Kathy Xu: 51 Methodologies on Embodied Artificial Intelligence, AI, and Venture Capital

Raise money while the bubble lasts.

🥷 Editor: Koji

🧑‍🎨 Layout: NCon

This week, Kathy Xu appeared on the Gao Neng Liang podcast: "Vol. 227: Embodied Intelligence Newcomers and the Queen of Investing — Innovation, Booms, Bubbles, and Cycles." In this episode, she shared a high-density stream of perspectives on embodied intelligence, AI, and investing — a rare public appearance with this level of depth in recent years.

A caveat: this reads more like a methodological self-portrait of a practitioner on the front lines. What's most worth absorbing is her mental framework for researching sectors, judging people, and allocating capital. As for specific judgments about which companies led on particular technical paths, which are stronger on full-system versus supply-chain capabilities, or whose data scale or technology is more advanced — these largely reflect Kathy Xu's personal perspective alongside her co-host Gao Jiyang (founder and CEO of Xinghai Tu). The industry may not share consensus on these points, so treat them as one informed opinion among many.

We've organized Kathy Xu's remarks from the full episode into 51 points across 7 categories.

I. How to Research an Unfamiliar Sector

1. Don't miss the forest for the trees

   The biggest mistake when entering a new industry is focusing only on a handful of star companies. To understand the "forest," you need to examine as many "trees" as possible. Before investing in Embodied Artificial Intelligence, Capital Today met intensively with a large number of teams in the space and compared them horizontally.

2. Early-stage investing is first and foremost about picking people

   When the existing market is small, future growth is large, and business data is still insufficient to judge by, the core is selecting the right founder team. She particularly values the combination of "strongest brain" and engineering mass-production capability: someone who can train foundation models and has also done hardware, supply chain, and batch manufacturing.

3. Three learning mechanisms

   Her methods for learning a sector include:

   • Learn from founders: Deep conversations every two to three months, focused on "what have you learned recently," "what are you seeing that others aren't," and "what's your biggest growing pain right now."

   • Winner Pattern Study: Study historical winners, leading labs, founder speeches, podcasts, earnings call transcripts, and technical roadmaps to identify winning patterns as early as possible.

   • Consumer/customer insight: Interview users, developers, and enterprise customers directly to understand real usage scenarios, willingness to pay, and pricing boundaries.

4. Information input quality determines cognitive ceiling

   She tracks a small number of the most outstanding large-model founders, embodied intelligence founders, and top scholars; for long podcasts, she starts with translations, summaries, and takeaways, returning to full transcripts only when needed. The point isn't maximum information volume but consistently listening to "the best people."

5. Set fixed time for thinking about long-term questions

   She believes in "fixed time for fixed things": daily meditation followed by review of the previous day's learning; monthly identification of a major question, allowing it to remain on the list for months and be revisited repeatedly rather than demanding immediate answers.

II. How to Judge Founders and Teams

6. Killer instinct: seeing what others don't

   The first common trait of excellent founders is killer instinct / tech vision: the ability to see deeper or farther than others, and the courage to make exclusive bets on a path. Richard Liu's early investment in delivery infrastructure and Zhilin Yang's relatively early bet on long-context and Agent capabilities are examples she cites.

7. Learning speed and execution speed must both be high

   Fast learning with slow execution isn't enough; truly excellent founders rapidly update both industry and self-awareness, immediately translating these into product and organizational moves. Her judgment of Gao Jiyang centers on: fast learner, strong executor, willing to make complex decisions.

8. Ambition and long-term patience

   What she values isn't just profit motive but whether founders want to build something strong, large, and enduring. Great companies require patience on a ten-year scale; short-term gold rushers struggle to attract equally excellent long-term partners.

9. Judging mission requires "unity of knowledge and action"

   To identify whether an entrepreneur is gold-rushing or mission-driven, observe whether what they say, believe, and do align; also look at what co-founders they attract. Excellent co-founders willing to bet the next decade — who know the lead founder better than investors do — constitute strong validation of the founder.

10. Top scientists create strong talent gravity

    The scarce resource in AI entrepreneurship isn't ordinary PhDs but scientists who simultaneously pursue truth, are willing to start companies, and possess tech vision. The best students tend to follow those who are technically strongest and most likely to succeed, not just compare equity and salary.

11. Credentials aren't the endpoint; "scientist status" is what drives the talent flywheel

    She argues the key isn't bachelor's, master's, or doctorate but whether the founder has sufficient scientific status to attract top researchers to follow. The core research team may need only twenty to thirty extremely smart people — provided you can attract, retain, and stabilize them.

12. The deeper the understanding, the more willing to bet on one direction

    Founders who truly understand technology can hear the frontline gunfire and know which path to bet on; those who don't understand deeply enough tend to want "both this and that," dispersing limited people, money, and time across multiple directions. She cites Anthropic, Moonshot AI, and DeepSeek as positive examples of focus.

III. Core Judgments on Embodied Artificial Intelligence

13. Embodied intelligence is harder than autonomous driving, large models, and traditional robotics

    It simultaneously requires foundation models, hardware bodies, supply chains, data collection, mass production, and real-scenario delivery. Strength in any single item is insufficient; complexity comes from the entire chain.

14. Must be "brain + body" integrated

    Body-only lacks the ability to continuously grow smarter; brain-only can't access real-world data. She endorses Xinghai Tu's approach of "intelligence defining the body, full system plus intelligence," believing integration is the hardest path but possibly the best long-term structure.

15. Hardware first, then brain training, is the reasonable sequence

    Without proprietary bodies, the brain lacks stable data feedback and iteration carriers; burning money on model training too early may prove inefficient without sufficient real-world data.

16. Physical-world data is the key bottleneck

    Language models can leverage existing internet data; robot interaction data must be actively collected. Whether embodied intelligence can expand from roughly millions to tens of millions of hours of data is a variable she repeatedly probes; but she explicitly acknowledges that required data volumes and timelines remain uncertain.

17. "The brain gets more valuable with use; hardware gets cheaper with production"

    Hardware must continuously reduce costs and barriers to use; models continuously learn and increase value through data and usage. Thus brain and body aren't simple addition but the combination of two value curves.

18. Currently still at an extremely early stage

    She estimates embodied intelligence technology penetration remains around 2%–3%, in the innovator phase, not yet fully crossing the chasm. She speculates data and hardware iteration may achieve key breakthroughs in the next two to three years but repeatedly emphasizes timeline uncertainty.

19. Commercialization should start with B2B dirty, hard, and exhausting work

    Traditional robots have already proven demand in handling, warehousing, logistics, and photovoltaic construction. General embodied intelligence costs more but can execute more complex and flexible tasks; near-term buyers should still be enterprises, not the To C market centered on home companionship.

20. After crossing the chasm, Get Big Fast

    Once disruptive technology crosses the chasm, three to five consecutive years of high-speed growth may occur. If a company belongs to the operating system or infrastructure layer and scales fastest during the surge, market share gained may prove highly durable.

21. Integrated operating system value may exceed pure infrastructure

    Her historical tiered comparison: PC era, Windows (Microsoft) was the operating system, Intel the infrastructure; Mobile Internet era, iOS (Apple) and Android (Google) were operating systems, Qualcomm the infrastructure. She generalizes that in mature structures, operating-system-layer companies may reach three to five times the market cap of infrastructure companies. This "three to five times" is her empirical induction from the episode, not a strict disaggregated comparison of individual company business values.

    Models combined with Agents form ecosystems with higher value ceilings than pure compute infrastructure; embodied companies that simultaneously control device and operating system have greater potential.

22. Embodied intelligence endgame may be highly concentrated

    Annual VLA model investment may reach hundreds of millions to roughly a billion dollars, plus real-world data and top talent — not many companies can stay at the table long-term. Whether it's three to five or more, she offers no definite answer.

IV. Judgments on AI Business Models and Competitive Landscape

23. Large models are a "good business"

    She judges that large models exhibit clear economies of scale: massive fixed training investment, scarce top talent, high barriers to entry, with competition likely converging to three to five players. As pricing rises, token costs fall, and Agents drive hundred-fold or greater consumption increases, model gross margins and profit prospects can work.

24. Leading models may develop network effects in post-training

    Pre-training may not have traditional internet-style network effects, but in coding and reinforcement learning phases, the best engineers training models on the hardest problems, with models improving to attract more high-value users, may create positive loops.

25. AI applications must be ten times better

    This AI wave hasn't replaced phones or other endpoints; traffic remains in big tech's hands. Startups pay both token costs and customer acquisition costs; if products aren't dramatically better than alternatives, unit economics struggle to work.

26. Startups without historical baggage can run faster

    When massive waves arrive, startups focused on single directions are often more agile than complex big tech. She contrasts Anthropic with Google, seeing the former as light and focused; but Google, through founder return, unified technical command, and resource concentration, can also restore competitiveness.

27. AI is "the era when scientists make money"

    This wave's higher-probability entrepreneurs tend to have strong scientific backgrounds. Compared to the product manager dividend of consumer internet, AI's barriers depend more on frontier scientific judgment and top research talent.

28. Token economy is new growth engine

    Traditional consumption and mature industries may only achieve low growth, while models, Agents, and token consumption may maintain multi-fold growth. She therefore believes VC capital should prioritize flows toward token production and consumption chains.

29. This AI wave is bigger than the internet

    She argues the most direct indicator of disruption is growth rate: internet and Mobile Internet commonly saw three to five times annual growth, while AI products may grow ten times annually. AI isn't just connection or tool; it will directly substitute some human labor, thus has a higher ceiling.

30. China may take the lead in embodied intelligence

    She believes China has three overlapping advantages:

    • Founder dividend of those who understand hardware-software integration;
    • Engineer and supply chain dividend;
    • Government dividend in infrastructure, data organization, and industrial execution.

    Large language models remain mainly catch-up, but embodied intelligence may offer a leading window.

V. Venture Capital and Capital Allocation Methods

31. VC invests in change; value investing invests in what doesn't change

    She sees Warren Buffett as seeking what remains constant long-term, while VC's object is change; though different, "focus" and "long-term holding" remain shared principles.

32. Investing in technology is like surfing, not mountain climbing

    Traditional industry investing is like mountain climbing — the opportunity is always there; technology investing is like surfing — miss the big wave and you wait for the next one. The correct approach: enter first with small checks of affordable loss, then rapidly add capital when you see the signal.

33. Home run intensity matters more than frequency

    A fund's number of great companies is roughly fixed, about three to five. Increasing project count and team size doesn't improve hit rate, instead diluting attention. What's truly controllable is ownership percentage and holding duration.

34. Double down, triple down, and hold for eight to ten years

    After confirming winners, add across multiple rounds and give time for compounding. She cites Moonshot AI — five additions over three years — and Xinghai Tu's continuous follow-on investment, believing eight to ten years qualifies as sufficiently long.

35. Excessive fund scale may damage returns

    Great companies are limited in number. After fund scale expands, achieving five to eight times returns becomes markedly harder; she believes roughly $600 million to $1 billion remains relatively manageable, beyond which forced diversification tends to occur.

36. Biggest regrets come from insufficient cognition, not valuation too high

    She passed on ByteDance at roughly $7 billion valuation early on, only understanding its model after studying Facebook's news feed and recommendation engine. The lesson: "you can't earn money beyond your cognition," so Winner Pattern Study must begin earlier.

37. Heavy Meituan bet came from judgment on human nature

    Capital Today once deployed roughly 55% of a single fund into Meituan. She used Toshifumi Suzuki's convenience store case to understand subsidies: convenient services, once repeatedly used, form habits; users won't easily return to more cumbersome old ways — the underlying driver is "human nature's laziness."

38. Losses aren't scary; cash flow breakage is scary

    When JD.com needed $2 million in funding, she gave $10 million, enabling it to expand categories, build delivery, and enter roughly 30 cities. The company then grew continuously at high speed; though persistently loss-making, it maintained cash flow through supply chain payment terms. Her principle: as long as cash flow is healthy and profitability is visible, losing money for growth is acceptable.

39. Look at incremental share, not just existing scale

    In rapidly expanding markets, "most leading" isn't current highest revenue but who captures the largest incremental market share and maintains speed. The essence of Get Big Fast is occupying the bulk of increment during the window.

40. Bubbles themselves aren't scary

    After the internet bubble burst, NetEase, SINA Corporation, and Sohu — which had raised money in advance — survived and gained roughly three years of competitive vacuum. Bubble bursts wash out cash-poor players; good companies may actually get better M&A and add-on opportunities.

41. Raise Money When You Can

    Capital sentiment oscillates between greed and fear. When greedy, people invest even at high prices; when fearful, even good companies struggle to raise. Today's market remains a greedy market, so founders should stockpile capital while the market is willing to give, not wait until they need it.

42. Pace matters more than pure aggression

    Early embodied companies that run too fast may burn out before revenue appears; too slow and they lose funding, talent, and leadership position. Correct pace is dynamic balance between "surviving until spring" and "maintaining leadership."

VI. Organization, Talent, and Founder Mode

43. AI-native organizations should be smaller and lighter

    She anticipates AI companies won't need the hundreds of thousands of troops of the consumer internet era; a handful of strongest talent directly deploying Agents, teams of twenty to thirty may generate 100 million RMB in revenue. Middle management will be drastically reduced.

44. Hiring standard becomes "is person + Agent better than Agent alone"

    She relays one founder's hiring formula: ordinary high-salary employees who, combined with Agents, don't show clear marginal gain aren't worth hiring; hire extremely strong talent who can command Agents, even at higher individual compensation.

45. Founders should spend most time on irreplaceable things

    Supply chain, HR, expense reimbursement, and operations should be systematized and delegated as soon as possible. What's irreplaceable for founders is technical direction, vision, resource mobilization, and attracting top talent. She advises technical founders to devote more than half, even roughly 80%, of energy to papers, frontier researcher exchange, and strategic judgment.

46. Lack of focus is fatal

    The most common mistake after raising big money is overestimating oneself and launching too many projects simultaneously. Betting wrong on model direction wastes time, capital, and talent simultaneously, letting competitors build leads. She believes the technology era cannot tolerate frequent major strategic errors.

47. Founder IP's purpose is attracting talent, not chasing traffic

    Founders need to let talent know their vision, convictions, and technical iteration, but shouldn't consume massive time on social media traffic. High-quality product launches and periodic public expression are sufficient.

48. Frugality is long-term operational capability

    She treats frugality as a virtue in tech companies, citing Meituan, JD.com, and others. Capital-intensive industries especially need to avoid organizational bloat and inefficient spending.

VII. AI's Social Impact

49. AI may widen wealth inequality

    The strong can use AI to further enhance capabilities; those who don't understand AI may face greater pressure. Between now and an idealized AGI society, there will be transition pains in employment, data security, and re-entrepreneurship.

50. Transition period needs values and institutional guidance

    She believes society needs to help affected people re-start businesses or find new work and interests, and discuss the transition period from philosophical, values, and public policy perspectives, not just focus on technological progress.

51. Don't assume humans will have nothing to do

    AIGC has already brought new work and new companies in music, video, novels, short dramas. Beyond token production, massive token consumption industries will emerge, so she doesn't worry about humans having nothing to do.

References

[1] Gao Neng Liang podcast: "Vol. 227: Embodied Intelligence Newcomers and the Queen of Investing — Innovation, Booms, Bubbles, and Cycles": https://www.xiaoyuzhoufm.com/episode/6a626f5fa3fec224d5a26510