Zhu Xiaohu Changes His Tune: DeepSeek Has Him Almost Believing in AGI
A year ago he said he would never invest in any Chinese large-model company; now he says if Liang Wenfeng opens a funding round, "I would definitely invest" — price no longer matters much, what matters is being part of witnessing the emergence of human AGI and AI consciousness.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
For the leader, open source is a curse — and soon it's too late
Zhu Xiaohu classifies OpenAI's flip-flopping on open source as "the leader's curse": when you're ahead you definitely want to keep it closed, and once others catch up, open-sourcing is hard. The mechanism he gives is cost — you spent ten times the cost to develop the foundation model, and a Chinese company can catch up at one-tenth the cost in at most twelve months, so whether to still pour that money in today is a test even for American tech giants and VCs. Hence his judgment that OpenAI will not open-source, because open-sourcing is already too late: programmers worldwide are already building on DeepSeek's open-source architecture, so open-sourcing again "wouldn't mean much."
— Zhu Xiaohu20 days to 20 million DAU, without spending a cent on ads
Zhu Xiaohu says DeepSeek is "the Android of the AI era has already appeared," because its growth speed is unprecedented: 20 days to 20 million DAU, and without spending a single cent on advertising, unlike many domestic companies that burn money on投放 — it relies entirely on word of mouth. He repeatedly stresses that its writing's "beauty and depth" astonishes users, and that he uses it every day to ask relatively deep, relatively hard questions, to see whether its feedback can offer insight. This is the factual foundation for all the judgments that follow — closed-source value, valuation reshaping, application explosion.
— Zhu XiaohuCompute and algorithms aren't the bottleneck — high-quality data is
Zhu Xiaohu says a hundred-thousand-card cluster trained for six or seven months and the results were "probably indeed quite ordinary," so the demands on compute and algorithms are not that high now; the core is high-quality data. He uses a chef analogy: what corpus you train on and what the parameter weights are determine the dish — some are Sichuan cuisine, some Cantonese. He speculates that DeepSeek's beautiful writing and profound answers in philosophy and quantum mechanics may be the team's DNA — they enjoy these kinds of thinking, so the training corpus is higher quality in these areas. He also mentions that the initial corpus needs PhD-level, domain-expert-level people to label it.
— Zhu XiaohuThe data flywheel isn't worth much — his biggest lesson in two years
Zhu Xiaohu says this is his biggest lesson in the past two years: he used to think the biggest moat in this wave of AI might be the data flywheel, but now it seems — including DeepSeek and OpenAI's own research — the data flywheel isn't worth much. The mechanism is that most user data is repetitive, low-information, meaningless; chit-chat doesn't produce intelligence. What truly has flywheel value is high-quality data that requires professionals from various industries to label. So the moat shifts to capturing user mindshare, building customer relationships, and workflow integration — acquiring 20 million users in 20 days without spending on advertising is itself a huge moat.
— Zhu XiaohuClosed-source models are clearly not worth that much money anymore
Zhu Xiaohu says the entire LLM industry needs to reshape its valuation system, because closed-source models are clearly not worth that much money anymore, especially America's OpenAI — if a hundred-thousand-card cluster brings no further breakthrough and only optimizes inference, then domestic players will catch up quickly, and the valuation definitely won't hold. On Trump's announced Stargate project, he thinks that still believes in the old Scaling Law, the era of compute as king, whereas today compute isn't such a big bottleneck, nor are algorithms; what matters more is professional data in various fields, so that investment "is meaningless," plus Elon Musk and others don't have that much money — it's a show performed for Trump.
— Zhu XiaohuIn China, you can assume the foundation model is free
Zhu Xiaohu's advice to AI founders is even more resolute than a year ago: any startup must never research foundation models, especially in China — you can assume the foundation model is free and already powerful enough, so look at what users actually need on top of it and seize the need to provide the best solution. He gives the call-center example — take over the entire call center, no matter how much is done by humans or by AI, and directly offer the client half the current price. He also reiterates his AIGC PMF view: the model is already this powerful, if 10 people can't find PMF, 100 people certainly can't.
— Zhu XiaohuByteDance switching to open source to catch up won't be easy either
Asked whether ByteDance could catch up if it made a fierce push, Zhu Xiaohu says "it won't be easy for them either," and switching to open source right away isn't easy either. The mechanism is that for a big company to do open source, unless it's as thorough as DeepSeek, if it opens like Llama — not so thorough — others still can't use it and may prefer DeepSeek. So he judges that the most important thing for DeepSeek is to keep moving forward, keep catching up to OpenAI, and establish its lead and open-source ecosystem, so that later big companies will find it very hard to catch up. He also suggests that this year we might see Tongyi Qianwen make its ecosystem compatible with DeepSeek, which he thinks would be a more meaningful landmark event.
— Zhu XiaohuAn Agent is just a program — don't be fooled by definitions
Regarding OpenAI's five technical levels from L1 to L5, Zhu Xiaohu says "actually it's just another definition," and an Agent is actually no different from an ordinary program — it's just that you communicate with it and have it arrange tasks for you. He thinks the core is still whether it can truly replace 50%, 80%, 90% of people in certain scenarios, truly without human intervention — that's the most core milestone. His progress report: programming went from maybe 30% last year to 50%, 70%, 80% this year, because the rules are relatively clear; more important is to look at areas where rules aren't so clear, like medicine, where someone has already used OpenAI's Deep Research to write two papers and found them very well written.
— Zhu XiaohuIn their own words · checked verbatim
The data flywheel isn't worth much, because most user data is repetitive, low-information, meaningless.
数据飞轮价值不大 因为大部分的用户数据都是重复的 是低信息含量的 没有意义的
Zhu Xiaohu26:14
For any startup, never research foundation models — especially in China, you can assume the foundation model is free.
任何创业公司来说 千万别去研究底层模型 尤其在中国 你可以假设底层模型是免费的
Zhu Xiaohu36:19
Figures
| Training duration of a hundred-thousand-card cluster | About six or seven months, with no obvious performance improvement | 14:02 |
| Cost for Chinese companies to catch up | At most twelve months later, at one-tenth the cost | 3:08 |
| Cards needed for a multimodal vision model | One to two thousand cards are enough to train | 37:19 |
| Lifetime price of Tongyi Qianwen AI hardware package | From a dozen-plus to twenty-plus yuan | 49:24 |
| Xiaohongshu's first check | $250,000 | 55:26 |
Glossary
- Scaling Law
- The principle that model performance improves as compute, data, and parameter scale grow.
- RL
- Reinforcement learning: a training method that lets models self-improve via reward signals; DeepSeek used it to boost capability at low cost.
- DAU
- Daily active users: the number of users who actually use a product each day.
- PMF
- Product-market fit: a product finding a market need that users are truly willing to keep using.
- Capex
- Capital expenditure: corporate investment in long-term assets such as compute and equipment.
How to listen
Founders and investors watching opportunities in China's AI application layer, especially teams still agonizing over whether to keep building their own foundation models or trying to read where the open-source ecosystem is heading.
After 54:26, the parts about Hangzhou, investment ratios, and the quick-fire Q&A are lower density and can be skipped.