The debate over the speed of technological progress and the speed of technology diffusion.
On September 16, Richard Zhu, Managing Partner of GSR Ventures, visited the Institute of Innovation and Entrepreneurship of Peking University and delivered a keynote speech titled *Venture Capital in the AI Cycle* that lasted for more than two hours.
This investor, who backed Didi and Ele.me and chose to exit in batches at the peak of the humanoid robot craze, directly set the reference frame for AI entrepreneurship back 20 years ago: when the speed of technology popularization and equalization exceeds the speed of technological progress, the only remaining decisive factors are humanities and business.
Let’s first clarify three points:
1. This article is organized by Tide Surge AI based on the on-site record of Richard Zhu’s speech at Peking University;
2. We have focused on sorting out the content of the AI business section;
3. Richard Zhu also talked a lot about the consumer Internet, including the traffic logic of Tencent and Alibaba, the competition between Didi and Yidao, and the history of group buying and Mobike. This part is not included in this article, and you can search for the full record online by yourself.
The core content of this transcribed speech:
Richard Zhu put forward the first framework for judging the timing of AI entrepreneurship (the competition between the speed of technological progress and the speed of technology diffusion), explaining why the customer acquisition magic of cutting-edge models has failed, why open-, and why customers will become competitors after technology diffusion
Richard Zhu, Managing Partner of GSR Ventures
He pointed out three AI scenarios that are already making huge profits (role-playing, reasoning, video), as well as GSR Ventures’ “stretch to reach” investment logic in underwater robots, flexible manipulators and warehouse sorting. In the second half of the speech, Wang Xiaochuan, founder of Baichuan Intelligence, joined the discussion. The two talked from “steering wheel and engine” to “phase transition from the scientific era to the intelligent era”.
The following is the main text of the speech record ——
(Note on compilation: The following is an oral transcription of the speech delivered by Richard Zhu, Managing Partner of GSR Ventures, at the innovation and entrepreneurship class of Peking University on September 16, 2026. It is an excerpt of the AI business section, including the discussion with Wang Xiaochuan, founder of Baichuan Intelligence, in the second half of the speech. Based on the on-site record, it is presented in the first person, and some oral expressions have been streamlined.)
I am very glad to be at Peking University. Last week, Xiaochuan talked about AI here. I want to say that technology is very important, but what is more important is humanities and business.
I have a technical background myself. I studied communication engineering in the pilot class of Shanghai Jiao Tong University, which is equivalent to the Yao Class of Tsinghua University.
To be honest, every technology cycle is very similar: At the beginning of the cycle, what matters is technology, technological leadership and infrastructure.
But there are too many smart people all over the world, and the speed of technology popularization and equalization is extremely fast. Once technology is popularized and equalized, what matters forever is humanities and business.
So why have I always been so enthusiastic about Peking University? Because I think there are countless opportunities at Peking University.
Today we are talking about AI. Many lessons from the Internet and mobile Internet over the past 20 years are worth learning from in the current AI era. After AI technology is popularized, the key performance indicators for entrepreneurship and investment will be very similar to those of the Internet and mobile Internet back then.
Speed of Technological Progress vs. Speed of Technology Diffusion: The First Framework for Judging AI Timing
To judge the timing of AI entrepreneurship, my framework is very simple: which is in the dominant position, the speed of technological progress or the speed of technology diffusion.
In the past two years, technological progress has definitely taken the lead.
At that time, every new cutting-edge model released could attract a huge number of users for free through word-of-mouth communication.
But why does OpenAI say it wants to “pause the frontier” and slow down today?
OpenAI CEO Sam Altman publicly supports “slowing down the speed of improving AI model capabilities”
Because releasing cutting-edge models can no longer acquire customers.
Up to now, the usage rate of GPT-5 has never exceeded 10%. I saw a more comprehensive data today: the usage proportion of cutting-edge models has dropped from less than 50% at the peak to around 45% today.
Why? Cutting-edge models do perform better, but they are too expensive.
Besides, models that are “good enough” can already solve the vast majority of my problems, so there is no need for me to use cutting-edge models.
In the first half of this year, everyone was talking about “token max”, that they had to run everything on the most expensive model first; today everyone is talking about “value max”, and no one mentions token max anymore. Token max costs too much money and creates very little value.
If releasing a cutting-edge model can no longer help me acquire customers or drive revenue growth, but instead exposes my capabilities and allows competitors to catch up easily, why should I release it?
This is the important reason why OpenAI wants to slow down. This is evidence that the large model track has been finalized and latecomers have advantages.
Last year, building a cutting-edge model in the United States cost 500 million to 1 billion US dollars, while in China it may only cost 1% to 2% of that cost, and you can catch up in two or three months.
History Repeats: Solar Energy, Electric Vehicles, and 200 Robot Companies
Once we enter the stage dominated by technology diffusion, history will repeat itself. First movers have to spend a lot of money to go through all kinds of pitfalls; but once they succeed, latecomers can catch up very quickly.
The solar energy industry has gone through three cycles of reshuffling. Do you remember Suntech Power? Back then, Dr. Shi returned from Australia to start his business, listed the company in the United States with a high market capitalization, but it no longer exists today. Who is the leading player now? Tongwei Co., Ltd. What was its main business before? It produced fish feed. Today, there is no technical barrier for solar panels at all. The barriers lie in whether you have cheap capital, whether you have bank credit lines, whether you have enough capital strength to compete with rivals, and who can survive to the end. This is the logic of competition after technology diffusion.
The electric vehicle industry is the same, it has gone through three rounds of reshuffling. Why can Xiaomi and Huawei take the lead today? They have supply chain advantages, channel and brand advantages, and capital cost advantages, which are exactly what startups do not have. Individual startups may have great product diversity, but in this case, startups are basically only proving the feasibility of products for large companies.
The robot industry is the same. There are nearly 200 robot enterprises in China now, and every one of their robots can dance and box. That’s why the stock price of leading humanoid robot companies has fallen recently, which I think makes sense: this market is far too small. I think the vast majority of the 200 robot enterprises may not survive for 10 years.
There are also world models. Yesterday alone, nearly 100 world models were released. Any professor from Tsinghua or Peking University who starts a company dares to ask for a valuation of hundreds of millions, billions or even two billion yuan. To be honest, the technical route of world models is still completely unclear. OpenAI’s GPT-6 may directly generate VLA models. When the speed of technological progress is still unclear, the diffusion speed is already so fast, which makes entrepreneurship very tough. For investors, these are all very high-risk investments.
Open-
Linux went through this path at the beginning of the Internet era.
Before Linux, there were three proprietary Unix systems from Sun, HP and IBM, all bound to their own hardware and sold at very high prices. Why did Linux succeed? It runs on x86 servers and is extremely cheap. After the rise of the Internet, Internet companies all pursue cost performance, x86 became the mainstream, and Linux spent 30 years replacing all Unix systems. Today, only a few proprietary Unix systems remain in very old systems such as banks. This process will be much faster in the AI era, possibly only 5 years.
Where is the logic? It lies in the cloud vendors’ account books.
If cloud vendors deploy closed-enue with model companies; if they deploy open-thier
The token price in the United States has been going down since June and July, because open-5% of closed-e price of open-o raise prices, but because of insufficient computing power
Token price in the US continues to decline
Liang Wenfeng made it very clear himself: if he does not raise prices, he can recoup the cost in 10 months with a pretty good gross profit margin, but he is forced to raise prices because of insufficient computing power.
So today, the revenue of open-, and how much revenue you can generate corresponding to that computing power
I think independent closed-he future
Liang Wenfeng also said a very good sentence: the AI market is so huge that it is impossible for you to take all the profits by yourself. You have to give up 90% of the profits and keep less than 1% for yourself to make the business a real success.
After Technology Diffusion, Customers Will Become Competitors
What is even more terrifying in China is that after technology diffusion, your customers will become your competitors.
Horizon Robotics is an example. Its stock price has dropped significantly after listing. Why? Because BYD may also develop chips on its own. People have done the math: as long as your annual shipment volume exceeds 1 million units, self-developed chips are more cost-effective than buying chips. Today there are many car companies in China with annual shipment volume over 1 million units, and developing vehicle-mounted chips is not that difficult, the technology has been fully diffused. Yu Kai put it very clearly: since you want to develop it yourself, I will cooperate with you and authorize you the IP. But the revenue scale of IP authorization is completely different from that of selling chips: you used to get 100 yuan for a chip, but now you can only get 20 yuan as copyright fee, the gap in revenue scale is huge. The same goes for Cambricon. When it was just listed, Huawei was its biggest customer, but later Huawei developed its own chips.
The same thing has happened recently in the United States. There is a well-known vertical AI company in the US called Abridge, which only does one thing: it automatically records the conversation between doctors and patients and writes it into the medical record system. There are two medical record system providers in the US, and Epic takes 50% of the market share. Epic is not only Abridge’s shareholder and investor, but also its channel. Today Epic says it will provide this function for free. Why? Because Epic thinks that if Abridge continues to grow bigger, it will provide medical record systems for free in turn, so Epic has to strike first. As its shareholder and channel partner, Epic has to provide this function for free to fight against Abridge.
Abridge did start early, it accumulated millions of medical records before, the error rate of its trained model is very low, and it can write medical records with an accuracy of over 95%. But today AI is developing very fast, and Epic also has a huge amount of medical record data. Therefore, developing such an application becomes very easy and simple later on.
We have seen many companies doing the same thing in China. I have told the founders directly: this function will definitely be provided for free in the future. Because to be honest, there is almost no threshold to do this now, the most worrying thing is that you do not have an entrance, you only make a very simple AI function, and you are just making wedding dresses for others.
Chinese Software Going Global: Open
Chinese software companies have had a tough time in the past.
Why is the return on investment of Chinese software companies much worse than that of US companies? The answer is very obvious today: software is equivalent to the service industry, and the currency you charge is the local currency.
US software companies charge in US dollars, while Chinese software companies charge the same amount of money in RMB. This is not because Chinese enterprises are unwilling to pay.
In addition, US software is international by nature, and it is very difficult for Chinese software companies to go global and make sales to large overseas customers.
In the current AI era, I think openell Chinese open-
Last year we invested in an enterprise. After a certain overseas coding tool was launched, this enterprise was the first to make an open-source localized version. That overseas tool is really easy to use, but it is too expensive, and enterprises are generally unwilling to hand over their data. Many recent articles have proved that it will read all the data. Our product is open source, all data is stored locally, and users can choose to connect via API. It reached 100,000 stars and more than 10,000 forks on Github in four months, and forks are more important than stars. It can gain tens of thousands of stars every month at a very fast speed. It achieved nearly 30 million US dollars in ARR just two months after commercialization, and will soon add collaboration and team cooperation functions. It is the fastest-growing Chinese software company.
I also gave an AI doll we invested in to our Japanese LP, and they liked it very much. Why? I said, do you remember Tamagotchi? The Japanese electronic pet in the 1980s was popular all over the world. What technical means could they do in that era? Only pixelated small squares, the product experience was very poor, but it was a global hit.
