In the AI Era, Networks Have No Natural Monopoly and Marginal Cost Doesn't Approach Zero
Internet products are giant towns: marginal cost approaches zero, so the winner takes all. AI costs one more token for every extra user, and that difference will rewrite the business design of traffic nodes.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Portals had already divided up the first wave of traffic before 2005
Zheng Qingsheng (郑庆生) cuts China's twenty years of venture capital into three phases: before 2005 was the portal era, when Sina, Sohu and NetEase occupied the first wave of internet traffic, and the landscape was already set before 2005; from 2005 to 2010, Tencent, Baidu, Shanda and that cohort seized the portal-era traffic, carved it up, and rose; after 2010, mobile internet opened new traffic entrances, the old entrances ceded a large share of the market, and a new wave of companies reshuffled the deck. He stresses that although the big traffic was held by the giants from 2005 to 2010, community-type innovation never stopped — it just couldn't capture the main traffic.
— Zheng QingshengShort video doesn't challenge long video, it challenges text itself
Zheng Qingsheng's judgment is that short video should never have been compared with long video at all — it has become a technological way for humans to understand the world. His argument: text is a very advanced form of knowledge product, and it takes a person a long time to master it, which is why every country has to run literacy campaigns; images are far simpler than text, and mixed image-and-text layouts tend to crowd text out. He gives the example of primitive humans — writing a passage to tell someone the moon is very big and very round is a very advanced move; the simplest way is to drag the person over there and look. So short video isn't challenging other media forms, it's challenging text itself, and he treats short video as a peer product to text.
— Zheng QingshengHumans' new behavioural patterns are, on the whole, unpredictable
This is the only generalisation Zheng Qingsheng draws from product history. He says people originally treated short video as a small category — you watch long video, I watch short video, he looks at O2O — but in the end short video's traffic was so strong that it could challenge many companies, even challenge text, and nobody knew this in advance. He likens investors to literary critics: I can't write a novel, but when you write one, my instinct tells me this is probably a masterpiece. Founders are the same: they can't provide a rational forecast or the basis for one, but they have instinct, and instinct can rise into conviction.
— Zheng QingshengAI's network has no natural monopoly, and marginal cost doesn't approach zero
Zheng Qingsheng gives three ways AI traffic nodes differ from the internet. First, the internet is a network with natural-monopoly characteristics, and marginal cost approaches zero — you're already at 1 million DAU, and the 1,000,001st user costs you nothing; but AI is constrained by token consumption, and every additional user is money, so business model design becomes different. Second, the old products provided some kind of service, attracted users to click, and step by step helped them reach an outcome, whereas efficiency-type AI goes straight to the outcome and is outcome-oriented; how well it works depends more on whether the provider can offer an outcome-oriented solution. Third, historically the principle level of technological revolutions has never been a black box; how this emergence happens and what changes it will bring are currently unclear.
— Zheng QingshengDeep digitalisation will bring new hardware opportunities
Zheng Qingsheng distinguishes two kinds of digitalisation: the old kind moved easily counted, already structured data online; this round is deep digitalisation, processing information that previously could not be processed. He gives the example of an automatic camera device from Microsoft Research — worn on the chest, it takes a photo whenever the light changes, producing tens of thousands of photos a month that you cannot process, so under the technical means of the time this kind of digitalisation was meaningless. Today AI can process it into meaningful online content, so he thinks hardware that provides deep digitalisation could become a new information node, a new traffic node. A phone can't do this, because the most important thing about deep digitalisation is that you have to be present all the time.
— Zheng QingshengThe AI era has no explicit two-sided network effects yet
Zheng Qingsheng thinks large models themselves have an implicit two-sided network effect — the more people use them and the more data there is, the better the answers new users get. But at the explicit level there is none: the old two-sided network effect required that when you speak someone listens, when you sell something someone buys, and it had to be assembled bit by bit; now everyone talks to a large model, a many-to-one relationship. So he observes that the first wave of social-leaning companies to come out in the US were not very successful — essentially there is no two-sided network effect, and once model quality declines while a rival's model quality rises, you're in a commercially unfavourable position. He states plainly that at present this is unclear, including whether OpenAI adding group chat will hit Meta; he thinks there is still a gap in the logic.
— Zheng QingshengWhether value settles in models or applications, he can only answer with conviction
Asked whether, after investing in Kimi, MiniMax and Manus, value ultimately settles in model companies or application companies, Zheng Qingsheng says this is the same question as the last era's debate over whether there would be new applications or whether it would all be the giants' era. Because he does early-stage investing he is optimistic, and firmly believes new innovation can eventually become a giant within one cycle, adding a few more names to the list of giants — it's just uncertain whether that's more than five or fewer than five, or fewer than three. He concedes that large model companies have already been added to the list, that what forms can be added he doesn't know, and finally says this can only be answered with conviction.
— Zheng QingshengThe CEO is the personified symbol of the organisation and its institutions
Zheng Qingsheng divides founders into two kinds of gift. The first is product sensitivity; from one to ten, product sensitivity matters more. The second is being able to become the personified symbol of the organisation and its institutions — he uses the general as an analogy: in ancient times a campaign had no advanced means of transmitting information, yet among 100,000 or 200,000 men he was the highest institutionalised symbol, maintaining morale and military discipline, serving as the organisation's imagination, firmly believing the organisation has an abstract life and that you are the representative of that abstract life. He says someone who hasn't made it as a very successful CEO may be a successful product manager but ultimately can't complete the leap, because this step requires a gift, and people with both gifts are rare. Even if you can't do it, you may still have to play the part.
— Zheng QingshengIn their own words · checked verbatim
So in the end, short video — I think this form of short video is no longer challenging competing companies, no longer challenging other media forms; what it challenges is text itself.
所以最后短视频 我是认为短视频的这种方式 它已经不是挑战 竞争公司 不是挑战其他的媒体形式 它挑战的是文字本身
Zheng Qingsheng31:18
Humans, whether it's the subconscious or whatever new behavioural pattern, are on the whole unpredictable.
人类无论它是潜意识 还是怎么样的新的行为模式 总体上是不可预期的
Zheng Qingsheng40:23
In all naturally monopolistic networks, marginal cost is missing — what marginal means is that the cost of each additional new user is zero.
在所有的自然垄断的网络里 边际成本都缺 什么边际就是 等于说你每增加一个 新用户的成本 是零
Zheng Qingsheng1:01:31
The important essence is that it goes deeper and acquires more of what used to be unimportant, or completely unimportant, information in the offline world — information that would otherwise drift away with the wind.
重要的本质是它更深入的 在线下世界里面 获取了更多 原来不那么重要的 或者完全不重要的信息 会随风飘散的信息
Zheng Qingsheng1:08:33
Figures
| Year Zheng Qingsheng started learning to program | 1984 | 4:05 |
| Length of time Zheng Qingsheng worked in auditing | three years | 12:07 |
| Length of time Zheng Qingsheng worked in management consulting | three years | 12:07 |
| Share of projects where Sequoia China was the first institutional investor | more than half | 3:04 |
| Year Zheng Qingsheng first met Zhang Tao | 2011 | 23:13 |
| Year Pinterest was founded (Zheng Qingsheng's recollection; he says he apologises if he misremembers) | 2010 | 27:16 |
Glossary
- two-sided network effect
- A moat in which users on both sides of a platform attract each other and which has to be assembled bit by bit.
- PMF
- The state in which a product matches market demand and the model works.
- survivorship bias
- Seeing only the samples that survived and ignoring those that vanished, which distorts judgment.
How to listen
Investors looking at consumer and AI applications, and founders building AI products — especially anyone who wants to understand the structural differences between AI traffic nodes and the internet.
The career autobiography from 2:00 to 17:17 at the start; low information density.