The world's first listed large-model company: Zhang Peng says there is no bubble in Chinese AI
There may be a bubble in the US capital markets, but not in China — China's AI spending is only about one-twentieth of America's, nowhere near enough. Zhang Peng's logic: we're still far from AGI, the money isn't enough, so we have to invest more; it won't happen on its own.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The first to eat the crab spent two years negotiating the restructuring
In January 2018 the state issued a directive allowing staff at research institutes to commercialize existing research results, and specifying that the returns be split between the original institution and the founding team. But how the ratio would be set and what the results were worth had no black-and-white answer — everything had to be negotiated. Zhang Peng says the state opened ‘a window’, not a wide-open door. As the first person from Tsinghua's computer science department to take this path, the university didn't know what to do either; the dean of the research institute and the senior colleague in charge of tech transfer sat in on the talks. From 2018 to the company's registration in June 2019, a year and a half was spent mainly on two things: how to split the returns, and how to assess the monetary value of the results — ‘in the past it was never judged with money’, at most people would say what level it was at.
— Zhang PengPerceptual intelligence is just a skill; cognitive intelligence has a brain
Zhang Bo drew a two-dimensional quadrant to distinguish the two generations of AI: the problems today's AI solves cover a very small range; one layer out is ‘I know that I don't know’, and further out is ‘I don't know that I don't know’. Perceptual intelligence solves single skills like recognizing images or understanding speech; in Zhang Peng's words, it ‘isn't actually a brain, it's just some skills’. Cognitive intelligence requires generalization — after learning from limited samples, it can transfer to situations it has never seen, draw inferences, including learning, logical reasoning, error recognition and self-correction. He uses driving as an example: someone who can drive a manual can quickly pick up an automatic, with no one teaching them — that's learning, feedback, trial and error, then generalization. But he thinks this problem is still unsolved today, because cognitive science and brain science themselves have no answer to what the essence of cognition is.
— Zhang PengThe investors' first question was how do you make money
In August 2022 Zhipu open-sourced GLM-130B, and Zhang Peng then went fundraising. He told investors they had trained a model comparable to GPT-3; the reaction was ‘I don't get it, I really don't get it’, and all the questions were about how you make money and how you commercialize this thing. One investor, chatting online, asked whether they could make money, and said look at how bad the environment is, how bad the economy is, why don't you halve your valuation. Zhang Peng didn't cut it. That round dragged on for nearly half a year, until ChatGPT took off in China — ChatGPT coming alive helped them a great deal; no one had to question what this thing was anymore, and later it flipped to investors asking when they could build that ChatGPT thing. He admits that period was quite hard.
— Zhang PengSculpting training cost down to one-fourteenth
GPT-3's training cost was analyzed externally at over 20 million US dollars. For Zhipu's training of GLM-130B, by Zhang Peng's account compute cost only 4 million RMB, and with labor about 10 million RMB, roughly one-fourteenth of the former. He attributes this to the advantage of Chinese teams — they dig into details, squeeze efficiency out of details. Cost reduction is a continuous main line at Zhipu: they were also the first to do inference of a hundred-billion-parameter model on consumer-grade cards, bringing the cost down from over a million to two or three hundred thousand. GLM 4.7 still has only just over 300B parameters; it's good not because the parameter count doubled again, but because training efficiency and data utilization are higher, and the model architecture and parameter count were designed with inference-time cost control in mind — 8 cards on a single machine is enough.
— Zhang PengOpen source is not free; only time can prove it
After DeepSeek went fully open source, many customers equated open source with free in their minds — you've open-sourced it and it costs nothing, so why should you still charge me. Some customers had already approved budgets, then said the open-source one is no worse than yours, maybe even better, so can we skip buying. Zhang Peng's answer: using open-source stuff is fine, but that's not the same as what you actually want, which is a commercialized service — don't mix the two up. Quite a few customers deployed DeepSeek all-in-one machines themselves, or got someone to help deploy them, and after a while most turned back, because the original vendor doesn't provide commercialized service and can't integrate with internal systems. His line is that only time can prove it, and this did delay them with some customers.
— Zhang PengThe internet bubble burst, but the infrastructure stayed
Asked about the view that large-model companies going public is a great escape, Zhang Peng retorted: how do you define bubble. His first layer of logic: suppose it is a bubble, going public won't save you — that's not saving me — so listing and whether the bubble bursts have no necessary connection. The second layer is an analogy to the internet: the bubble burst, and what remained is the network infrastructure, technological innovation and many products people use today. He thinks what's really feared is not that investment produced no real productivity, but whether I can recoup my investment returns in time. Hence his conclusion: from a capital-markets perspective America may have a bubble, but in China ‘this thing doesn't exist, it's not enough, it's far from enough’ — China's investment is about one-twentieth of America's, and much of it is spread across infrastructure, unlike America where it's concentrated in a few leaders.
— Zhang Peng50 people, 200 people, 500 people — every hurdle can be fatal
Zhang Bo told the team that startups have several hurdles: the 50-person mark is generally passable, as long as you make money, profit or loss isn't the key, the key is that team confidence doesn't scatter first; at 200 people division of labor starts to appear — commercialization, R&D, product, daily operations — and division of labor brings management costs of communication and alignment; if coordination fails, people can each mind their own patch and the company falls apart; at 500 or more, layers and middle managers appear, information transmission gets longer, alignment gets harder, compliance and security get more troublesome. Zhang Peng didn't quite understand at the time, but after experiencing it himself he finds it very true — the key isn't the specific number, but the stage of the company's development. He gives a picture: back when they were at Kejian with just over a hundred people he knew everyone; after moving to this two-floor office, there was a batch of people in the company whose names he couldn't call.
— Zhang PengNo big launch event, yet high praise overseas
When GLM-4.5 was released in July they held a launch event for a few hundred people; for later versions they did almost no promotion, with only a few dozen people on site, relying mainly on online release and open source, but online — especially overseas — the reviews were very high. Then a series of things Zhang Peng calls magical happened: US companies directly used their models; Windsurf took down its interface and put in Zhipu's model; and manufacturers took their open-source model to distill and prune, stuffing it into their own models to serve customers. His summary: this is to some extent also the effect of DeepSeek teaching everyone — no matter how you boast or how you promote, in the end it comes back to the actual effect it should have. People are also a bit tired of big launch events now.
— Zhang PengIn their own words · checked verbatim
A window was opened, and everyone saw that there could be this path outside, but it wasn't a wide-open big door saying go ahead, walk wherever you like.
开了一盏窗户 就大家看到了 外面还可以有这一条路 它并不是开了一个很敞开很大的一个门 说你随便走吧
Zhang Peng15:06
I remember very clearly, one investor chatting online said, this thing, can you turn it into money? And now look at how bad the environment is, how bad the economy is, why don't you halve your valuation, how about that.
我印象特别深刻 有个投资人往线上聊的 这样东西 你们能变人钱吗 啊 现在你看这个大环境这么差 经济这么差 要不你们把估值降一半 怎么样
Zhang Peng43:19
In many customers' minds, they put an equals sign between open source and free. The impact on us is: you've open-sourced it and it costs nothing, so why should you still charge me?
很多客户的脑子里面 他就把开源和免费 就花等号了 给我们带来的影响 就是说你那都开源都不要钱了 你为什么还要收我钱呢
Zhang Peng1:35:59
Even if the internet bubble burst, what did it leave behind? Much of what people use now, much of what they enjoy, is what that bubble era left behind.
就算互联网那个泡沫破了 它留下了什么 现在大家用的很多的东西 享受的很多的东西 是那个泡沫的时代留下的东西
Zhang Peng1:56:12
Will it happen naturally? It won't. So that's that — you still have to invest.
它会自然发生吗 它不会 那不就完了吗 所以还得投资
Zhang Peng1:57:12
We can have very lofty ideals, and we will never give up on those ideals, but it's not that we have only ideals and no idea how to pursue them.
我们可以有很远大的理想 这个理想我们一直不会放弃 但是又不是说我们空有理想 也不知道怎么去做
Zhang Peng2:13:19
Opportunity always favors the prepared. Even if you're drifting at sea and a piece of board floats past you, you still have to flail a couple of times to grab it.
机会永远是留给有准备的人的 就哪怕是你在海上飘着 有一块布板从你沿前飘过 你也要扑腾两下才能把它抓住
Zhang Peng2:24:24
Figures
| Ratio of China's AI investment to the US | About one-twentieth | 1:57:12 |
| Customer structure | 9 of the top ten internet companies are customers, 60% are enterprise customers, government accounts for 20% | 1:54:10 |
Glossary
- MaaS / Model as a Service
- A concept Zhipu proposed around 2021, then covering cloud API, on-premise deployment, software-hardware integration and other forms.
- Test Time Scaling
- Spending compute at the inference stage rather than the training stage to improve model performance.
- VLA / Vision-Language-Action model
- Unified modeling of vision, language and robot actions, used to control robots.
How to listen
Investors and founders watching Chinese large-model commercialization and listing paths; practitioners who want to understand the trade-offs between open source and 2B/2C.
The first 15 minutes on Tsinghua lab history move slowly; you can start from the negotiation over tech transfer (15:06).