The model is a refinery; the money is in the chemical plant and the car company
If there are two or more foundation models, none of them can make money — they get competed down to the price of electricity. The real moat is context data nobody else saved, plus the last mile that big companies can never do because of how promotions work.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
From an idealistic founder to a founder who can do the books
Wu Minghui says his biggest change over these three years was going from a purely idealistic founder to a founder who can run a business. At the most extreme point, the company's cash on hand hovered around 100 million yuan, while he was burning 100 million yuan a month — one month of runway. Now the company generates its own cash and is profitable. He attributes the shift to being forced to hold the cash-flow line, not to any deliberate learning.
— Wu MinghuiHe built a Copilot in 2020, but not even L1 was ready
Wu Minghui says the product concept the company built at the end of 2020 was today's XGBT and Microsoft Copilot; back then it was called Xiaoming Assistant. But that kind of product needs an agentic model to be ready — by Sam Altman's five levels of AGI, it needs an L3 model, and at that moment not even L1 was ready; chat-style AI itself was still immature. His retrospective: the biggest problem was being too optimistic in his technical forecast, and building too early.
— Wu MinghuiRaised $100M, but only 200M was actually usable
He broke down where the financing went: compensation to historical claimants alone exceeded 200 million, with a third paid out first to former employees; another part went to repaying bank loans — banks hadn't lent for years, and would even trick you into paying off a loan first and then refuse to renew it. So of the $100 million raised, what actually stayed in the company and was usable was maybe 200 million, and cash on hand was never comfortable. A shareholder remarked that he could still ship product with resources this tight, which showed he had become a mature CEO, not a CEO who burns money.
— Wu MinghuiData is what you saved; if you didn't save it, it isn't data
He distinguishes foundation models from the data he collects himself: foundation models grab snapshots for pre-training; what he collects is continuous change, like the daily follower count of a KOL. He uses Tianyancha as an analogy — the data at the State Administration for Industry and Commerce is a snapshot at any single point in time, showing only today's equity structure, but scraping it every day lets you see a company's historical evolution and past disputes. His exact words: "Data — it's a datum, and ju means evidence, it means what's been saved; in many cases, if you didn't save it, it isn't data." When a client came during the 2012 Olympics to look up advertising around the 2008 Olympics, scraping it again in 2012 was already meaningless.
— Wu MinghuiOnce F is smart enough, competition becomes a fight over X
He offers the framework Y equals FX: human decisions come from a good model F and the corresponding evidence X; F is the model, X is the context. When F is already smart enough, competition between companies, between individuals, between countries is ultimately not over F but over X — whose context is stronger, whose data is better. From this he derives: a general foundation model trained on public data cannot make money as long as there are two or more of them; it gets competed down to the price of electricity, because others can distill it. So their strategy is to go into vertical industries and rebuild the data moat from day one.
— Wu MinghuiScaling law hits a size where basal metabolism can't take it
He uses two analogies to rebut the idea that scaling law is unlimited: first, the blood-vessel branching described in the book Scale — the aorta branches down to capillaries in roughly 26 levels, and at the end it gets so fine it approaches a limit; bandwidth has a limit. Second, the Finnish elevator maker KONE: past a certain building height, the cable already weighs much more than the car, the cable is pulling itself, so tall buildings must split into multiple elevators and cannot go higher. His conclusion: "scaling law is a bunch of nonsense — once it gets to a certain scale, its basal metabolism can't take it." He suggests that people doing training should read more life-science books; he is especially fond of Darwin's On the Origin of Species.
— Wu MinghuiBig companies don't do slow-feedback work because no owner can get promoted
Wu Minghui gives an organizational-mechanics explanation: big companies don't do the last mile not because it has no value, but because for slow-feedback work there is no owner in the organization who can get promoted off it. Everyone's first thought is not the company's long-termism but whether they can get promoted next year and how this quarter's review will go. So the smartest people all get pulled into the fastest-iterating work, and people doing slow work never get promoted, so eventually they stop doing it. From this he derives a harsher judgment: competition between big companies is not all the same — the competition between Tencent and Alibaba is very different.
— Wu MinghuiThe software industry's rules are dead in China; models are a bit better
Wu Minghui says that in China, once a big enterprise client makes money, it often demands "give me the source code"; competition in the industry is fierce, and if you don't give it, a competitor will. No matter how good the software is, once the source code is handed over it can be replicated, and departing employees can take it out and start another company. But models are relatively much better, because a model is a black box: you can open-source the parameters, but the data will definitely not be given to you. His exact words: the software industry's rules are completely dead in China; many people hand over source code, everyone thinks it's cheap, and they even think of themselves as cheap. In the AI industry, so far nobody has dared tell him to hand over the model — and to some extent, even if they took it, they wouldn't understand it.
— Wu MinghuiIn their own words · checked verbatim
I think I went from a purely idealistic founder to a founder who can run a business.
我觉得我从一个纯理想主义的一个founder 变成了一个会经营的founder
Wu Minghui0:00
But it's like there was always this不服气 in me — why can't I get this thing done?
但是相当于是我始终心里面是有种不服气的 就是凭什么我不能把这事干成
Wu Minghui1:30:58
The robot era hasn't arrived yet, so I really can't bear to sell these assets.
机器人世代还没来呢 所以我是真的不舍得卖这些资产的
Wu Minghui1:43:02
Data — it's a datum, and ju means evidence, it means what's been saved; in many cases, if you didn't save it, it isn't data.
Data这个事 其实它是一个数据 据就是凭证 存下来的意思 很多情况下 你没存下来就不叫数据了
Wu Minghui1:48:08
So let me give you an example — scaling law is a bunch of nonsense; once it gets to a certain scale, its basal metabolism can't take it.
所以这就我给你举个例子 就是scaling law 是一个特别瞎扯的一个事 就是他大到一定规模 他的基础代谢就受不了了
Wu Minghui2:11:24
Because the owner of that slow-feedback work simply cannot get promoted, so that kind of person doesn't exist in that kind of organization.
因为那个反馈很慢的那个事的那个owner 他根本就不可能得到晋升的 所以他就使得他这种组织里面就不会存在这种人
Wu Minghui3:11:05
We got pass@1, and then topped the leaderboard in one shot, so they really were shocked.
我们就pass at 1了 然后直接一把搜它 所以对方确实是很震惊的
Wu Minghui3:23:09
Figures
| Latest valuation | $1.5 billion | 2:12 |
| Total funding raised | Over a billion US dollars | 2:12 |
| Age | 43 | 2:12 |
| Cash on hand at the most extreme point | Burning 100 million yuan a month | 0:00 |
| Compensation to historical claimants | Over 200 million | 1:44:06 |
| Amount raised | $100 million | 1:44:06 |
| Usable funds left in the company | Maybe 200 million | 1:44:06 |
| Number of blood-vessel branchings | Roughly 26 | 2:11:24 |
| GPUs bought before the pandemic | 1,000 (including some consumer-grade cards) | 3:22:08 |
| OSWorld submitted result | pass@1, topped the leaderboard on the first submission | 3:23:09 |
Glossary
- agentic model
- A model that can plan autonomously, call tools, and execute multi-step tasks, as distinct from a single-turn chat model.
- pass@1
- An evaluation metric for whether the model gets it right on its first attempt, emphasizing no score-padding through repeated sampling.
- MOA
- Combining several models that each specialize, so the whole exceeds a single general model.
- context
- The specific data and background information a model relies on when making decisions; Wu Minghui uses it to mean X.
How to listen
Founders building enterprise AI products, investors watching data moats and the model-competition landscape, and engineers who want to understand the structural problems of China's software industry.
The early details about Mint and the Nightingale acquisition can be fast-forwarded; the core judgments come after 1:48.