The Real Differences in Chinese and U.S. AI Venture Capital | A Conversation with Jenny, Partner at Leonis Capital
Summary
The U.S. AI application battleground is shifting from “万能 agent” to stable, controllable enterprise workflows that can replace expensive labor. Large enterprises may reject even 95% accuracy because human review can leave the product “creating a lot of extra work”; what sells is breaking a process into small steps and giving AI only limited choices at each one. In the U.S., annual contracts of $100K-$200K can still count as small deals, while large contracts reach $1M-$2M or even several million dollars. The pricing anchor is how much work AI can replace for roughly $10K that previously required a $70K-$80K employee.
The China-U.S. split between To C and To B is driven first by market structure, not simply by founder preference. Jenny believes the U.S. consumer market is highly fragmented by ethnicity, age, and geography—even the payment apps preferred by Black, white, and Asian consumers may differ. At the same time, expensive labor and tipping culture have made Americans comfortable spending $3-$5 to solve a small pain point, and more willing to pay $20 a month for software. U.S. prosumers therefore often start as individual users before spreading into enterprise procurement through a Notion- or Figma-style adoption path.
Application-company moats usually combine technology, data, and industry depth, rather than simply “training a model too.” Building an in-house RL or full AI team is expensive, and the next upgrade to the underlying model may erase the value of that work; without both substantial proprietary data and a strong technical team, it is rarely worth doing. Jenny believes coding is too tightly coupled to model capabilities, leaving many products exposed to model vendors; verticals such as law, finance, biology, and maintenance still offer depth that model companies struggle to cover.
Jenny believes AI companies should trade at lower multiples than traditional SaaS, and that both the private and public markets are currently in a bubble. SaaS often trades at 20-30x revenue, but AI companies incur token costs with every additional service, leaving many “looking like software on the surface but actually being infrastructure companies”; she therefore considers Perplexity and Cursor overvalued. High valuations help companies use stock to compete with OpenAI and Anthropic for talent, but also expose employees to the risk that their shares could become “worthless” after the next down round.
Jenny expects the U.S. AI bubble to burst in 2026, with Nvidia and OpenAI the two most likely pins. Any quarterly Nvidia sales miss could hit the public markets; if Google makes its now-improved Gemini completely free, she estimates it could take at least 30% of OpenAI’s users. She says roughly 60%-70% of OpenAI’s revenue comes from the $20 monthly subscription, so a 20%-30% user loss could drag down valuations across the entire private market: “It’s a very big balloon, just waiting for a pin.”
Raising money in the U.S. is itself a test of localization: a Delaware structure, one or two months living in Silicon Valley, warm introductions, and a dense process matter more than a single pitch. Warm introductions result in meetings with roughly half of funds, while cold emails may convert at only 10% or less; Jenny recommends approaching 40-50 funds simultaneously, with the goal of ultimately securing 2 or 3 investors. First meetings are usually capped at 30 minutes, and 3 or 4 meetings can be completed within 1 or 2 weeks, alongside reference checks with former employees, investors, and mutual contacts.
The hardest decision for cross-border founders is choosing early which market to serve and which side’s capital to accept. Jenny prefers founders who can clearly map competitors in categories A, B, and C and explain their wedge, rather than claim “only I can do this”; when a sector is clearly U.S.-oriented, she advocates “short-term pain over long-term pain” and absorbing the migration and fundraising costs upfront. Without an international image and strong marketing capabilities, $1M in ARR may only be the starting point, and large cultural gaps may require several million dollars of ARR; for a To B company already at $5M ARR, she considers an $80M-$100M valuation negotiable.
Deep dive
1. Jenny Chose to Leave OpenAI as a Choice About Her Path to Influence
Jenny Xiao grew up in the U.S. and lived in China for 8 years, completing Chinese middle and high school before pursuing a PhD at Columbia University. She worked at OpenAI for roughly 1 year in 2021-2022, during the period before ChatGPT launched and as the company was changing fastest. About 1 week after ChatGPT went viral, she decided to leave and start her own fund.
At OpenAI, she repeatedly asked herself who had the greatest influence over the company. The answer was Sam Altman, who had previously worked in venture capital. Around the same time, she met J, a partner with extensive investing experience. Jenny brought AI and technical judgment, while J brought investing training; the two began working together in late 2022, though J had already been building the fund since 2021.
Jenny still sees herself as learning while investing: “Every investor has to maintain a beginner’s mindset.” She criticizes some Silicon Valley investors in their 40s and 50s for entering a “semi-retired” state and losing touch with young founders. On one side are established funds in empty offices talking about golf; on the other are newer funds where founders still have an edge at Blue Bottle. She also notes that experience does not necessarily mean losing one’s vitality.
2. As the Foundation-Model Consensus Fades, Applications Move from Wrappers to Real Delivery
In 2023, the shared narrative among Chinese and U.S. investors was foundation models, with capital pouring into OpenAI and Anthropic. In hindsight, that was not a mistake: foundation-model companies grew fastest and captured most of the profits.
Lioness was already looking at applications, but the market broadly dismissed them as low-value, lightweight “wrappers.” By 2024-2025, companies such as Cursor and Perplexity had shown that applications could develop distinctive products and some degree of defensibility, materially improving the capital-market narrative around the application layer.
Agents then became the hottest theme of 2024-2025, but large enterprises generally stalled on reliability after trying them. Jenny sees the market moving from the grand story that “an AI agent can do XYZ” to the practical questions of how to deploy it, reduce errors, and lower review costs. Her conclusion is that founders have “become more pragmatic.”
Asked by 曲凯 about the absence of a new consensus narrative, Jenny did not point to another grand scaling law. She focuses instead on niche data: in healthcare, biology, and science, high-quality data for fine-tuning can still produce major capability gains. Roughly half of the messages she receives on LinkedIn ask whether she can arrange hourly data labeling in the senders’ PhD fields; people with PhDs around her commonly encounter the same demand.
3. What Enterprises and Governments Really Buy Is Constrained Reliability
While visiting a long-established Japanese bank, Jenny heard about a product—referred to in the source transcript as Robo—that helps investment bankers organize financial documents and handle Excel. The bank recognized its value but considered even 95% accuracy insufficient: it was “actually creating a lot of extra work,” because employees still had to check every item.
Her engineering solution is to break the workflow into discrete stages, allowing AI to choose among only a small number of options at each step rather than letting it “do anything and say anything.” Once error boundaries, review mechanisms, and accountability are controllable, large enterprises can turn a demo into a purchase.
To G offers a similar opportunity. Jenny describes much U.S. government software as resembling “software from the 1980s and 1990s,” while governments demand products that are as close to foolproof as possible. Outdated installed systems therefore create opportunity, but they also raise the reliability and delivery bar beyond what ordinary agents can reach.
4. The U.S. Is Not One Consumer Market, and Prosumers Often End Up Being Paid for by Companies
曲凯 observes that U.S. founders disproportionately build To B and To G products, with some investors openly saying they do not understand To C. Jenny attributes this to market fragmentation: different demographic groups may use different payment tools, and AI products need to target a single profile such as “U.S. Gen Z” or “rural conservatives.” It is difficult to gain a unified mass market naturally, as products can in China.
U.S. companies have an exceptionally strong willingness to pay for software because labor is expensive. If software can replace employees or increase their output, purchasing it is a rational decision. Combined with the everyday habit of paying $3-$5 for almost everything through tips, paying $20 a month to eliminate a small pain point feels natural. Feedback from founders suggests that Europe and Japan are far less willing to pay than the U.S.
Jenny classifies prosumer products according to who ultimately pays. In China, individuals paying out of pocket makes the product closer to To C. In the U.S., Notion and Figma often begin with individual adoption, then spread “one to ten, ten to a hundred,” before the company buys enterprise accounts for 50 or 100 users. The commercial end state therefore looks more like To B.
Lioness-backed Motion primarily serves individuals and companies with fewer than 20 employees, and these customers are already materially stickier than individual users. Another portfolio company, Jace, uses AI for email and productivity, and Lioness uses it internally as well. Jace’s larger customers often follow the same path: an individual saves time with the product, then spreads it across the company, allowing PLG to feed into enterprise sales.
5. Customer Concentration Can Be a Starting Point—or Hide Services Revenue
曲凯 notes that at some U.S. companies, the top 3 customers, or even a single customer, account for 60%-70% of sales. Jenny says this is common early on, but investors worry that the company has not found true PMF or is doing substantial extra services work for a small number of customers.
“Using investors’ money to do services” is a major taboo in U.S. venture investing. By the A and B rounds, companies generally need to diversify. Jenny considers a combined share of roughly 30%-40% for the top 5 customers a healthier profile.
The exit path is therefore clearer for To B and developer tools. If customers such as Microsoft or Google use a product successfully, they may acquire it outright and fold the functionality and team into their platforms. Mid-sized companies valued at several billion dollars also frequently acquire smaller teams. To C products have user profiles and independent brands that are harder to embed elsewhere, making them less likely acquisition targets.
6. Strong Founders Must See the Secret, the Competition, and Their Own Wedge
Jenny first looks for a founder’s unique insight into an industry: “You have to tell me why you know something other people don’t.” Good ideas in China can quickly attract 10 or 20 copycats. The U.S. has more fragmented verticals and stronger willingness to pay, leaving room to dig deeper into information gaps.
She asks directly whether someone else could build the product. If a founder says only they can do it and everyone else is incapable, she silently marks them down. The ideal answer maps the strengths of competitors A, B, and C, then explains where the company can enter and how it will move “from a point to a line to a surface.”
Jenny is wary of the Silicon Valley accelerator-style claim that “we’re the best and everyone else is garbage.” Learning quickly is not the same as blind confidence; it means constantly recalibrating the market map. She also dislikes people who ask for funding before quitting their jobs, comparing the behavior to “a married person joining a dating app,” promising to get divorced once they know someone will be there.
The investment stage determines whether the investor looks at the person or the business. Angels and accelerators can invest purely in the founder and care little about the initial idea. Lioness invests at seed and typically expects a small team of 2 or 3 people plus some customers, so direction, market, and early validation matter as well.
7. Behind the Age of Young U.S. Founders Is Earlier Responsibility
曲凯 questions why U.S. funds are willing to back large numbers of founders who have not yet graduated. Jenny explains it culturally: American teenagers may drive themselves to school at 16, and families and society give them independence earlier. Her husband started a company at 13 and ran a sizable cloud-services business at 16 or 17; experiences like that are “pretty common” in the U.S.
曲凯 says that after attending YC Demo Day every year, his impression is that founders are “younger every cohort.” He has seen many 17- or 18-year-old high-school dropouts, with the youngest around 15 or 16. He also says his company analyzed the 100 best-performing U.S. AI companies and found that founders most commonly were 26 or 27, with a median around 28 or 29; those aged 18-25 accounted for roughly 10%-15%.
Dropping out is not necessarily an irreversible gamble. U.S. schools allow leaves of absence, so failed founders can return years later, and many companies do not care much whether a candidate has a degree. Jenny left Columbia for OpenAI, later returned to finish her dissertation, and received her PhD that May. Her husband left Stanford after 1 or 2 semesters.
Jenny’s warning is: “Where you drop out from still matters a lot.” She backed a 19-year-old Chinese-American founder who had dropped out of the University of Pennsylvania. It was his first startup, but he already had a substantial open-source track record. By contrast, some Stanford students take leave to join YC for their resumes and return to school after raising money. Age should only trigger deeper questions about learning speed and the history of real projects.
8. Silicon Valley Favors Technical Talent in the Late 20s, but Valuation Has No Unified Rational Formula
The most popular profile today is around 27 or 28: either a Stanford PhD or similarly strong academic who is better suited to be a CTO, or an engineer who joined a large tech company at 22 and has accumulated roughly 5 years of experience, combining technical ability with the energy to be a CEO. Repeat founders with prior success are also highly sought after, but Jenny acknowledges age discrimination in Silicon Valley: people over 35 without significant achievements face more questions.
She calls Silicon Valley valuation “a form of metaphysics.” A little over $10M is normal for an angel round; $20M is considered expensive. Seed rounds commonly land around $20M-$25M, while YC companies can reach $30M-$40M.
Founders from major tech companies or with exceptionally strong backgrounds can completely break those ranges: even with nothing else, a valuation can jump directly to $100M-$200M. By Series A, current valuations generally exceed $100M, but the ARR required to reach an excellent pre-Series A profile has risen from roughly $1M to $3M-$5M.
5 years ago, seed-to-A typically took about 18 months and could be compressed to 12 months in a fast case. Today, both the revenue threshold and Series A valuation have risen, leaving many companies to run out of cash and raise seed extensions. Rising valuations are therefore not merely a matter of paper wealth; they also change cash planning and fundraising cadence.
9. AI’s Marginal Cost Means Its Multiples Should Not Simply Copy SaaS
Jenny’s core valuation objection is that once a traditional SaaS product is built, selling 1M or 10M copies requires almost no proportional increase in cost. AI companies still pay token costs for every additional service, and many products effectively carry infrastructure expenses, so their revenue multiples should be below the 20-30x commonly seen in SaaS.
She believes some leading projects maintain high valuations on momentum even if their user counts decline. Perplexity and Cursor both have structural problems that revenue growth has obscured. 曲凯 mentions later rounds where small amounts of new money or secondary-share pricing are used to mark up the valuation; Jenny explains that a high valuation helps a company compete for talent with stock.
Cursor has to compete with Anthropic and OpenAI for the best AI coding talent. Only a sufficiently high valuation makes the equity attractive. Employees often see only that “the higher the valuation, the more valuable the stock,” overlooking the possibility that it could become worthless after the next down round.
10. U.S. Fundraising Starts with Local Structure and Relationships, Not a Formal Pitch
Jenny recommends starting with the company structure. Many mainstream U.S. funds essentially invest only in Delaware companies, while Cayman structures can become an immediate red flag. If existing domestic shareholders temporarily prevent restructuring, founders should explain the migration plan candidly; she estimates that 80%-90% of U.S. investors will still engage after receiving a clear commitment to make the change.
The second step is to live in Silicon Valley for at least 1 month, preferably 2, and first meet founders, investors, and employees at major tech companies. Staying only 2 weeks and rushing to finish the raise exposes a founder as an outsider through countless details—for example, knowing only WeChat when exchanging contact information and not knowing how to scan a QR code on LinkedIn.
Using an FA to reach early-stage investors can be a negative signal in the U.S., implying that the founder lacks local relationships and cannot raise independently. Warm introductions result in meetings with roughly half of funds; cold emails may convert at 10% or less. Jenny recommends contacting 40-50 funds at the start and ultimately securing 2 or 3 investors. Silicon Valley has “several thousand” institutions, so the number of funds is not the bottleneck.
First meetings are strictly limited to 30 minutes, and being cut off does not mean there is no interest. The fund will usually bring in other partners and ask focused follow-ups on technology, data, or internal disagreements. Lioness typically decides after 3 or 4 meetings over roughly 2 weeks, or 1 week in a fast case; it can signal interest within 48 hours of the first meeting. At the same time, the fund contacts former employees, former investors, and mutual LinkedIn connections. A request for the data room usually signals strong interest.
11. Angel Endorsements Help, but Institutional Support Remains Limited
Before approaching institutions, Jenny recommends finding industry angels. An AI design company, for example, could ask people from Canva or Figma to invest $100K-$200K. The amount is not large; its value lies in professional validation and giving institutional investors one additional credible industry contact during diligence.
曲凯 believes active U.S. angels effectively perform part of an FA’s role by connecting founders with institutions, offering advice, and signaling conviction through personal investment. Jenny agrees. Founders can also ask funds directly how many companies they invest in each year, their stage, check size, valuation, and level of involvement, then tailor the pitch accordingly.
Asked what a fund can actually help with, Jenny gives a restrained answer: the most valuable help is introducing the next-round investor, customers, or employees, but investors do not have time to handle every task. Accelerators offer more systematic support, but take a larger stake at a lower valuation; “if you want investors to help, you have to pay a lot for it.”
12. The China-U.S. Product Aesthetic Divide Runs from Single-Point Tools to Integrated Systems
After returning to China, Jenny was most struck by the gap in product aesthetics. U.S. users are willing to let Granola do one thing—take notes—while Chinese users ask why Feishu cannot combine the functions of Notion, Granola, and a dozen other tools. The investment preferences follow the same pattern: the U.S. favors pure software and To B, while China is more drawn to To C and software-hardware combinations.
She still uses ChatGPT most often and says she has “made it sing.” Her life is divided into 8 areas—personal, health, work, research, and others—each with a long custom prompt, while each conversation has its own information configuration. Since the memory feature launched, she has been especially pleased with ChatGPT’s overall user experience.
For coding, she uses Claude Code more often and is bullish on Anthropic because its enterprise orders are growing quickly. She is less optimistic about Cursor and Perplexity, which she considers too close to the model vendors. By comparison, Lioness-backed Kepler uses AI for biological-data analysis. Jenny cannot validate the product herself, but her mother, who researches cardiology, uses it every day and rates it highly—reinforcing Jenny’s confidence in specialized verticals.
13. The Bubble Began in 2023, and Later Rounds Are Closer to the Impact Zone Than Seed
Jenny dates this valuation surge to the release of ChatGPT. Crypto-related events had previously depressed U.S. venture capital, while ChatGPT gave investors a new direction and narrative. The increase was relatively limited in 2023, accelerated most sharply in 2024-2025, and was pushed further by the participation of sovereign wealth funds.
She describes the market as “a very big balloon, just waiting for a pin.” If Nvidia has a weak sales quarter, the shock will spread through the public markets; if OpenAI’s revenue falls short, the impact will hit the private market directly. If Google makes Gemini completely free, she estimates it could take at least 30% of OpenAI’s users, while OpenAI relies on the $20 monthly subscription for roughly 60%-70% of its revenue.
Pre-seed and seed are furthest from real-market pricing and may take the smallest hit. Her comparison is $10M for pre-seed in 2021 versus roughly $15M-$18M today, and $15M-$20M for seed versus $20M-$25M today. The real pressure is in Series B and C. That is why Lioness avoids chasing projects that are already extremely expensive and did not accelerate its investment pace in 2025 simply because the market was hot.
14. Once Model Intelligence Hits a Plateau, Applications Can Defend Only Through Industry Depth
Jenny believes scaling laws have reached their endpoint. Most researchers may agree, while investors farther from the technology have not yet incorporated that view into valuations. RL, post-training, and vertical fine-tuning can still improve capabilities, but they look more like incremental repairs than the simple, direct, and universal gains produced by scaling laws in the past.
Her user experience is that the difference between GPT-3 and GPT-4 was enormous, while GPT-4 to GPT-5 was not a comparable leap. 曲凯 points out that GPT-5 and later versions added multimodality, retrieval, and other capabilities. Jenny sees these as more functional improvements: the product experience is clearly better, but that does not mean the model’s underlying intelligence is still improving at the same rate. She even says OpenAI and other major AI companies seem “a little out of tricks” on this front.
Jenny’s threshold for an application company to build its own model is “a large volume of exclusive proprietary data plus a very strong technical team.” Otherwise, hiring a complete AI team and doing RL or fine-tuning is too expensive, while the next-generation foundation model may naturally achieve the same result and wipe out the value of the earlier investment.
Asked how applications can ultimately compete with models, Jenny gives a blunt answer: “Most application companies will be eaten by model companies.” Coding is particularly exposed because it is so closely tied to the underlying model. Sectors such as law, finance, and biology require complex industry knowledge and workflows that OpenAI and Anthropic will struggle to cover comprehensively.
15. The Enterprise-Productivity Debate Will Ultimately Be Tested by Labor Replacement and Cross-Border Choices
Jenny summarizes 2 seemingly incompatible claims: an MIT article says 95% of enterprise AI applications ultimately fail, while research from Harvard and Yale says junior entry-level jobs on LinkedIn fell 20% or more from 2023 to 2025. She interprets the former as describing large enterprises building AI internally, while the latter reflects AI’s real impact on employment.
Large enterprises struggle to hire top AI talent and may lack the resolve to carry transformation through. Partnering with startups may produce much higher success rates. Lioness therefore looks for vertical companies with technology, data, and industry moats rather than multifunctional agents with polished demos.
The best-performing company in the fund’s first vehicle uses AI to manage maintenance. For companies such as McDonald’s, Starbucks, and Marriott that have internal maintenance teams, it records repair work, flags pipes or facilities that may need attention, provides recommendations, and schedules time. It does not perform repairs itself; it helps enterprises manage maintenance workflows that were previously handled by people.
Jenny’s investment standard is clear: “It has to replace people, not augment them.” If a product only improves employee efficiency by 20%-30%, customers may decide their existing headcount is sufficient. If it can compress a $70K-$80K annual role into roughly $10K of AI spending, willingness to pay rises sharply. In the U.S., annual contracts of $100K-$200K can still count as small deals, while large customers commonly pay $1M-$2M a year or even several million dollars.
16. Chinese Founders Must Choose the Market Before Choosing Which Side’s Money to Take
Jenny’s broad recommendation is to integrate into U.S. venture culture rather than simply carry a Chinese résumé into the market: live locally, enter the community, and build local relationships. She believes China has no shortage of excellent founders; the problem is that domestic software directions often lack the right environment. If a product is naturally suited to the U.S., the team needs the confidence to truly put down roots there.
曲凯 offers the most practical counterargument: excellent Chinese teams may easily raise more money at home, while moving to the U.S. means relocating, rebuilding credibility, accepting lower valuations, and facing greater uncertainty. Once domestic capital is taken, shifting toward the U.S. becomes harder. Jenny’s response is to let the market decide. For directions where the U.S. is clearly superior, founders should choose “short-term pain over long-term pain.” If conditions in China are overwhelmingly better, a China-Singapore path may make more sense; there is no need to raise U.S. capital merely to sell to U.S. customers.
Data can offset some of the discount for cultural and background differences, but the threshold varies. A U.S. To B seed company may raise money with $100K-$200K in revenue because investors still focus mainly on the team. With a purely Chinese background and a product style that differs materially from the U.S., a company may need at least $1M ARR—and in many cases several million dollars of ARR—to convince investors that “you can succeed wherever you go.”
曲凯 mentions a case referred to as “madness” in the source transcript. Jenny then uses the halo around “Mallence” to explain that its fundraising advantage came from more than data: it had a strong team and successful U.S. marketing. English-language videos went viral in the U.S. before spreading back to China, creating the first impression of an international Chinese team. For a To B company with $5M ARR, she believes a valuation of roughly $100M can work; conservatively, it could still negotiate at least $80M. The final figure depends on the sector, corporate structure, and other fundamentals.