Zhipu CEO (Chief Executive Officer) Zhang Peng: Chinese AI startups and large tech companies have multiple relationships, which are relatively complex.

Data strategy is becoming one of the new competitive forces.

On September 5th, Zhipu AI, an AI big model company established for only five years, announced the completion of a new round of financing, with a financing scale of billions of yuan, led by Zhongguancun Science City, which is an investment platform established by the Haidian District Government of Beijing.

This is the third round of financing completed by Zhipu AI this year.

The pre-investment valuation of the latest round is 20 billion yuan, and the post-investment valuation has not yet been disclosed.

Since its establishment in 2009, Zhipu has completed 11 rounds of financing and is currently the highest-valued AI big model unicorn company in China.

Advertisement

The core members of Zhipu AI mainly come from Tsinghua University.

After 2022, Zhipu has opened source multiple basic big models, and its business model is mainly B-end services, including customized services and API interfaces.

Driven by Microsoft and OpenAI, the global AI industry has developed rapidly in the past two years.

Currently, there are 37 generative AI-related unicorn companies worldwide, with 17 new ones added in the past year.

Among them, there are 27 in the United States and 5 in China.

The five in China are Zhipu AI, Baichuan Intelligence, Moon's Dark Side, Minimax, and Zero One Everything.

On September 5th, at the 6th Bund Financial Summit, Zhipu CEO (Chief Executive Officer) Zhang Peng accepted an exclusive interview with the "Finance and Economics" magazine and answered these widely concerned questions in the industry.

Compared with American AI unicorns represented by OpenAI, the scale and development path of China's big model unicorns are different.

The most obvious difference is the business model, where American AI companies use more software models, which can quickly increase in volume; Chinese companies are more about customized services.

Many people believe that this is the main reason for the difference in valuation between the two sides.

Zhang Peng mentioned that this is determined by the market environment.

Doing B-end market in China is "grumbling and doing it at the same time," but this is not without benefits.

Zhipu's investors include investment institutions such as Hillhouse Capital, Qiming Venture Capital, and Legend Capital, as well as many large internet factories such as Meituan, Alibaba, Tencent, and Xiaomi.

Zhang Peng said that the relationship between startups and big factories is complex, with competition and cooperation coexisting, but everyone's understanding of technology is not quite the same.

Zhipu is currently a relatively smoothly developed big model startup in China.

In addition to obtaining continuous high financing, in the first half of this year alone, it won more than 18 government and enterprise bidding projects, ranking first among the five unicorns, even surpassing some big factories.

At present, the financing, computing power, data, and applications of China's big model industry have strong policy guidance factors.

In Zhang Peng's view, there are still many challenges in the development of China's AI industry, and as a startup, it needs to cope with various pressures in a complex environment.

China and the United States have become the two highlands of global AI development, and big models are also called the next generation of technical infrastructure.

The outside world pays attention to the development and application of big model technology.

Zhang Peng and we focused on four topics: the following is the dialogue arrangement.

On the question of the differentiated development of AI between China and the United States, Zhang Peng said: The AI industries of China and the United States have never been the same, and mutual reference and learning have always existed.

The first is the basic situation is different.

This wave of AI big models originated in the United States, and we play the role of a follower.

Although the overall development in the country has been very fast in the past two or three years, in any case, in the most top technology, we are still in the process of catching up.

From the development model, it is also quite different.

The United States converges faster, and China is more prosperous and more diverse.

Asked why China is more diverse, Zhang Peng said: The United States is indeed more clear in the division of labor from the entire ecosystem.

They are more accustomed to ecological division of labor and cooperation, so I said they converge faster.

China has more parallel and vertical.

What I mean is that diversity is seen from a broader perspective, not just from the perspective of startups.

In the era of mobile Internet, China has achieved good development by relying on a large number of application innovations and more popular network infrastructure.

In the development of AI, everyone actually hopes that we can rely on a larger market, more application outbreaks, and the diversity of applications and markets to accelerate our development.

Asked what the specific difference between American AI companies being software models and Chinese AI companies still focusing on projects is, Zhang Peng said: There are indeed these two models, but it is not that the United States only has software and China only has customized projects.

These are two different models, and there is no distinction between good and bad.

The United States' technology development is called "technology first" and "knowledge density first".

Everyone respects basic innovation and maintains a very strong attitude of encouragement and promotion towards new things.

There is a relatively complete innovation and market ecological chain for value distribution between each other, to jointly promote this matter.

In many domestic situations, it is another model.

Everyone has higher requirements for safety, controllability, continuity, and autonomy during the development process.

So in the initial stage, everyone may be more willing to say that I need close service, I need personalized customization.

Many customers' mentality is to hope that technology and data are completely in their own environment, and data needs to be physically isolated, which is related to China's culture and market background.

On commercialization and price wars, Zhang Peng said: This market situation determines that you must focus on customization?

This is a problem that must be considered.

Doing ToB business in China is to grumble and do it at the same time, but it is not necessarily all bad, and there are benefits.

Asked where the benefits are, Zhang Peng said: You can quickly open up a line from technology to market landing, and you can go to the market faster to create value.

Although this may be vertical, you need to spend a lot of energy and resources to do this in the initial stage, which is very tiring.

But the advantage is that as long as one person can go through, everyone is relatively easy to follow and copy.

We also do recognize some problems in this regard.

As a startup, it is impossible for us to do a lot of vertical fields ourselves, and the cost cannot be borne.

You can only grasp the core value in this, and then empower this experience to the entire industry and industry ecological partners, and they will help you replicate.

Asked whether we are now discussing the landing of big models, will the speed be too fast?

The technology itself is still iterating, and it may be that tomorrow's new technology will quickly subvert previous applications?

Zhang Peng said: You are right, this is a very realistic problem.

In fact, we have been discussing this internally for a long time.

For enterprises or users, when they land this technology, they must first think clearly whether your product is highly compatible with AI.

Secondly, because the ability of the model itself has not reached an ideal state, you must do some engineering and customization on it, which cannot be the ability that will grow itself during the model evolution process.

For example, data, evaluation, and the construction of the entire system, team construction, and architectural construction, the ability of the model and the ability of the software should be clearly separated, so that you can retain as much early investment as possible.

No one will tell you where this line is clearly, and you can only explore and continuously rub to find a reasonable opportunity.

Asked why the price war started so quickly, Zhang Peng said: Everyone now cannot find a differentiated value point, and the difference in services provided is not so great, so they can only compare prices.

Asked whether the big model is still very expensive today, what impact will the price war have on the development of the industry?

Zhang Peng said: There are two aspects of impact.

On the one hand, low prices or even free will definitely allow more people to use it; on the other hand, it is still necessary to follow the business logic.

The price war is indeed a great pressure for us.

If one person sells at a low price, others can only follow suit, and I actually do not agree with this approach.

If your technology and product competitiveness are strong enough, you do not need to take the initiative to fight the price.

Asked whether from the perspective of investors, they will pay attention to whether you have big customers, so some companies will first get customers at a low price, and then rely on financing to promote?

Zhang Peng said: We do not do this kind of thing, and there are indeed some people who do this, but the market has a correction ability.

Good customers will pay more attention to sustainable development, and they will pay attention to the strength of technology, products, and partners, and they do not want to do something that is very optimistic and then disappear.

Asked whether many industry-leading companies are doing big models themselves?

Zhang Peng said: From last year to the first half of this year, there were indeed more, and large enterprises were doing it themselves.

But recently we have seen some turning situations, and they found that this matter is not so easy.

It is not a matter of forming a team and taking an open source model to run through.

It is better to purchase.

Asked whether for startups, it is necessary to continue to invest in technology and to commercialize, will the burden be too heavy?

Zhang Peng said: It is indeed a challenge.

However, AI technology has come to today, which is actually an interdisciplinary field, and it is very particular about engineering practice.

You can't just study in the laboratory every day and ignore the actual market demand.

So when the technology reaches a certain level, you can explore some reason scenes, collect feedback, and promote rapid iteration forward.

You need to know what the market needs.

Of course, this is a process of weighing and pulling, and you can't completely follow the market requirements, nor can you imagine it completely.

Asked how to balance, Zhang Peng said: I often chat with frontline customers, what they want to do.

But I will not do whatever they say.

I especially like to ask a question: Why do you want to do this?

We need to find the essence behind this matter.

For example, smart city projects, most of the time the first step is to summarize all the data together.

But when you go back to the origin to see, the purpose of doing a smart city is to make the experience of people in the city better.

Is it necessary to collect all the data?

The cost is very high, how to efficiently process these data, how to protect data security, it seems to bring more problems.Now that we have more advanced AI technology, we can use new technologies to avoid old problems and forge new paths.

Question: After dealing with clients, do you have to deal with the technical department back at the company?

Zhang Peng: This kind of situation is very normal, and enterprises will encounter such problems.

The client wants to do this thing, but the technical staff say it's meaningless.

If we talk about it from the perspective of technological evolution, these are not problems, what we actually need is just time.

There are better solutions to this problem, but the client can't wait, and the market can't wait either, what to do?

We can only use some engineering methods to solve the problem first.

Regarding the relationship with big tech companies, Question: Currently, in public information, Zhi Pu has won the most bids among startups, why is that?

Zhang Peng: I think the main reason for winning so many bids is that the clients affirm our core technology.

They don't have to worry about us suddenly disappearing, because many companies actually rely on open-source technology from abroad.

What if it's not open-sourced one day?

Clients who think about this kind of problem may come to us for cooperation.

Question: Compared with big factories like Baidu, Alibaba, and iFLYTEK, what's different about you?

Zhang Peng: First, our core technology is completely under our own control.

Second, we don't regard AI large models as a tool for making money.

People have different understandings of this matter, and there will be differences in the direction of investment, resources, and the final results.

I can't say what the goals of big factories are, but I feel their goal is to have this technology be sufficient, and it's enough to make money, not necessarily to push the technology to what extent.

Question: Don't you want to make money?

Zhang Peng: It's not that we don't want to, we will balance it.

Making money is to pursue higher goals, to prove that this matter can be done, so there will be some differences in cognitive logic.

Question: Alibaba is one of your investors and also develops large models by itself.

What is your relationship with Alibaba?

If we compare it with the relationship between OpenAI and Microsoft?

Zhang Peng: As I mentioned earlier, there are some differences between China and the United States.

Big factories in China tend to do everything themselves.

They have resources, talents, and data, so it's more difficult for everyone to cooperate (from the perspective of startups).

In the early years, Microsoft, in order to support OpenAI, withdrew its internal large model development team, but now it has started again.

Microsoft is a builder of the ecosystem.

He doesn't think he has to do everything himself.

If the ecosystem is better, I will naturally be better.

This is a difference in cognition.

On the other hand, the market-oriented mechanism led by capital in the United States gives him the confidence to dare to do this.

Microsoft invested $10 billion in OpenAI, which seems a lot, but compared with his returns in the secondary market, the cost can be covered in a minute.

Question: So for you, are big factories just a simple investor?

Zhang Peng: It's not just that, there are competitive relationships, investment relationships, and technical cooperation relationships.

They didn't invest in us out of thin air, but because we have long-term various kinds of cooperation, they know what we are doing, and understand our capabilities.

It's a more complex relationship.

Regarding the impact of policy on the AI industry, Question: To this day, is computing power still a bottleneck?

Zhang Peng: It is an important influencing factor, but in the long run, this matter is solvable.

You can see that national policies and capital are actively solving this matter, but the effect may still need time.

Question: For example, various local governments are building computing power clusters.

What direct impact does this have on the large model industry?

Zhang Peng: Macroscopically, it is definitely beneficial, and we do rely on large resources to do this.

There is a view that everyone is worried that it can't be used after it's built, there is redundant investment, or it is very scattered.

In fact, many times it's like flying a kite, let it fly first, as long as the line in the hand doesn't break, the problems that arise in the process can naturally be solved.

For example, some people ask me, the depreciation of AI chips is only 4-5 years, and the update speed is very fast, why do you have to spend so much money to buy computing power, why not wait for new chips to come out and then buy?

Although these assets have a certain service life and are updated very quickly, these resources will not be wasted.

When there are new versions of chips, the old ones can be used for inference or other things.

Technological progress brings more choices, and you can't just focus on the present.

You can't design everything perfectly based on the current situation.

Some people say, what can be done with so many computing power centers?

For example, a cluster of 1000 cards, a larger model can't be done.

But do we definitely need every model to be so large in the future?

Can computing power centers only be used for model training?

Will training and inference definitely be separated in the future?

It's not necessarily the case.

Of course, I'm not encouraging everyone to invest in computing power, after all, energy consumption is also a problem.

Question: Apart from computing power, is data a problem?

Zhang Peng: Data is a very important factor restricting the development of China's AI industry.

A large amount of our data is privatized, or has very clear IP restrictions.

The cost of processing data is indeed too high, and everyone is still hesitant about the investment.

However, when the technology has developed to a certain extent, the model itself has been leveled out due to open source and other reasons, and at this time, data will become a more critical competitive advantage.

So people are not willing to come out and say their data strategy, including data ratio, data source, data processing, etc., which has become a part of the enterprise's competitive advantage.

From a policy perspective, the country has been emphasizing the role of data elements, and the industry will also pay more and more attention to data.

We can see that many people have started to start businesses or other related work in this area, and the data problem will also be improved in the future.

Question: So what impact has the country's overall policy brought to the large model industry?

Zhang Peng: In the current environment, it is very necessary for national policies to pay attention to it.

In the future, for a long time, AI will definitely be the core of economic development and digital and intelligent transformation.

But there are many tests and challenges in the specific implementation process.

Are we going to buy cards to build a machine room first, or support technological innovation companies to do research and development, or do we not need to catch up with the forefront, only focus on doing applications?

Everyone has different ideas here, and they all meet the policy requirements, and there will be different paths in the implementation process.

Everyone is in a different position and has different cognitions of new things, and can only do their own territory first.

From our perspective, when talking with investors, everyone has obviously become more focused on AI, and the policy will also lead some customers in, such as the financial industry, which has obviously increased its investment in AI due to policy influence, and overall it is a good thing.