You don't even have a secretary, so why do you need an AI one?
Zhu Xiaohu believes that 99.9% of daily work doesn't require frontier models, and open-weight models will eventually consume more tokens than closed-source ones; but Personal Agents lack high-frequency use cases and may repeat the failed subsidy playbook from ride-hailing a decade ago.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Frontier models are wasteful for everyday work
Zhu Xiaohu says you fundamentally don't need frontier models for 99.9% of daily work. He canceled his Claude subscription months ago because the combination of domestic Workbody and DeepSeek already takes his work "beyond expectations." He judges that open-weight models will eventually consume more tokens than closed-source models—the reason is straightforward: most people's everyday needs simply don't require such powerful models.
— Zhu XiaohuFoundation model margins far exceed cloud computing
Zhu Xiaohu gives a counterintuitive figure: a single GPU running inference for one hour generates $38 in revenue with costs of around $5, yielding gross margins exceeding 80%—higher than AWS cloud computing's 40%-50%. He also mentions Anthropic's staggering projected annual loss of $78 billion, but emphasizes that the company has actually turned profitable in the past two quarters—including both R&D and inference costs—showing that foundation model companies aren't margins-less but continuously investing in next-generation research.
— Zhu XiaohuPersonal Agents lack real high-frequency use cases
Facing the Personal Agent craze, Zhu Xiaohu counters: ‘Do you have a secretary?’ Most consumers lack sufficiently quantified routines and simply don't need a secretary. A decade ago, Laiyi (莱野) ran a similar business but ultimately survived only through ride-hailing subsidies and food-delivery subsidies. He observed that even users of American muse treat it like chatting with Doubao (豆包)—casual conversation rather than actual scheduling or report generation. Without high-frequency use cases, the business doesn't hold up; China's app silos and WeChat ecosystem only make it harder.
— Zhu XiaohuSpeed is just the entry fee; retention is the real moat
Over the past year, the startup ecosystem debated whether an AI product's moat is speed or retention. Zhu Xiaohu's conclusion: speed is just the entry fee, retention is the real moat. With Vibe Coding and better product development tools, almost every company can now drive iteration velocity—meaning speed itself no longer creates differentiation. The American investor benchmark is: for companies with under $2K in annual revenue, growth below 5x won't get looked at; 5x is the pass line, 10x is excellence. The two Haidian companies he invested in roughly double every two months.
— Zhu XiaohuCompanion businesses are more profitable than time-saving tools
Zhu Xiaohu believes that emotional companionship products are better suited for startups than ‘help you save time’ productivity tools, because companion users are price-insensitive while productivity tools get price-warred into irrelevance by big players. His example: Pax Historia, a YC-backed game that uses AI to interpret historical narratives, reportedly hit $100 million in ARR with a $300 million valuation—he was skeptical until he realized he personally spent $2,000 a month on it. Female users crave romantic fantasy, male users crave conquest fantasy—these time-killing needs are actually more valuable than time-saving ones.
— Zhu XiaohuThe best founders combine engineering excellence with business and human nature
Zhu Xiaohu summarizes the shared profile of successful founders over the past 20 years: excellent engineers who also understand business and human nature, with Zhang Yiming as a one-in-a-hundred example. He points out that China has no shortage of excellent engineers; what's scarce is the combination of "engineer + understands business + understands human nature." This is a reminder for ecosystems like Haidian, where engineering and research culture is thick but product and operations are weak: the AI era's demands on founder core competencies are as high or higher than the mobile internet era, not lower just because technical barriers have shifted.
— Zhu XiaohuHe passed on Chinese foundation models, and he's not sorry
Asked whether he regrets not investing in any Chinese foundation models, Zhu Xiaohu shifts the answer from judgment to liquidity: primary and secondary markets differ—in secondary markets you can dump a bad call the same evening, but primary market exits require waiting for someone else to buy. Over the past 12 months the trend has been consistent: post-lockup, prices are 80% off the peak, sometimes 90% off, and many projects are essentially write-offs even after IPO. So he prefers small positions to test, then adding after confirmation, rather than all-in betting on hot tracks—that style of investing would stress him out and "give him gray hair."
— Zhu XiaohuHe predicts 20-30 Chinese AI apps will hit $500M ARR next year
For next year, Zhu Xiaohu gives a specific forecast: China will see at least 20 to 30 AI application companies hit $500 million in ARR—lower and more achievable than the $1 billion threshold he mentioned previously. The prerequisite is going global; the domestic market alone can't support that threshold. He specifically mentions that Manus has already exceeded $500 million in ARR and is attempting to transform into a truly global-serving, Haidian-headquartered Chinese AI startup—he thinks it would be interesting if that happens.
— Zhu XiaohuIn their own words · checked verbatim
For 99.9% of daily work, you fundamentally don't need frontier models.
今天我觉得对于99.9%的日常工作 根本不需要用到 Frontier模型 顶级的
Zhu Xiaohu2:00
A single GPU for one hour generates $38 in revenue, costs maybe $5—so the gross margin is very high.
这个一块GPU 一个小时 创38美金收入 成本可能在5块美金 5块钱 对吧 所以他的毛利是很高的
Zhu Xiaohu4:01
Do you have a secretary? A Personal Agent? A Secret? Right?
你有秘书吗 你有Personal Agent 你有个Secret吗 对吧
Zhu Xiaohu8:03
Speed is the entry fee. Without it, no investor will even want to look.
速度是入场券 你没有这个速度根本就没有投资人会愿意看
Zhu Xiaohu12:03
Emotional companionship is probably better, because those users don't care about price.
情绪下的陪伴可能是最更好的 因为这些用户对价格不敏感
Zhu Xiaohu14:05
The best founders are excellent engineers who also understand human nature and business. It's really hard. Zhang Yiming is one—truly one in a hundred.
最好创业者就是优秀的工程师 在同时懂人性懂商业的 很难啊 张一鸣是一个 真的是百里挑一的
Zhu Xiaohu17:06
Post-lockup, prices are 80% off the peak, or even 90% off.
解禁以后价格是高峰的80%off 或者甚至90%off
Zhu Xiaohu18:06
Figures
| AI application ARR benchmark (Zhu Xiaohu) | $1 billion | 3:01 |
| GPU hourly inference revenue | $38 | 4:01 |
| GPU hourly inference cost | $5 | 4:01 |
| Foundation model gross margin | 80%+ | 5:01 |
| AWS cloud computing gross margin | 40%-50% | 5:01 |
| Post-lockup stock price decline from peak | 80%-90% | 18:06 |
| Pax Historia ARR and valuation | ARR ~$100 million, valuation ~$300 million | 15:05 |
| Zhu Xiaohu's monthly spending on Pax Historia | $2,000 | 15:05 |
| US investor growth benchmark (companies with sub-$2K annual revenue) | 5x = minimum, 10x = excellent | 12:03 |
| Predicted Chinese AI companies hitting $500M ARR next year | 20-30 | 21:08 |
Glossary
- Personal Agent
- AI private assistant that helps users manage schedules and process information
- ARR
- A company's annual recurring revenue
- Vibe Coding
- Development approach where natural language conversation lets AI generate code directly
- Open-weight model
- A model with publicly available parameters but not necessarily public training data
How to listen
Investors tracking the AI application space, entrepreneurs building Personal Agents or companion products.
Skip the opening pleasantries and review of last year's predictions; jump to the argument at 1:04.