Summary

锦秋基金’s size and duration follow a deliberate logic: $500M for a 12-year term was designed that way. A $200M fund can only make spot investments and lacks the capacity to add in later rounds; a $1B fund “would inevitably end up chasing big projects.” The 12-year term reflects the view that this AI wave is “an infrastructure-type variable,” not an industry cycle. With a seven-year fund, “probably 70-80% of AI software applications would be too risky to touch,” as would many compute-in-memory and embodied-AI projects.
The hardest contrarian call was to explicitly avoid foundation-model companies starting in 2023. Even if 智谱, Mini Max, and Kimi go public next year, there is no regret: “People often overlook one fact: even with 100M DAU, Doubao is still free today.” Competition is brutal, resource requirements are enormous, and the rivals are not just startups. OpenAI itself forecasts a $78B loss in 2028. “There was no need to pick a hard bone to chew,” when the risk-reward looked better in embodied AI, applications, and compute-in-memory.
The strategy boils down to “fast certainty.” Research narrow sectors upfront—compute means only compute-in-memory and optical computing; applications focus on AI+marketing, creative tools, and healthcare—then build a base of knowledge about “what works and where the good people are.” When the right person appears, move quickly with capital and clear feedback. Since mid-2024, the current fund has spent two years reviewing 4,000-5,000 companies and teams and invested in 51, roughly 1%; missing deals is acceptable as long as it “draws its own boundaries and stays disciplined about what to invest in and what not to invest in.”
Manus was the signature miss. There was an opportunity to invest in its earliest round, but “we spent our attention evaluating a plugin product.” What was missed was the Xiaohong team’s ability to move from Monica to Manus by “fighting disorder and remaining antifragile.” The lesson is that people still carry substantial weight in AI application investing. The traits rank resilience > learning ability > execution, and “know-how is not a moat; continuous, rapid learning is.” If any one of the three is weak, the investment is a pass.
Competition surged in the first half of 2025. Sequoia and Hillhouse were “looking at everything and investing in everything.” First-round check sizes rose from 2024, as founders from big tech carried higher valuation expectations; in 2024, more of the investments were independent developers. Of the 50-plus deals already closed, more than half have completed their next financing round. Excluding projects constrained by SAFE terms, 8 of the remaining dozen-plus received follow-on investment, and 2 received multiple additional investments.
The 2026 outlook is optimistic. Several model companies and some embodied-AI companies may go public, leading funds are gradually completing fundraising, and “voices are returning for mega funds”—a category that “had effectively disappeared from the Chinese market for a time.” China’s capital markets “could be quite lively next year.” AI entrepreneurship is an air-quoted “Shangri-La”: macro, FX, and interest rates need less attention, but that will attract more entrants. “There is no need to rush, but time is not infinite either.”
If only one direction could be funded next year, the choice would be expert-knowledge applications deployed across industries. The form does not matter; if the product must interact with the physical world, add hardware or robots. As for the criticism that embodied AI has “not found a landing scenario”: “That is what the next phase of embodied AI needs to solve.”

Deep dive

1. Leaving ByteDance: Choosing the Work I Loved Most Over the Company I Loved Most

杨洁 worked on strategy and investments at ByteDance in 2022. When internal restructuring made investing difficult, “between the company I loved most and the work I loved most, I chose the work I loved most.” Several colleagues in similar situations did not look for jobs; they joined her to build 锦秋. All 3 partners seem to be INTJs.
Her quick-fire self-description: “I am still at an age where I dare to place bets and can carry the consequences.” She believes in MBTI; astrology charts provide emotional support, while “postmortems take care of cash flow.”
Her feel after 3 years of entrepreneurship: she underestimated the difficulty at first, but once the problems were broken down and made more specific, “we worked through them step by step, and every problem could be solved.” It also gave her “a stronger feel for the real state of entrepreneurs”: every day is about finding problems and solving them.

2. Betting on AI Before ChatGPT: The Signal Was in ByteDance’s 50-Plus Investments

Many people assume 锦秋 was founded after the foundation-model boom. In fact, its 2022 fundraising pitch was already about AI and robotics: the BP modeled the global populations of construction workers, blue-collar workers, and repetitive white-collar workers, along with the share AI could take on.
The signal was empirical. The team had invested in more than 50 minority-stake companies at ByteDance, most related to AI or robotics, with “clear revenue already taking off” in their deployment scenarios. 未来机器人, for example, still doubled revenue every year during the pandemic. The direction was unfashionable at the time, leaving an open space in the market.
She attributes the result to timing: “At the moment the wind started blowing, we were already standing there, with everything prepared.” In 2023, the team immediately began betting aggressively on the transform direction and quickly launched a 12-year fund.

3. $500M and a 12-Year Term: Both Were Designed by Calculation

The size was deliberate. A $200M fund can only make spot investments and lacks the capacity to add in later rounds; a $1B fund “would inevitably end up chasing big projects” and be forced to demand stable, predictable returns. $500M is enough to invest early and keep adding later. Roughly half the portfolio is invested at the first round, with the rest at Series A or later.
The 12-year logic is that this technology variable “is not an industry cycle but an infrastructure-type variable.” It will reshape industries one by one, and a longer cycle improves the return profile. With a seven-year fund, “probably 70-80% of AI software applications would be too risky to touch,” as would many compute technologies and embodied-AI projects.

4. The Method Is “Fast Certainty”; Speed Is Earned Through Upfront Research

Speed and certainty sound contradictory, but they are not: “Front-load enough work, and the final result can come quickly.” That means studying directions in advance and building a base of knowledge about “what kinds of things work, where the good people are, and what it takes to outrun the competition.” When the right person appears, decisions come quickly. Directions without that preparation are “sometimes worth looking at less.”
锦秋 is not a broad-scanning fund. In compute, it invests only in compute-in-memory and optical computing, the 2 small areas aimed at breaking the compute bottleneck. Applications center on AI+marketing, creative tools, and recently AI+healthcare. Embodied AI is reviewed sector by sector, “very, very narrowly,” including even later-stage areas such as “language and math.”
On whether narrow sector choices might exclude companies that later grow into something bigger: there will be misses. “But there is no way around that—early-stage investing does not require being too afraid of missing something. The main thing is to draw your own boundaries and stay disciplined about what you invest in and what you do not.”

5. Missing Manus: The Signal Was Antifragility

锦秋 had a shot at Manus in its earliest round, but “we spent our attention evaluating a plugin product.” The retrospective lesson was that the team’s transition from Monica to Manus showed an ability to “fight disorder and remain antifragile” while building “more interesting products”—a trait the team failed to observe early on.
That is why people still carry substantial weight in AI application investing. The 3 traits are resilience, learning ability, and execution. Resilience ranks first because “everything else can be learned. If you cannot code, you can learn it from Bilibili or YouTube—but resilience is, to some extent, innate.” Execution matters more than before because AI tools amplify the output of a strong team.
How do you assess resilience? Look at what someone achieved and under what conditions, then “ask about things that did not go as expected and see how they attribute them.” If any one of the 3 traits is weak, the investment is a pass: “Yes, I do that. I mainly think there is no need to fear missing.” In AI, “know-how is not a moat; continuous, rapid learning is.”

6. A Real-Life Resilience Sample: Closing the Excruciating LP Spreadsheet

The 2022 fundraising was the moment she was “most likely to give up.” ByteDance and Sequoia were both anchor LPs, “but even with that, fundraising was extremely difficult.” After the Russia-Ukraine conflict, the strategy had to shift toward Chinese LPs. “People often say down time is the best time for a startup—true, but that best time is painful.”
Her response was counterintuitive. The LP list in the multi-dimensional spreadsheet was “excruciating. If looking at that spreadsheet makes you miserable, then do not make the spreadsheet.” In the short term, “do not review the past yet”; put the focus on what can move forward. 天宇’s response was: “Just keep moving forward. Talk to whoever needs to be talked to.”
The mindset changed after raising 2 funds: “How big is our fund, really? How much money is there in the market looking to invest in the primary market? It does not matter if some people reject us—we can always raise the fund.” It was also her first concentrated experience of mass rejection: “Maybe before that, I was usually the one rejecting others.”

7. No Foundation-Model Investments: Doubao Is Still Free at 100M DAU

Starting in 2023, the fund explicitly decided not to invest in foundation-model companies, and “we still believe that decision was right.” Even if 智谱, Mini Max, and Kimi go public next year, there is no regret: “You can run the numbers—if we had invested in them in 2023 or 2024 and held to an IPO, how much would we actually have made? Would that really be an investment where the risk and return were proportionate?”
The core argument: “People often overlook one fact: even with 100M DAU, Doubao is still free today.” Model competition is brutal, the rivals are not just startups, and the resource requirements are enormous. “There was no need to pick a hard bone to chew.” One more reality check: OpenAI forecasts a $78B loss in 2028.
锦秋 made differentiated bets in parallel, investing aggressively in embodied AI, applications, and compute-in-memory in 2023-24. “For the risk and return of early-stage investing, that may be better.” Her view of current hot areas—embodied AI, dexterous hands, and consumer electronics—is restrained: “There may be nothing wrong with them. These are all real businesses. But whether they fit our investment strategy is another question.”

8. Valuation Matters Less Early; Follow-On Investment Requires Discipline

Valuation matters, but early-stage investing is not highly valuation-sensitive. She has never abandoned a deal she liked because of price: “If something ultimately generates a very high multiple return, paying a little more at the early stage does not have that much impact on the later return.” Investing in growth can mean “prepaying 1-2 years of growth”; pure concepts are “indeed rarely funded.”
First-round check sizes were clearly larger in 2025 than in 2024, for 3 reasons: competition intensified, founders coming out of big tech already had higher valuation expectations, and the valuation had risen by the time of follow-on investment, requiring more capital to add.
Of the 50-plus deals already closed, more than half have completed their next round, with a dozen-plus currently fundraising. Excluding projects whose SAFE terms prevented additional investment, 8 of the remaining dozen-plus received follow-on investment and 2 received multiple rounds of additional investment. The logic remains “fast and certain”: keep investing capital and post-investment resources in directions the fund likes. Asked whether continued investment in the wrong projects can produce bad outcomes, she conceded: “It can.”

9. Sequoia and Hillhouse “Invest in Everything”; AI Helped Work Out the Differentiation

In mid-2024, competition on both the China and US sides did not feel particularly intense. “In the first half of this year, we suddenly felt competition take off.” Sequoia and Hillhouse became highly active, and “it is not easy to find a differentiated space to compete with them—they now look at everything and invest in everything.” But not every founder needs Sequoia or Hillhouse’s money; “people still look for a fit.”
The central question at the team offsite was how to differentiate. ChatGPT supplied a framework for breaking the discussion into rounds and angles; Claude and ChatGPT wrote the questions and set the scoring rules, while AI even had a seat on the judging panel. ChatGPT also suggested revisiting the first day’s questions on the final day. The conclusion: “Strengthen our own investment capabilities and do a better job of providing value-added services to portfolio companies.”
Founders choose 锦秋 for speed and certainty: “We make decisions quickly and give feedback quickly. We can clearly explain whether we are investing or not—we will not say, ‘You still need to meet the next partner.’” There is also differentiated insight, such as how to turn a demo into something scalable without it breaking during the scaling process. The criticism that 锦秋’s investments are “a bit watery” is not especially concerning: reviewing 4,000-5,000 projects and investing in 50, roughly 1%, “is not necessarily high compared with most early-stage investing.”

10. 锦秋 Is Itself an AI Company

“In the past, traditional institutions invested in AI companies. Now we are also an AI company investing in AI companies.” Its AI sourcing system uses fixed sources such as paper authors, university lists, and award lists, while also identifying deal opportunities from posts by preselected WeChat accounts. The team uses AI for full-web analysis of a target’s public commentary, paper-quality assessment, code audits, and competitive analysis—“it is more efficient than people and gives us broader information.” Work is largely based on Feishu and multi-dimensional tables; “it is already difficult to separate which parts are AI.”
Her 2 most recent aha moments both came from ChatGPT. Purse generated content in response to her situation: a roundup of trending AI tools on Twitter, a “90-day personal-brand execution plan for an early-stage AI investor,” and even nutrition and care advice related to her father’s treatment. After 5.1 launched, she used agent mode to start a new Twitter account, have it follow the 20 builders recommended on the website, and update their activity weekly. The first step worked.
How much money would it take for her to go a month without using AI? “I might not even want that. During my father’s treatment, all of his treatment data was put into ChatGPT, where I created a project.”

11. From RL Research to 红红模拟器: Founders in the Wild

朱哲清 was still at Meta in 2024. 锦秋 was tracking RL—“there were not that many people paying attention to RL at the time”—and found him among the Chinese students of a prominent Stanford professor. The RL opportunities and new business models he shared were highly valuable. After he started a company, 锦秋 invested, and the company later received strategic investments from Samsung and Intel.
王登科 built 红红模拟器 in 2023, making him “one of the earliest people to use large models to build fun small products.” At their first product conversation in a café, he hand-built an English-learning product on the spot. Many people tried to talk him out of building hardware. 锦秋 supported him: “This will not hurt his core business. He rarely gets to pursue something he wants to do, so there is no need to pour cold water on it.”
Founder profiles are becoming more diverse. In Wuhan, 方瑞 “single-handedly” built an AI plugin for multi-dimensional tables with a 40-50% share. Talent is no longer hidden in a fixed set of valleys, which is precisely what the AI sourcing system is meant to solve. The example of a wild idea eventually proving itself is “Poblasty” (the original name): “At first, we thought a startup had little chance in search, but it has done surprisingly well.”

12. The Investor as the Founder’s “21st Partner”

The phrase came from a colleague: “AI may be the 21st partner of our fund.” She borrowed it to define the investor-founder relationship: “You are a partner, so you matter; you are ranked 21st, so you do not matter that much.” Entrepreneurship deserves respect. “We do not casually criticize or instruct a founder. Our experience is simply information we give them; it does not mean they have to do things that way.”
Her 3 requirements for the team are straightforward: “Ask when you do not understand, give feedback promptly, and do not point fingers at founders.” The small dinners and CEO summit are designed to create a “field”: founders do not particularly need a “teacher” but do need a place to find partners, people, and customers. Non-portfolio CEOs are not invited to the CEO summit—“the protagonists on this AI stage are all of you.”
She tries to assess what drives a founder. Working 10-plus hours a day with little rest during the week is painful “if you are doing something you do not love, chasing a higher valuation, or being pushed along by external factors.” 杨洁 recalls her first job as a product manager with 王鑫: during campus recruiting, young people were told to choose work that was “interesting, advantageous, and useful,” with interesting first. “That had a major influence on many of my choices afterward.”

13. Making People Work from Passion, Not Anxiety

The primary market naturally lacks a sense of security. At Sequoia in the US, the first onboarding ritual is copying the line, “You are only as good as your next investment.” Her answer is to choose people who want to win and are smart—“they know when they need to work with whom”—point out problems while offering suggestions for next time, and require colleagues to “also have confidence in me; they should not decide that everything is impossible because of one project.”
The most important move is to admit her own mistakes: “If everyone sees that you make mistakes too, they will not feel pressure about making mistakes themselves.” She has told the Manus miss story internally many times to “desensitize everyone to mistakes.” Information is shared as broadly as possible, and project visits and discussions are relatively open. The goal is to avoid the fund’s most common failure mode: “Everyone is an island, nobody knows anything, and all they can do is charge ahead.”
Her underlying attitude: “If I can make it happen, I hope people work because they love the work, not because they are anxious.” Her own relationship with VC has also changed. For the first 10 years, it was about “learning different industries quickly from CEOs”; now it is about building this AI wave together with CEOs—“I do not want to be a bystander.” Financial returns and social status? “Probably negative. If you want financial returns, there are better things to do; if you want social impact, there are faster ways.”

14. 2026: If I Could Invest in Only One Direction, It Would Be Expert-Knowledge Applications

The INTJ-style answer to the subtraction exercise is “expert-knowledge applications in specific fields”: AI applications deployed across industries, in whatever form works. If they need to interact with the physical world, add hardware or robots. To the criticism that embodied AI has “not found a landing scenario,” she gives one answer: “That is what the next phase of embodied AI needs to solve.”
Her view for 2026 is that AI “is still at the beginning of a long cycle.” As model capabilities improve and compute costs decline, deployment will spread further. Many products have entered the deep-water phase of refinement, so the market can expect products with “better usability, higher completion quality, and higher project success rates.”
The capital markets “could be quite lively next year.” Several model companies may go public, some embodied-AI companies may do the same, and Hong Kong has already seen several AI listings this year. Active funds are gradually completing fundraising, while the late-stage mega funds that had been quiet for 2 years “have started making noise again.” “Mega funds had effectively disappeared from the Chinese market for a time”; restoring that later-stage capital is important.
Her view on pace: “It will continue for a long time, penetrating industries one by one. There is no need to rush, but time is not infinite.” AI entrepreneurship is an air-quoted “Shangri-La”: macro, FX, and interest rates matter less, but that is precisely why more people will want to enter. “If you start late, competition will be tougher later, but the opportunity is long-term.”

15. Epilogue: The AI Builder Turning Investing into a Product

The most product-minded of her 10 “I am” statements is: “I am an AI Builder who wants to turn investing into a product.” The product “can provide capital that is fast and highly certain, while also providing dependable support.”
The most revealing statement is: “I am someone who can make decisions amid imperfection.” At ByteDance, the team was large enough to “collect as much information as possible and converge late.” Now, even with AI, the granularity and depth of information remain imperfect. “But decisions still have to be made. We make them every day.”
On the human side, she owns 6 cats, specifically buying cats that have been unsold at pet stores for a long time and kept in cramped spaces, plus 1 tortoise and 2 lizards. “Every year, make sure you do not do what you did last year”—which is why VC suits her. One AMA project post added 7,000 followers on Xiaohongshu. Her closing line: “I am someone who worries about my father’s health every day.”

E219 | Dynastic Feuds and an Acquisition Battle: How Did €14B of Hermès Shares Disappear?“How We Can Eliminate Crime” | Ben Horowitz and Garrett Langley