Zhu Xiaohu: China's AIGC applications have already exploded — they're just all hidden in to B
Zhu Xiaohu says China's AIGC applications have already exploded, it's just all happening in enterprise services: delivery-rider accident insurance payouts fell from 150 yuan to 60, video interviews hit 1 million sessions in a year, and consumers never see any of it.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The moat is data, not the model
Fancy Tech is the example he keeps pulling out: when he invested in 2022 revenue was just over 10 million, last year it did over 50 million, up five or six times, and it's profitable. The barrier isn't at the model layer — anyone can tweak Llama — it's in data: the US has no product short-video data, US e-commerce is basically still stuck on images, while China converted entirely to short video over the past three years. The second layer is the effect loop: roughly 60% to 70% of customers authorize him to monitor ad performance, so he knows which videos suit Taobao, which suit Xiaohongshu, which suit Douyin, and catching up later isn't easy. The third layer is sales management capability.
— Zhu XiaohuNeither money nor headcount buys you PMF
Zhu Xiaohu breaks the AIGC barrier into two things: finding PMF, and managing sales. The first has nothing to do with headcount or cost — in his words, if ten people can't find it, a hundred people can't find it either, so throwing tens of millions of yuan at a large model is equally futile. His own cadence is to invest 20 to 30 million RMB first, then add more if it goes well. The second scarce capability is sales management: most founders building large models don't know how to manage sales and don't understand commercialization; the AI video interview company he backed raised only one-tenth what its competitors did, the competitors all died, and it is the only company in all of China that can do human-machine double-blind testing, and the only one that can do machine follow-up questions.
— Zhu XiaohuNo scenarios, no data
In the first half of 2023, while the Four Little Dragons crowd was still riding out the cycle, he concluded these large-model companies were hopeless and didn't talk to a single one, including Wang Huiwen, whom he knew well. The reason wasn't technology — no scenarios, no data — and on top of that the valuations were that expensive from the start, a billion dollars in the first round. He judged the outcome might be worse than the Four Little Dragons: the Dragons had an early golden period and revenue climbed fast; large models have no revenue and still have to do research. His most piercing question: if you spend big money to build a GPT-4-level model and someone open-sources it, all that money is wasted — apart from the big platforms, which startup would dare?
— Zhu XiaohuAntitrust has sealed off the large-model exit
Someone asked him how he'd invest if he had 100 million dollars and absolutely had to put it into large models. He said he told Wang Xiaochuan at a conference: without antitrust, he'd be willing to invest in Wang Xiaochuan — solid technology, solid person, decent relations with the big platforms, at least sellable to Tencent or Alibaba. But with antitrust, he doesn't know how to exit, so he doesn't invest. He contrasts it with Dianxin: in the early mobile internet days everyone built localized OSes, Dianxin ended up selling to Baidu and the fund made a little money, because there was no antitrust then. Today large-model companies hoping to be acquired by a big platform find that road sealed off.
— Zhu XiaohuOpen source catching closed source is only a matter of time
His argument is one of engineer orders of magnitude: OpenAI has only one or two hundred engineers, while open source has millions, tens of millions of engineers using it worldwide — how could you stay behind forever? Closed source leads today because OpenAI's technology iteration curve is still fairly steep, maybe a year or even a year and a half ahead; but the curve will flatten someday, and then open source catches up in one leap, just like Android caught iOS. His test is hard-edged too: whether 100,000 cards can still brute-force a miracle. If they can, then sure, impressive; if they can't significantly improve performance, iteration slows and open source catches up immediately. He also cites the Silicon Valley line: the secret is inside OpenAI and Anthropic and a few companies like them, they have no talent drain, so catching up short-term is hard.
— Zhu XiaohuLarge-model pricing hit the floor within a year
This is his most concrete strike against large-model commercialization: in early 2023, private deployment of a large model was quoted at 10 million; by June it was 5 million; by year-end not even 1 million, a central SOE deployment for under 1 million RMB. In one year, the price hit the floor. He contrasts this with the Four Little Dragons: back then monetization was far easier, because the solutions were markedly better than traditional ones and there was little competition — only five, six, seven, eight players in the market, with a two-to-three-year golden window to build revenue first, and only then did they start undercutting each other. Large models didn't even get that golden period.
— Zhu XiaohuThe explosion has already happened, it's just all in to B
This is his most counterintuitive line: applications have already exploded, it's just all to B, so consumers have no idea. The examples are concrete: delivery-rider accident insurance — insurers used to collect 100 yuan in premium and pay out 150, because verifying a claim cost too much; now the rider makes a call and films a walk-around video of the scene, payouts drop from 150 to 60, and insurers start making money. Another: China's traditional marketing calls, a market of four to five hundred billion a year — US carriers are monopolized by FaceTime and don't support C2C video calls, China does, so an ordinary phone call turns into a virtual livestream, you can have a celebrity video-call you live, and conversion at least doubles.
— Zhu XiaohuEveryone who listened to us is still alive today
His one refrain to founders from start to finish: don't burn money. This isn't PR talk — he says he's been telling portfolio companies not to burn money since 2021, those who listened adjusted pretty well, and those who wouldn't listen were the ones who raised too much money and thought they were hot stuff, and now they're facing reality. The way he teaches companies to budget: in the base case assume revenue drops 2% and you can break even; never assume revenue growth, because the moment you assume growth you spend the cost first, and if revenue falls short you're in trouble. He says that's the consensus in the investment world this year: no company should burn money, treat every round as the last.
— Zhu XiaohuIn their own words · checked verbatim
For PMF, if 10 people can't find it, 100 people can't find it either, right? So it has nothing to do with headcount or cost.
就ARG在PMath 你10个人找不到 投100个人 你同样找不到 对吧 所以和人数和成本没关系的
Zhu Xiaohu8:04
Without antitrust, I'd be willing to invest in Wang Xiaochuan — at least he could sell to Tencent or Alibaba. With antitrust, I don't know how to exit.
如果没有反总站的话 我愿意投王小山呢 他至少能卖给腾讯或者阿里的 有反总站以后 我不知道怎么退了
Zhu Xiaohu16:05
This is what I think of as the romanticism of a technical guy. I think this is the most typical romanticism of a technical guy.
这就是我觉得技术男的浪漫主义 我觉得这是最典型的技术男的浪漫主义
Zhu Xiaohu18:06
Right, it's already exploded. Nobody says this, you know? Everyone knows, because it's all to B, so consumers have no idea.
对啊 已经大爆发了 没有人这么说 你知道吗 大家都知道 因为都是土壁 所以消费者都不知道
Zhu Xiaohu28:09
Never assume revenue growth. If you assume revenue growth, you spend the cost first, and if revenue falls short, you've already spent the cost.
你千万别假设收入增长 假设你收入增长 你成本就先花出去 收入万一达不到 你成本都先花出去了
Zhu Xiaohu34:15
If you think you're impressive, fine, but back then you really couldn't tell it was impressive, right? From past experience you couldn't tell it was impressive, from the way you talk you couldn't tell it was impressive.
你如果觉得很牛那也行 但是 英明那时候 真的看不出来很牛 对吧 过去经历也看不出来很牛 说话也看不出来很牛
Zhu Xiaohu42:19
The ones who wouldn't listen raised too much money, way too much money, and all thought they were hot stuff, right? Now they're all starting to face reality.
不愿意听话都是融了太多钱的 都融了太多钱了 都觉得自己很牛逼的 对吧 现在都开始面对现实了
Zhu Xiaohu47:15
Figures
| Share of Fancy Tech customers authorizing performance monitoring | About 60% to 70% | 6:03 |
| First-round valuation of large-model companies | One billion US dollars | 13:04 |
| GSR Ventures' per-deal size in AIGC applications | 20 to 30 million RMB first, more added for good projects | 17:06 |
| Market size of China's traditional marketing calls | Four to five hundred billion yuan a year | 27:09 |
| Delivery-rider accident insurance payouts | Per 100 yuan of premium, payouts fell from 150 yuan to 60 yuan | 28:09 |
Glossary
- PMF / Product-Market Fit
- Product-Market Fit, the state where a product happens to hit a real market need.
- Llama
- Meta's open-source family of large models, which companies can fine-tune themselves; it is the foundation for this batch of application companies.
- POC / Proof of Concept
- Proof of Concept, a small-scale trial a customer runs before formally purchasing, to verify results.
- C2C video call
- A video call initiated straight from the phone's native dialer; in the US this is monopolized by FaceTime, while Chinese carriers support it.
How to listen
Early-stage investors and founders building to B AI applications, especially teams agonizing over whether to train their own model.
44:00-46:30, personal preferences and McKinsey-style training — low information density.