Liang Wenfeng Investor Meeting (Full Transcript)

Liang Wenfeng

Welcome, everyone. When we first started this company, our original intention was not about how much money we would ultimately make, going to the capital market, going public, or anything like that. We didn’t have that motivation. The first few dozen people never thought about it that way. If they had, they wouldn’t have come.

So generally speaking, we approached this endeavor with a great deal of goodwill towards the world. We felt this would be useful for humanity—something beyond monetary gain. Of course, later on, as the potential benefits became enormous, other temptations arose, which is a separate matter. But our starting intention, our vision, and the vision we maintain to this day are not pursued in a way that maximizes commercial interests.

I think this is key. About twenty years ago, the person I most admired in management was Jack Welch, the former CEO of GE. Looking back now, most of what he said may no longer hold true, but one crucial point he got right: the most important thing for a company is its vision. Managing a large company doesn’t rely on your rules and regulations; it relies on vision. What is vision?

Vision is not a slogan hanging on the wall. Vision is what you do, not what you say—it’s how you actually operate. I’ve forgotten Jack Welch’s exact words, but that’s the gist. So how do we manage so many people and how are we organized? In truth, we have no organization. We are driven by vision, organized by a shared vision. We are organizationless. This has both advantages and disadvantages.

In the future, we will find ways to leverage our strengths and mitigate weaknesses, but this is our characteristic. We don’t operate by saying “I want to achieve this KPI” or by having assessments. We only have vision. This vision isn’t even formalized—it’s not written down. Nothing has ever been written. This vision is embedded in how we do things, in our attitude toward the world.

Everyone in our company may understand this vision differently. Each person’s vision may have nuances, but there is alignment in the broad direction.

I still believe it comes from a place of great goodwill toward the world, wanting to do something meaningful. This is how we organize. I’ll speak first, then we’ll open up for questions. I’ll probably frame everything I say around this vision. The vision is real, not fabricated. We genuinely think and act this way. Otherwise, you can’t explain many of our actions. Why do we insist so strongly on open source?

Because the vision itself demands open source. Without this vision, you can’t rally people. For example, **also open-sources, but their open source is different from ours. Their open source feels forced, as if it wasn’t their original intent. For us, it is our original intent. And on the matter of open source, we were very clear from the start.

First, there’s the vision; second, we believe that to commercialize AI successfully, open source is beneficial. This sounds contradictory and counterintuitive, because historically open source and commercialization have been in conflict. But I think AI is different.

Historically, a software company’s market might be a few billion USD a year. If they open-source it, the market might shrink to a few hundred million or a billion. But AI is huge enough that it could eventually account for, say, 10% of global GDP. That’s a massive number.

You can’t monopolize this. You have to share; otherwise, you definitely won’t survive. This is different from previous open-source software where the market wasn’t that big. But AI is too big. If we try to monopolize the benefits, history will abandon us. I think this is an objective law, a historical perspective.

It’s not that if I don’t open source, I can monopolize the market. That simply doesn’t align with objective reality. You will face many obstacles, and other methods will prevent you from achieving that goal. In this situation, you don’t necessarily have to follow traditional business thinking. You need a mechanism to ensure that the benefits you capture are limited, so that you can succeed. You need restraint. I think restraint is necessary.

If we want to succeed in AI on our watch, the first thing is restraint. You can’t think that a certain percentage of humanity’s GDP or China’s GDP belongs to us. The more you think that way, the less likely you are to succeed.

So from the outset, we felt restraint was necessary. The more restrained you are, the more likely you are to succeed. This is a business consideration, a macro-level one. I think it’s intuitive—at least it’s intuitive to me, or at least I genuinely believe it. We don’t have many other advantages. We’re not special. We’re not richer than others, nor do we have better people than other companies. We don’t.

Think about it: when we founded this company two years ago, we didn’t have much money, many GPUs, any reputation, or appeal. We were just a group of very ordinary people.

We really are a group of ordinary people. If there’s a narrative we like, it’s that ordinary people have done extraordinary things—not that geniuses did extraordinary things. This is closely tied to our restraint and our vision. So, does open source conflict with commercialization? I think in the case of AI, if you’re not restrained, you won’t succeed.

Open source is part of restraint. Our restraint isn’t only reflected in open source; it’s evident in many aspects. But generally, we don’t need to think about open source or restraint. The more restrained you are, the easier it is to succeed—at least this has been proven so far, and it’s logically consistent.

Otherwise, there’s no way to explain how we succeeded: we had no weapons, a very low starting point, very few resources, and our people were just a random group of ordinary individuals. I myself am just a university graduate, not from the top-tier school. This restraint is also part of our vision. AI is too huge, the benefits are too large.

We are very restrained. As long as we succeed, the benefits will be enormous. You can just take a small portion, and it will be plenty. So there’s no need to think about which portion to take or how to take it. I think we shouldn’t even consider it, because the benefit is big enough. Taking just a little is sufficient. That’s why we said earlier: we only take a reasonable profit. It depends on your intention, not the size of the profit. That’s different.

This is not about our API pricing. Our API pricing is based on what we consider a reasonable profit: roughly recovering the cost of purchased equipment in ten months. We think that’s reasonable.

Under current circumstances, considering risks and upfront investments, if we depreciate a server over three or five years financially, but commercially we feel recovering costs in ten months is enough—yes, that’s enough. So that’s the logic behind our current API pricing. Our V3.2 Flash and others all recover equipment costs in ten months.

That’s our standard. It’s not profit maximization. If it were profit maximization, we would set higher prices. Because in this price range, user demand is inelastic: even if we double the price, token consumption wouldn’t change much. If we double the price, total revenue would nearly double. Wait, let me think. Oh, that’s great.

Let me tell you a story about one of our models. Initially, we worried about too much demand, so we set a relatively high price. The team wasn’t happy. Later, I lowered the price to a quarter of that, and everyone was delighted. I think that reflects our true thinking.

It goes back to the vision I mentioned earlier. We want this to be useful for people—not to maximize our profits, but to ensure that while we earn a reasonable profit, everyone can afford it. I think that’s the sentiment among others in the company at the time. When we lowered the price, many people in the company group chat cheered. Everyone was very happy. Because this is the purpose of all our hard work and dedication to making this model excellent.

The purpose is to make it very cheap, very effective, so that everyone can use it fully. We find that very fulfilling. That’s our motivation, our vision, the consensus that brings our company together to do this. That’s the consensus inside our company. This is probably quite unique, because for our competitors, lowering prices is certainly not a good thing—they definitely wouldn’t cheer.

Because your revenue and ARR would drop by half if you cut prices in half. Yes, that’s a difference. We think this is enough. Internally, recovering costs in ten months is already very satisfactory for me commercially. Externally, we believe this price makes everyone happy and willing to see. It’s a win-win for the company, society, and everyone.

I think, okay, someone commented on the screen saying that a ten-month payback period means profit is too high. Indeed, there is still room to lower prices. There is also room for model optimization, so overall price reduction potential is still significant. But this cost—ten-month payback—we can achieve it, but others cannot. For example, Alibaba or Tencent may not have our optimization; their costs should be several times higher.

There is still a lot of optimization work to be done. Why don’t we keep lowering prices? Because demand is inelastic. If I lower the price further, demand won’t increase much, or it will increase very little. Because at this price, everyone can afford it. People are satisfied with this price and won’t stop using it because it’s expensive.

So lowering prices won’t bring the company more revenue, and it won’t add much social value either, since everyone is already satisfied. If the price goes even lower, it won’t significantly increase societal happiness. Okay, so on this pricing issue, we are definitely not starting from the point of maximizing company revenue or profit.

This is part of our restraint. Because in the short term, higher prices might mean more revenue; but in the long term, it’s hard to say. I think restraint is a strategy. To me, restraint is a strategy. Sometimes you can give up something to gain something else.

Open source is the same. It can be seen as our pressure or our concession. First, this concession makes our company happy. Employees feel a sense of achievement, and it strengthens our cohesion. And this concession benefits society. Society, other peers, and ordinary people are all happy.

So I understand this restraint, in the long run, increases the probability of us achieving AGI. When considering something, I have no doubt that AGI will have enormous commercial value. On that basis, my priority is not how to increase my share or how to get a larger slice. My priority is how to increase the probability of my success. This restraint may also be reflected in many other aspects.

For example, last Spring Festival we suddenly had a surge of users, but we didn’t chase after retaining those users, monetizing them, or grabbing commercial benefits from them. We didn’t compete for users or try to make money. But we worked hard to serve them well.

We never thought, “I want to become the next super app, compete with someone, become the next ByteDance or Tencent.” Not at all. We could have, but we didn’t. My understanding is that this is also part of restraint. Don’t try to grab everything. If you have users, it seems you could become the next ByteDance and consume everything.

I think that is commercially viable and possible. Last year, if we had spent a lot of money to compete with ByteDance for users, that would have been a strategy. But we chose a very restrained approach: we don’t compete for this. Because there is a bigger watermelon later; what’s ahead might just be sesame seeds.

I shouldn’t grab all the sesame seeds. Of course, maybe the sesame seeds are big, but I think the later AI opportunity is much bigger. Looking back, it might have been right that we didn’t go all-in on the consumer side last year. Because we can see there really is a bigger watermelon behind, and what’s ahead is just small sesame seeds. If I had a lot of money last year and made things very big, what benefit would it bring? You wouldn’t have gained anything.

These are my real thoughts. Because I think the AGI opportunity behind is extremely big. The AGI opportunity is always extremely big.

I don’t even need to consider whether I will have a position in it or what my business model will be. We don’t need to think about it at all. As long as there is such a huge commercial opportunity, you will find a way. As for the sesame seeds ahead, we will pick them up casually. But we won’t stop or treat it as an important matter. So last year’s consumer DAU, I think is probably a small matter.

But we did pick it up. We maintained user usage at a relatively low cost, because it might be useful later. Although we don’t know what use these users will have now—it’s purely a cost—it might be useful later. Since we can get it easily, we will get it. Looking at this year, it’s very possible that our ARR from API or AI will also be an opportunity.

If demand continues to expand, and if we can buy more GPUs, then achieving an ARR of several hundred million USD is very possible. If AI can reach a billion USD ARR, then our company’s cash flow could potentially break even, covering our R&D expenses and all costs.

So this is also possible. But we don’t prioritize it. We will do it, but I think it’s an important matter. It’s not our first priority, nor is it what we really care about today. The bigger opportunity should still lie ahead. The earlier opportunities include last year’s consumer side and this year’s B2B side. I think these need to be done well, but they are not our goal.

Or rather, most people in our company don’t think this is a very important matter, not as important as AGI. Let’s talk more about open source, because many questions have been about it. First, I think we will open source. Our strongest models might also be open-sourced. Because I don’t see any benefit to keeping them closed-source—no necessary benefit.

ByteDance’s model is closed-source. What benefit does it bring? I don’t see any benefit. Even if you open source and tell everyone everything, the barrier to entry is still very high. For others to use it, the barrier is high. It’s difficult for them to use it; and even if they do, making the cost low is also very, very difficult. It’s not easy.

Just because I open source doesn’t mean others can easily achieve the same deployment cost as me. There’s still a lot of work to be done. Although the working principles are understood, not every company is willing or has the capability to organize manpower to achieve this goal. I’m used to this. They may not be good at it because there is too much resistance. It’s hard for them to control costs. They have many management and physical constraints.

This is also the advantage of a startup. If a startup is too small, it doesn’t have the power to do this; if it’s a large company, it’s hard to organize.

Both have difficulties. So this is the sweet spot for companies of our scale. If we were larger, we might not have other problems; if smaller, our power would be insufficient. As for open source, I think we should set the price. Currently, we should recognize that we won’t coerce anyone.

Because for the pricing model, I won’t charge a very high fee. I might also charge based on a ten-month payback. With a ten-month payback, it will already make independent deployment by third parties unprofitable. Third parties cannot achieve this cost; they definitely cannot. So open source won’t affect my revenue. Of course, if I wanted to make a hundred times the profit, then open source would be…

Moderator

We heard you, but the video seems to have dropped, boss.

It’s probably the phone line.

Liang Wenfeng

So open source, I think, has no impact on our business model. The premise is that we only take a six-fold profit—ten-month payback corresponds to roughly six-fold profit. If we only take six-fold profit, open source won’t have any impact. But if you want to take a hundred-fold profit, then open source will indeed affect that, because third parties can deploy at, say, twenty-fold cost, which is lower than you.

Is this model sustainable long-term? I think it is. Under our vision, I think open source is sustainable long-term, or we plan to do it this way. You can think of it as having restraint, which also brings long-term benefits. This strategy gives us more opportunities in technology and increases the probability of us achieving AGI. We can be more relaxed.

Think about it: we don’t even need to work overtime, because it’s not that hard. But for other companies, it might be very hard because they have too many thoughts. Actually, it’s not hard. It’s not hard at all. It’s just beginning… Externally, it might look like we chose a difficult model—doing research, doing the hardest things, like a “hard” mode.

But in reality, we have given up a lot in other areas, which makes us very efficient and allows us to work very easily. So on the matter of open source, my judgment is that it’s sustainable. There is no conflict between open source and commercial payment, provided we are at a six-fold profit. Six-fold profit seems high, but it’s actually not that high. Given the high efficiency of AI today, a reasonable profit might be around that level.

In the future, it might drop to, say, four-fold or three-fold. I think that’s about as low as it can go, but there will still be significant profit. Even just selling API, though I don’t think selling API is that attractive. But this example shows there is no conflict.

And I don’t worry about others deploying our model and competing with us. Not at all. We hope they can deploy it. We try our best to help the open source community deploy our models. I’m not worried they will steal our business, because the market is big enough. I’m only worried they can’t deploy it well—some details might be wrong, making the performance worse or costs higher.

Yes, there is no conflict here. Last year, when we had more B2B inquiries, people asked: if we open source, will that conflict with our consumer side?

Because I don’t have the advantage of traffic. Tencent, for example, has a lot of traffic. If they deploy our open-source model, they might take all the consumer users away. But that doesn’t actually happen, for many reasons. And the question is: is the open-source model we provide the same as the one we deploy ourselves? Yes, it’s the same.

We don’t give a worse model as open source and keep a better one for ourselves. They are the same. This also shows there is no conflict. Last year, we basically open-sourced everything on the consumer side, and we didn’t see any conflict. Really, no conflict. So that’s the open source part.

Now, regarding the company’s long-term vision: I think our goal should be AGI. Everyone’s definition of AGI may differ, but that doesn’t stop us from taking AGI as our goal. From a technical roadmap perspective, the path to AGI is relatively clear.

With the current generation of AI technology, if you can describe a problem very clearly with complete context and instructions, it already surpasses humans. But there’s a premise: you give it complete context and complete instructions. This definition is hard to achieve.

For example, in our meeting today, we have a long and strong context—maybe decades of context between us. AI doesn’t have that. What AI can do now is, within a limited context, outperform humans. But it still can’t replace humans. One missing piece is continuous learning. Humans can continuously learn.

When you hire an employee, they might spend two months getting familiar with the company environment and their work. After two months, they can start working and do many things. They understand when you say, “Call Xiao Wang,” and they know who Xiao Wang is. But for AI, without those two months of learning context, if you say “Call Xiao Wang,” you’d have to tell AI who Xiao Wang is, their position, where they are, how to find them, and what to pay attention to. You’d have to give AI all the context. In that situation, AI could do it, but you can’t possibly provide all the context—it’s not realistic. So AI cannot replace your employee.

But if AI had continuous learning ability, like your employee, learning for two months at the company, then it could replace everyone. So the next step is solving continuous learning.

The development of AI can be understood as a staircase. Last year’s step was CoT (Chain of Thought). Because we found that through Chain of Thought, intelligence could reach a higher level. By thinking on its own, it raises the upper limit and enables AI to do more. We crossed that step.

This year’s step is Agent. Because we found that with Agent, even more tasks can be done. Its capability range expands, and its intelligence ceiling rises. Why is it a staircase? Because each step builds on the previous one. Agent uses CoT, and CoT uses the previous step—language models. No step is wasted.

So the development of AI intelligence has a clear trajectory. This year’s step is Agent. But the Agent step will also be completed. After solving all problems within its reach, it still cannot replace your employee, but it will have reached its ceiling.

Just like CoT: after reaching its ceiling, it already surpasses the best humans at solving Olympiad math problems and writing code. But it still stops there. That technology does not yet reach AGI. So you see, the trajectory of AI intelligence is traceable.

After Agent, we think the next problem to solve is continuous learning—how to enable models to learn continuously, rather than requiring a strong one-time training. They should be able to learn over a long period like humans. This problem is related to task completion and others; they address the same issue.

We are now at the Agent stage and can see the next bottleneck: continuous learning. The next problem to solve is how to achieve continuous learning. This is visible and relatively clear. It’s an obstacle ahead that you must cross, and there must be a way to cross it, but it takes time. After continuous learning, we might arrive at a singularity.

That singularity is: when the model can learn continuously, it can do everything humans can. It can develop its own version, do research, and develop its next version, creating even more advanced AI models. So it will reach a singularity of self-iteration. But this singularity is not a singularity; it’s also a gradual process.

This process might be a long gradual change, not a sudden shift. But conventionally, people think it’s a singularity because early prophets predicted a singularity there. In reality, it’s not a singularity; it’s a continuous process. After this step, I think comes embodied intelligence.

This is our speculation. We think the timeline should be: first solve continuous learning, then the singularity of self-iteration, then embodied intelligence. After embodied intelligence, it enters the physical world, can do housework, take care of the elderly. We think this is an ideal roadmap. But everyone’s view differs—no right or wrong. We just think this roadmap is the easiest.

Because each step requires very little new effort. With this roadmap, we don’t need to work overtime. But if the roadmap were reversed—say, achieving embodied intelligence first—that would be very laborious, a tough job. We don’t want that. We want to do it easily.

If we solve continuous learning first, then the self-iteration singularity, then embodied intelligence, the path is easy. Because later on, you can use earlier technologies to help develop later ones. After the singularity, doing embodied intelligence wouldn’t require human effort—the model would do it itself. So that answers what our long-term goal is.

Let me tell you: this is our long-term goal, which we call AGI. Let’s return to reality. Last year’s biggest reality was that everyone wanted to do chatbots, compete for consumer traffic. This year’s reality is that everyone wants to grab B2B revenue and participate. If you don’t participate, you’re not at the table, right?

But we don’t think it’s an important matter. Inside the company, what we truly care about is the AGI roadmap I just described, and how to achieve the next technological breakthrough. But strangely, the thing you want most is the hardest to get. Things you don’t care about that much come relatively easily.

There is a strategic advantage here: our minds are on AGI, and we are doing AGI. When we then do applications, consumer or B2B, we don’t need to devote much effort. Very little energy is required. I think it’s like you’re at a high technical level, and doing something at a lower level is a kind of dimensionality reduction.

At least last year on the consumer side, we saw that. We didn’t put much effort into the consumer side. We even considered not maintaining those users. But users wouldn’t leave. They really wouldn’t leave. So they all stayed. And now this year’s B2B revenue looks quite optimistic. I think compared to peers, this number should be relatively good, I estimate.

But we haven’t put much effort into it. We haven’t

done anything extra. It’s just a byproduct. When we do the step towards AGI—which is a necessary step—we serve this technology through API. We haven’t done anything extra. We are still doing AI. This is a byproduct.

We just need a few people to maintain the API. We don’t even have customer service, no sales, nothing. Users come by themselves.

Or consider consumer and B2B users: they are all byproducts on our path to AGI. They are intermediate outputs and don’t conflict with our AGI work. We are not doing it for the consumer or B2B market. We are doing AGI for AGI’s sake, and we happen to generate these outputs, which we then commercialize. This is different from other companies.

Other companies make models to serve consumer or B2B users. For us, the initial intention is different. We are still pursuing AGI. I think this is a kind of dimensionality reduction. AGI is a bigger vision that can attract more talented people and has stronger cohesion.

So I have an organizational advantage, and I use that advantage… it’s a dimensionality reduction. If you are a commercial company with a vision to serve consumer users, that’s a different story. You have other advantages—product, user service, traffic—but not in technology. Currently, the favorable situation is that model technology is the most important.

You need to make the model good. Everything else… yes, this can explain our development trajectory. We really chose AGI, and we never thought about having many users. When we suddenly became popular last Spring Festival, it wasn’t in our script. We never thought about it. We just wanted to improve the technology.

But then I found that our organization, compared to fully commercialized product-oriented organizations, had extra advantages in talent and organization. That’s amazing. At that time, everyone was fighting fiercely for consumer users, but in the end, it was someone who wasn’t fighting who took it away. This indeed confirms the logic I mentioned earlier: talent and organizational advantages.

The talent advantage is not that our people are smarter than others, but how I organize these talents, how to motivate them, and how they collaborate. That has an advantage. Because bringing smart people together doesn’t automatically mean they will cooperate and be passionate about a goal. You need a vision. Our past experience taught me that the AGI vision is very powerful.

Okay, that’s the question. Next question: what is the importance of core interests?

I said earlier that we need to be very restrained in many aspects. But what is our core interest? Actually, our core interest is only one: maintaining team stability. This is our biggest core interest, arguably our only core interest.

As long as I can maintain team stability, we will definitely succeed, we will definitely achieve AGI. It’s that simple. As long as everyone stays and we can continue, we will… basically no risk. It’s just a matter of sooner or later. We may encounter setbacks, but if everyone stays, we can continue. Money is certainly not a problem. Resources are not a problem. Other elements are easy to obtain.

For us, there is only one core interest that cannot be compromised: we must maintain team stability. This is also a very big challenge, or I think the biggest risk. Of course, this risk has been significantly alleviated by our recent fundraising. Because everyone received relatively many options, the amounts are quite large.

From a team stability perspective, as long as the most important and oldest employees are stable, others are unlikely to leave. Even if others have fewer options or lower income, they won’t leave, because they are not only after money. Everyone hopes to work in an environment where AGI can be achieved. So it’s still attractive to talent.

Historically, our talent turnover has been relatively low compared to peers. It has always been lower. But this is still our biggest challenge, the only challenge, you could say. Everything else is just a matter of time. At worst, we might delay half a year or a year, but we won’t fail. We definitely have money, definitely have resources—we lack nothing.

So many of the things we do now are to maintain team stability. Apart from this, I think we can give up everything else. We can be restrained. We have always been very restrained and unwilling to become an opponent of any major or minor internet company. I hope we can empower them, or help them do this, and assist everyone.

This is also part of our commercial significance. The premise is that everyone doesn’t…

Under this premise, we are very willing to assist and help anyone, even our competitors, including Alibaba, Zhipu, and Moonshot AI, to do better. Because we don’t lose anything. We are already open source. Open source is also about making the boundaries as clear as possible. We hope you can replicate it. If you can’t, ask, and I’ll tell you how. That’s part of open source. It doesn’t matter if you are a competitor.

Of course, if you’re a partner, I’ll do more. But on large interests, there is no conflict.

When dealing with the outside world, our attitude is: we only do the main line of AGI. This is what I just said—GPT, CoT, Agent, etc. We only do the main line. The AI field is broad. There are many things we feel are not on the main line, such as 3D and video generation. I think these are not very relevant to the main line of intelligence. We won’t do them.

There are also things like world models. I think they are not very important for the ceiling of intelligence at this stage. So we won’t do them either. But if others do them, we are happy to help. Whether we have time is one thing, but there is no conflict of interest. We also hope these AI technologies can be used in various production environments to improve social productivity and help various industries.

We are very motivated to do this. Whether we have time, manpower, or our colleagues are interested is another matter. But there is no conflict of interest. We hope to achieve this goal, and we believe it has no commercial conflict. We still get the profits we deserve.

I think our previous attitude—we haven’t lost anything because of it. We haven’t gotten less of anything because of open source, our goodwill, or our help to others. For example, last year’s consumer users: we still have many, and they are stable. This year’s B2B side, I think it’s optimistic.

We haven’t affected our commercial interests because of our goodwill. No effect at all. It might even add points. This seems counterintuitive, but it’s true. Or think about it the other way: if we went against it, could we get more? No. One question is: how do you understand that world models are not related to raising the ceiling of AI intelligence? I say that at this stage, that’s our judgment.

We have seen that from our own AI roadmap, it’s not the only one. From our understanding and judgment, the most important thing right now is to do AI training well. Doing AI training well doesn’t require world models, or even multimodality. If you narrow the scope of AI training, without multimodality, you just can’t do some tasks, but that doesn’t affect the algorithm’s viability. Multimodality will eventually be needed.

What’s important now is training. Next is continuous learning. Then self-questioning. But this roadmap doesn’t include world models or video generation. When video generation first appeared, it was very hot, as if it were necessary—if you don’t do it, you’re not an AI company.

I found that strange. If you think carefully, it has nothing to do with the intelligence roadmap.

In fact, after Sora came out, everyone did it—big and small companies. But small companies later cut it. It’s unrelated to the ceiling of intelligence. Commercially, it’s a good business, but it’s not related to intelligence. We won’t do it just because it’s a good business. We only do things that are on the intelligence roadmap.

Video generation is relatively clear, so I use it as an example. World models have a less clear meaning, because many things could be called world models. In our judgment, world models and intelligence are not the most important at this stage. The most important are AI training and solving continuous learning after training.

This is our company’s judgment. Of course, each company has different judgments. I just said that for our company, the most important issue is personnel stability. From another dimension, what are we lacking? What is the gap between us and the US? Actually, the gap is only one thing: resources. We don’t have as many GPUs. Our number of GPUs is relatively small.

We currently have about 20,000 H-equivalent effective compute. Most of this arrived recently—in the last month or two. Many machines may not have arrived yet. Our total compute last year was relatively low. This year we are aggressively expanding compute. We now have about 20,000 H-equivalent. In the next few months, we will buy large quantities of machines, mostly NVIDIA.

How many GPUs do we need? Definitely more is better. Within our affordability, the more GPUs, the better. That’s without doubt. So our strategy now is: buy as many GPUs as we can at reasonable prices. After this financing is used up, I’ll buy as many as I can. The speed of spending is not planned; as long as the price is reasonable, I’ll buy as many as I can.

If I spend all the money in half a year, I think that’s a good thing. If I spend it all in half a year, that would be too ideal. Actually, spending so much money in half a year is very difficult. It’s hard to buy that many GPUs. Prices are high, and we can’t pay extremely high prices; we need to ensure the price is reasonable.

If I spend all the money in half a year, that might be most ideal. Because turning money into NVIDIA GPUs is definitely better than leaving it in the bank. In the bank, I might get two percent interest. But buying NVIDIA GPUs has a ten-month social cost.

So definitely, buy as many as you can. If we buy GPUs first, later we can generate cash flow through services or something. With cash flow, we can survive. We don’t need to keep a lot of cash on hand. So we only worry about not being able to buy enough GPUs.

If we could turn all money into GPUs, we would unreservedly do so, and we are willing to pay a certain premium. We are willing to pay a premium to turn money into GPUs, because it’s so cost-effective. Even after paying the premium, it’s hard to achieve this goal. Objectively, if we can spend 20 billion RMB in one year, that would mean our procurement department performed extremely well.

The gap between us and the US is mainly in resources. The gap in people is not large. There is almost no gap in people, because they are the same group—Chinese people. When Chinese people go abroad, some stay abroad, some stay home. It’s not that smart people all go abroad. It’s relatively random.

The smartest people—maybe not more than half go abroad, but less than half stay in China. China doesn’t lack talent, and we have a large base, with many new people each year. Talent is not a bottleneck. Resources are the biggest bottleneck. Resources first affect talent cultivation. Because with less compute, we have fewer experimental opportunities, so our talent overall lags behind the US. The talent gap is essentially due to the compute gap.

On the largest models today, we actually can’t afford to train them. Even if we spent all 50 billion RMB, we couldn’t afford to train them. Even if we could stack them, we couldn’t use them. The largest model currently has about 800B activation parameters. Domestically, we are still at tens of billions of activation parameters. That’s an order of magnitude difference.

If I wanted to train a model as large as AI’s, I’d need about 50,000 GB300 GPUs, or 200,000 Huawei 950 GPUs. That’s just for training, not including research. So the biggest gap between us and the US is resources. Our current resources, including what we’ll have in the next few months, including the big resources coming soon, are only enough to do more experiments at the scale of tens of billions of activation parameters.

Because at the tens of billions activation scale, we still have many experiments to do and many things to figure out. We are still far