wave would be rapidly mobilized, because the To C applications were “so good.”
Under the traffic framework, the winning foundation-model companies will certainly form the next generation of gateways. What remains unclear is whether AI-native 2-sided networks, image communities or high-quality content platforms can emerge, allowing value to extend beyond the model layer.
张小珺 asked whether AI had made him excited again. 郑庆生 said he was equally passionate about To B and SaaS, but To C was a more durable interest. After 2 rounds of traffic gateways, seeing a third reshuffling was itself a rare personal historical experience.
18. AI Networks Simultaneously Rewrite Marginal Cost, Product Responsibility and the Epistemology of Technology
The first difference is marginal cost. For a traditional power grid or the internet, the cost of adding a user is nearly zero relative to the enormous upfront investment; with AI, every call consumes Tokens—“each additional one is a sum of money.” Subscription models, free strategies and user selection therefore need to be recalculated.
The second difference is the result orientation. Traditional Apps provide services and operating paths, and poor outcomes can sometimes be attributed to the user. Efficiency-oriented AI produces an answer directly from raw information, so providers bear greater responsibility for result quality and may restructure the processes and structured information on which traditional SaaS depends.
The third difference is the black box. The basic principles of technological revolutions such as steam engines and electricity are explainable, but it remains unclear why emergence occurs in AI or how capabilities change. Prompting textbooks even recommend adding certain “magic words” to improve output, making use “an incompletely rational process.”
郑庆生 therefore believes there is still no answer as to whether the artistry, metaphysics and unpredictable emergence in applications are native AI characteristics. Cost can be handled through business design, but responsibility for results and the black box will directly change product forms and industry boundaries.
19. AI Turns Information That Could Not Be Organized into Operable Data Assets
Earlier digitization mainly moved data that already existed on paper and was easy to count online, then structured it further. The “deeper digitization” triggered by AI begins processing unstructured information that could not be effectively used even after being captured.
In the past, a person recording 100 hours of audio could not organize it; a chest-mounted camera generating tens of thousands of photos in a month could not be reviewed. 郑庆生 mentioned that Microsoft Research had once used changes in light to trigger automatic photography: the technology could collect material but could not turn it into meaningful online content.
AI can process these materials into meaningful online content, allowing products based on all-day recording and automatic photography to command high valuations and perform well. Conversations, scenes and behaviors that once dispersed into the air may become meaningful online content for the first time.
If deep digitization is fully realized, the amount of information owned by each person could “increase by hundreds, thousands or even tens of thousands of times.” Add models iterated over several generations, and the outcome remains impossible to predict—but that is precisely the opportunity space for integrated hardware and software.
20. Dedicated Hardware’s Value Lies in “Continuous Presence”; the Cost Is Rewriting Privacy Boundaries
A phone can also record and transcribe, but it was not designed for continuous capture; users are unsure whether it is still recording. Dedicated wearable, attached or carried devices provide the certainty that “it is always there,” making them better suited to deep digitization.
These devices enter the offline world and capture information that was previously “not that important, completely unimportant or destined to scatter in the wind.” Large datasets, usage habits and subsequent ways to use the product may accumulate into a hardware moat, turning hardware into a new information and traffic node beyond the model.
张小珺 asked about privacy: does a public speech imply permission to record, and what about saying something in public that everyone nearby can hear? 郑庆生 offered no simple answer, saying only that laws, social norms and collective attitudes may all need to change.
His judgment is conditional: behavioral norms will change with technology as they did in the short-video era, but no one knows how far the change will go. The hardware opportunity and privacy friction are not separate issues; they are 2 sides of the same expansion of information.
21. Foundation Models Have Implicit Network Effects, but AI Social Products Have Not Built an Explicit 2-Sided Market
Foundation models have implicit network effects because they absorb more data, converse with more users and receive more feedback. Traditional platforms, however, require “someone to listen when you speak and someone to buy when you sell,” with both sides of supply and demand assembled incrementally. That kind of explicit network has not yet stabilized in AI products.
Today, people and models look more like many-to-one: everyone chats separately with the same model and does not need one another. 郑庆生 uses this to explain why the first wave of social-leaning AI companies in the United States failed to build strong moats. A decline in model quality or a rise in a competing model could quickly change their commercial position.
It remains unanswered whether AI should plug into existing 2-sided networks or form new networks composed of people and Agents. If a 2-sided network cannot be built, “the world as a whole may become a world of models.” But humans may eventually stop caring whether the entity they are chatting with has autonomous consciousness or a soul.
That shift in cognition is the larger unknown: if users know the other party is virtually generated, will the emotion still be real? The apparent logical gap today may be crossed by the default assumptions of the next generation.
22. The Boundary between Models and Applications Will Blur, but Non-Technical Advantages Will Not Automatically Go to Zero
郑庆生 believes AI’s commercialization starting point is better than the early internet and mobile internet. SaaS has already educated the market to accept subscriptions, products generally generate revenue earlier and teams directly compare ARPU. Current revenue still does not match the enormous infrastructure investment, but once a product reaches a certain user-satisfaction threshold, he remains optimistic about the commercial loop.
Model companies must first prove that their technology can continue to evolve. Application companies face more complicated uncertainty: products built on today’s models may be rewritten by the next capability upgrade. Technology and products advancing together turns entrepreneurship from a “2-variable linear equation” into a “5- or 6-variable equation.”
Model companies can build applications, and application companies can train models. The ultimate competitive question is: “To what extent can technological progress offset non-technical product advantages?” Some product moats can be penetrated by model advances; user relationships, operations and habits cannot necessarily be solved by a stronger model.
In his description, OpenAI already displays strong traits of a product company: large volumes of user data accumulate, and users habitually open the product and speak directly. A foundation model can be a product, and an application can possess technology. The real boundary will only emerge as technology develops.
23. Agents Are Global by Birth; Investment Is Moving from Models to Applications and Hardware
Unlike Apps, which require language conversion, local habits and local operations, Agents are naturally suited to global expansion because AI products adapt natively in areas such as language. 郑庆生 believes this is the first time Chinese and overseas Chinese founders can offer efficiency-oriented and entertainment-oriented To C products to the world from day one.
His sequence is to scan foundation models, infrastructure and vertical models first, then look for Agent applications and subsequently focus on rapidly developing intelligent hardware. Using the mobile internet as an analogy, he sees 2025 as 2010 and 2026 as 2011, meaning applications should become increasingly prosperous.
Startup geography is also shifting from a Beijing-centered model to multiple hubs across Beijing, Shanghai, Shenzhen, Hangzhou and overseas. Beijing, Shanghai and Hangzhou lean toward software, while Shenzhen is rising quickly on its hardware base. Overseas Chinese founders no longer need to return to China to serve the domestic market; they can address the whole world directly.
His core bet therefore lands on “the beginning of the great overseas expansion of Chinese founders.” Some old platforms will be replaced, while others will be improved by AI and continue expanding. The early investment conviction is that the list of giants will change; the unknown is whether 3, 5 or more new names will be added.
24. Sequoia’s Early Strategy Was Not Sector-Wide Coverage but Case-by-Case Searches for Irreplaceable Strengths
After Sequoia China established its seed fund in 2018, it further strengthened its focus on “small checks, early stages and technology,” while internally tracking whether it was the first institutional investor. 郑庆生 says such companies now account for more than half of all investments.
Responding to the criticism that Sequoia had “covered most Chinese foundation-model companies,” he denied using capital strength to cover an entire sector and emphasized a case-by-case approach. Kimi was connected to the earlier lineage of Moonshot AI’s Circular Intelligence; MiniMax had entered his field of view before the GPT boom; 智谱 was also a company he met very early.
Manus was followed when the team was still working on Monica. The investment case was the team’s and product’s capabilities, as well as the subsequent evolution they described—not waiting until Manus became a hot direction and then adding exposure. 郑庆生 summarizes the continuity of research and project evolution by saying, “There are no leaps in nature.”
Founder selection in the AI era is fundamentally unchanged, but technology now carries far more weight. Papers and citation counts were once supporting evidence of excellence; for technology-oriented companies, they may now be core judgments. Even the strongest era label, however, does not replace identifying each team’s product and capabilities one by one.
25. Bubbles Provide Innovation Liquidity; the Real Divide Is Whether Products Can Charge and Deliver
郑庆生 considers “Does AI have a bubble?” an easy talking point, because every technology cycle has a bubble at the beginning. Grand visions attract capital and talent, especially giving early-stage projects room to experiment. Without liquidity, many innovations that had not yet proved themselves would never start.
“Bubbles are like bubbles in the ocean”—entirely normal in themselves. The only question is whether products can ultimately be delivered. AI could hit user pain points and begin charging from an early stage, and many projects had healthy cash flow. That is at least more comfortable for investors than burning for user scale for a long time before looking for a business model.
He distinguishes a technology bubble from a pseudo-concept. Historically, some things were pseudo-concepts from the start; AI is clearly “a massive technological revolution in human history.” Valuations and capital may rise and fall with the tide, but the underlying value does not disappear.
On where the cycle stands, he rejects the phrase “the eve of the explosion,” because competition has already begun across all major fronts. A more accurate description is “the front end of an application boom.” That excites him rather than worries him: experiencing 3 cycles—PC, mobile internet and AI—from the front line is rare.
26. Great Founders Must Move from Product Talent to Organizational Persona
郑庆生 believes every strong entrepreneur should be product-sensitive. From 1 to 10, understanding demand and building a product are especially important. Beyond 10 or into the dozens, the leader’s task becomes organizational construction, and product sensitivity alone no longer explains differences in scale.
His central analogy is the ancient general: “You ultimately have to become the personified symbol of your organization and its institutions.” Before modern communications, a general was the highest representative of the institutions, discipline, morale and shared vision of 100,000 or 200,000 people. A large-company CEO performs a similar function.
Even if a person’s natural temperament is different, “as long as you play that CEO,” the role must be performed. A company is an abstract living entity created through collective imagination, and the CEO is where its personality lands. Someone who cannot make this transition may become a successful product manager but not necessarily a successful CEO.
Product talent and command talent appearing together is rare. A founder can be complemented by a team, but the larger the organization becomes, the less it can avoid this personifying role. Otherwise, the community lacks a credible spiritual representative.
27. MBTI Can Help Discuss People but Cannot Replace Investment Judgment
郑庆生 finds MBTI “quite interesting.” N, representing abstraction, may be an important underlying trait of a strong product manager; NF leans toward emotion and user experience, while NTJ combines intuition, logic and planning and may be better suited to organizational management.
He also preserves plenty of exceptions. A genius product manager may be INFP; the romanticism and sensitivity of NF and NFP may not suit the daily work of a CEO but can produce extremely strong products. S may be important for operations, while E or I can be partly offset by deputies and team structure.
When meeting founders, he occasionally asks about MBTI, favorite books or the number of Apps on their phone to open a discussion about the person. But when 张小珺 asked whether it could become an investment criterion, his answer was unequivocal: “No.”
28. Resisting Survivorship Bias Is the Only Way to Judge Effectively in an Unknown Era
郑庆生 treats survivorship bias as a first principle. History is written by winners or survivors, and the beliefs winners describe are necessary conditions, not predictive ones. Investing requires finding sufficient conditions by reconstructing the A, B, C, D and E options available at the time and the information behind each one.
He warns that humans naturally prefer drama. Historical texts, founders’ retrospectives and observers’ accounts all rewrite complex choices as inevitabilities. The genuinely difficult work is returning to the moment before the decision: “I knew the answer 4 months later, but at this moment I still didn’t know.”
2 recommended books extend this perspective. 《美国增长的起落》 traces technological revolutions before the internet, and he believes it is worth spending a year reading. Fudan’s 《从中国出发的全球史》 reconstructs global history differently, and he thinks it is worth spending 2 years on.
Looking toward AI, what he understands least is not any individual feature but what AI users themselves will become. If the core of learning becomes “learning how to ask questions,” broad perspective, cataloging and logic may matter more than detailed memorization. Once education, language and cognition change, they will in turn rewrite product demand.