Organized by | Gu Lingyu, Tencent Technology
DeepSeek recently completed its first external funding round since its founding. This round raised a total of over RMB 50 billion (approximately USD 7.4 billion), with a pre-money valuation of about RMB 367.5 billion (approximately USD 54.3 billion). Among the investors, DeepSeek founder Liang Wenfeng personally contributed RMB 20 billion, Tencent invested RMB 10 billion, CATL invested RMB 5 billion, NetEase, JD.com, andIDG Capital each invested RMB 3 billion, and the National Artificial Intelligence Industry Investment Fund invested RMB 1 billion.
Prior to this, Liang Wenfeng had proposed the principle of "no financing, no IPO, no commercialization." This large-scale fundraising marks DeepSeek's official entry into the capital market and has sparked widespread industry attention regarding its commercialization path and technological vision.
At a recent investor exchange meeting, Liang Wenfeng elaborated in detail on DeepSeek's organizational culture, open-source logic, technology roadmap, and views on the competitive landscape.
The following is a compiled transcript of Liang Wenfeng's remarks from the nearly 4-hour exchange meeting, obtained by Tencent Technology, categorized by topic, totaling 118 items. The text preserves the original meaning as much as possible, with only minor editing.
01 Vision and Restraint
1. When we first started this company, our initial thought wasn't about how much money we would ultimately make, going to the capital market, going public, or anything like that. The first few dozen people never thought that way. If they did, they wouldn't have come.
2. We embarked on this with a great goodwill towards the world. We believe this is useful for humanity, it's something beyond money. The original intention we started with, our vision, and the vision we maintain to this day, are not pursued in a way that maximizes commercial interests.
3.Managing a large company depends not on your rules and regulations, but on vision. Vision is not slogans on the wall; vision is about how you act, not what you say. It's about how you actually operate.
4. We have no organization; we are vision-driven, organized by a vision. We don't operate with a "I need to achieve this KPI, no assessments" approach. Only vision.
5. This vision isn't even documented; it's not written down anywhere. This vision lies in our methods of doing things, our attitude towards the world.
6. We don't have many other advantages. We have no special skills. We are not wealthier than others, nor is our staff better than other companies'. Actually, no. When we founded this company two years ago, we didn't have much money, many GPUs, any fame, or any appeal. We were just a group of very ordinary people.
7. The more restrained you are, the easier it might be to succeed, or at least so far, it has been proven, so far it can be explained. Otherwise, there's no way to explain why we could succeed: we had no special weapons, a very low starting point, very few resources, and our people were just a random group of ordinary individuals.
8. AI is too big, the interests are too vast. We are very restrained. As long as we succeed, the final benefits will be enormous. If you just take a small portion, the benefits are huge, so there's no need to consider now which part of these benefits to take or how to take them. I think there's no need to consider this at all, because the benefits are sufficiently large.
9. Last Chinese New Year, we suddenly had many users, but we didn't pursue retaining these users, monetizing them, or cashing in on the commercial benefits from these users. We didn't fight for users, didn't try to make money, but we worked hard to serve the users well.
10. We don't have thoughts like, I want to build the next super app, then compete with someone, become the next ByteDance, become the next Tencent. No such thoughts at all. I think the futureAGIopportunities should be very large; the future AGI opportunities will always be very large.
11.Restraint is a strategy. It lies in sometimes being able to give up some things to gain more other things.Not open-sourcing is actually the same; it can be considered our pressure, or considered our concession.
12. This restraint, I understand, can increase our probability of achieving AGI in the long run. When considering something, I have no doubt that AGI will have enormous commercial value. On this basis, my priority is not how to get a larger share, how to take more; my priority is how to increase the probability that I can succeed.
13. We have always been very restrained, unwilling to become an opponent of any internet giant or small company. I hope I can empower them, or hope I can assist everyone in doing this, hope to help everyone do this.
14. I think by holding this attitude before, we didn't lose out on anything because of it. We didn't lose anything because I open-sourced, because of our goodwill or because I provided help to others. Instead, it might have added points. This seems counterintuitive, but it is indeed like this.
15. We aim for AGI, but we have always been doing commercialization, so we have C-end users and B-end revenue. Based on historical experience, this strategy has been successful.
02 AGI Roadmap
16. If you can describe a problem clearly, give it complete context and instructions, it has already surpassed humans. But there's a definition, a premise here: you give it complete context, complete instructions.
17. AI cannot replace your employees. But if AI has the ability for continuous learning, it can learn at a company for two months like your employees, then it can replace anyone in the world. So we are still one step away from that: continuous learning.
18. The development of AI can be understood as a staircase. The step taken last year was Chain-of-Thought. Because we found that through Chain-of-Thought, intelligence can reach a higher level.
19. This year's step is Agent, because we found that using Agent, even more things can be done, its capability scope will be larger, its intelligence ceiling will be higher. Agent needs to useCoT, and CoT also needs the previous step, the previous step being the language model. So no step was wasted.
20. After Agent, we think the problem to solve should be continuous learning, which is how to make the model learn continuously, rather than giving it a strong training. It should be able to learn continuously over a relatively long period, like a human.
21. After continuous learning, we might reach asingularity. This singularity is that when the model can learn continuously, it can already do everything humans can do. It can develop its own versions, conduct further research, and then develop its next version, develop more advanced AI models.
22. This singularity is not actually a singularity; it's also a gradual process. This process might also be a relatively long evolution, not a sudden mutation. But habitually, we all think it might be a singularity.
23. This is our speculation. We think this timeline should be: first solve learning to learn, then reach that intelligence singularity of self-iteration, thenembodied intelligence. After embodied intelligence, it enters the real world, can do housework for you, can provide elderly care.
24. If we solve continuous learning first, then solve the self-iteration singularity, then solve embodied intelligence, the path becomes very smooth. Because after that, you can use earlier technologies to help develop later ones.
25. We only focus on the main line of AGI. The AI field is broad; many things we think are not on this main line, like 3D, video generation. I think they may not have much to do with the intelligence main line; we won't do them.
26.Video generation became very popular when it first emerged, as if it's something you must do; if you don't, you're not an AI company. So I find it strange. Actually, if you think about it carefully, it has little to do with the intelligence roadmap.
27. Commercially, it's a good business, commercially it's a good business. But it has nothing to do with intelligence. We won't do it just because it's a good commercial opportunity; we'll only do it if it's on the intelligence roadmap.
28. According to our judgment, world models and intelligence are not the most important things at this stage. The most important are AI training and how to solve continuous learning after AI training. This is our company's judgment, of course each company's judgment is different.
29. We now relatively believe in a narrative that AI can accelerate AI research. That is, it's not linear because you can use AI to accelerate your own research, so it might be non-linear later.
30. I think embodiment must eventually be entered, ultimately embodiment. Because for a normal person, their needs aren't a computer, right? Normal people need food, drink, entertainment, clothing, housing, transportation; they don't need a computer. What they need, so they still need embodied intelligence to solve specific human needs.
31. What do we hope AGI can do? It can help me iterate the next version of the model, can help me iterate the next version of the model. If we have embodiment, what we hope it does is also to let it iterate the next version of embodiment, let it make the next version of robots.
32. The core capability of the next-generation model must have continuous learning ability; only then can it be called the next-generation model. Before that, what we can do is cost, then make the effect better, speed faster. But for a major breakthrough, it should possess continuous learning.
33. The current limitation of Agent's ability is because it cannot learn continuously; it cannot effectively learn continuously. If continuous learning can be solved first, then AI's ability is very strong; it can greatly enhance our own research efficiency.
34. If continuous learning is achieved first, general intelligence might become easy; using it to do it becomes easy. So I say this is a result we'd like to see; we save effort, we become relaxed. Otherwise, manually doing general intelligence now is relatively tiring, labor-intensive, a data-intensive, labor-intensive task with low cost-effectiveness.
03 Team and Talent
35. Our previous experience taught me that the AGI vision is very powerful. This talent advantage is not that my people are smarter than his, but how I organize these talents, how I motivate them, and how they collaborate.
36. Gathering smart people together doesn't mean they naturally can cooperate, naturally can passionately pursue and achieve a goal. So you need a vision.
37. Our biggest core interest is to maintain team stability. This is our biggest core interest, perhaps even the only core interest. As long as I can maintain team stability, I will definitely succeed, definitely achieve AGI. It's that simple.
38.Money is definitely not a problem, resources are not a problem, other factors are easily obtainable. For us, there's only one core interest, only one thing we cannot compromise on: we must maintain team stability.
39. This is also a very big challenge we face, or I think it's the biggest risk. Of course, this risk has been largely mitigated by our recent fundraising round. Because everyone received quite a lot of options, the amounts are still relatively large.
40. From a team stability perspective, as long as the most important employees, the oldest employees, can stay stable, others are less likely to leave. Even if others have fewer options, lower income, they won't leave. Because they didn't come solely for money; everyone hopes to work in an environment where AGI can be achieved.
41. Everything else is a matter of time; at most, other things might cause us to be half a year late, a year late, but not that we can't achieve it. Definitely not short of money, definitely not short of resources. Actually, we lack none of these.
42. Our gap with the US is mainly in resources; the gap in people is not huge. Almost no gap in people, because they are the same group of people, likely Chinese. When Chinese people go abroad, some stay domestically, some stay abroad, some go abroad. It's not that the smart ones go abroad; no.
43.Talent is not the bottleneck; resources are the biggest bottleneck. Resources first affect talent cultivation, because with less computing power, we have fewer experimental opportunities, so our talent overall lags behind the US. The talent gap is essentially due to the computing power gap.
44. The shortage of AI talent is also temporary, and we've already seen it greatly alleviated. Because AI talent really isn't lacking; each company quickly cultivates people; cultivating people is fast.
45. There are still too many companies doing models domestically, still too many. In the US, maybe three companies; China has too many doing foundational models. In the end, certainly not that many people need to do foundational models; they will converge.
46. Our company's management actually has two lines: one is top-down, one is bottom-up. Bottom-up means each person thinks about what they want to do, does it themselves, no one manages them, no KPI.
47. Generally, we hope employees still have half their time unassigned, doing whatever they want. This is a research scope, allowing them to explore on their own, based on what they think is important, without predefined requirements.
48. We generally don't work overtime much either. Overtime has two reasons. First, research requires a relatively relaxed environment. If you push too hard, you can't do research. Since it requires your own interest, you need to think about these problems in your spare time, so it must be in a relatively relaxed environment to explore.
49. Second, we are very focused. Being very focused means we have very few things to do. Then I don't have that many things to do, so I don't need overtime. This is consistent with the earlier restraint.
50. Our company overall is built on consensus; I don't decide everything alone, but I seek consensus. My authority and influence within the company are built on consensus.
51. This decision-making mechanism is actually a consensus-seeking mechanism; it's not that I can push something; it must be consensus, then I can push it, then I will push it.
52. As personnel increases, we will make adjustments. Actually, we need to make this adjustment soon because I'm already doing this adjustment. If we don't make this adjustment, many things cannot proceed. Indeed, many departments should have organizational structures.
04 Computing Power and Resources
53. How many GPUs do we need? Now definitely the more the better. Within what we can afford, definitely the more GPUs the better, no doubt. So our current strategy is, at a reasonable price, buy as many GPUs as we can.
54. Actually, spending this much money is very difficult; you can't buy that many GPUs, hard to buy, and prices are high too. Can't buy at extremely high prices; must ensure the price is reasonable. If we can spend two hundred billion this year, then our procurement department's performance would be super good.
55. Our biggest gap with the US is in resources. Computing power resources: on one hand, domestically we simply can't buy GPUs; on the other hand, our capital investment is less than the US. Our capital investment level is much lower. Salaries based on talent account for a very low proportion. Look at their salaries like a hundred million USD, but calculated, talent salaries still account for a small portion; the bulk is computing power.
56. All the differences we see, including talent differences, model capability differences, application differences, can be considered due to differences in computing power resources.
57. Our gap with the US might be lagging behind by 12 months, lagging behind maybe 12 to 18 months, or say 6 to 12 months. Simply put, lagging behind the US by two years, but achieving this with one-twentieth of the US's computing power.
58. This narrative is lagging one to two years, but using only one-twentieth of their computing power. In the future, we want to rewrite this narrative: using a fraction of their computing power, but shortening the time further to 6 months, 3 months. I think this is a goal.
59. Scaling, we believe in Scaling; definitely the larger the scale, the better the effect, unlocking more functions. What prevents us from Scaling is actually computing power, not that we don't want to Scale, but we don't have that much computing power to do this Scaling.
60. We train such large models not because I think models this large are enough, but because I happen to have this many resources. I calculated based on my resources, what size model I can accept and train; that's how it's calculated, not that this model is enough.
61. When Silicon Valley says Scaling is reaching its end, that's for Silicon Valley; for Chinese people, we are far from that; we haven't even Scaled to that extent. This Scaling includes data Scaling, model size Scaling, and training cost.
05 Domestic Chips and Ecosystem
62.NVIDIA's CUDA moat is rapidly eroding. On one hand, now with AI, and with AI, establishing this ecosystem is much easier than before because AI can write code.
63. Now the computing card market is already larger than the gaming card market, so there's no reason these two still need to be coupled. The trend now is, in the future they will no longer be coupled. Then specialized chips, whether Huawei or NVIDIA itself, in the future will be specialized chips, not these previous things.
64. Domestic AI chip replacement now has a historic opportunity. We believe within a year, we will see something proven: the domestic chip ecosystem is completely fine. Previously it was thought problematic, thought unusable, not easy to use, but in the future I think within a year, we can reverse this perception, or use facts to reverse this.
65. Domestic AI chip hardware and ecosystem are not problems; the only problem is insufficient production capacity. Domestic card adaptation, no obstacle here, NVIDIA cannot block it. In a normal business environment, if I could buy NVIDIA cards, then domestic replacement is relatively difficult; but when NVIDIA cards are unavailable, everyone is forced to work on domestic chips.
66. When training V3, it still used NVIDIA cards, but didn't use NVIDIA's ecosystem. V3 used NVIDIA cards but didn't use NVIDIA's ecosystem. Instead, we first wrote a high-level compiler called TileLang, then completed everything else based on the TileLang ecosystem, already almost independent of NVIDIA's ecosystem.
67. I'm relatively optimistic about domestic computing power. I think on this point, NVIDIA is digging its own grave. Huawei's super node, Huawei's 950 super node, in performance and price can completely substitute NVIDIA's GB200, GB300.
68. Four Huawei cards equal one NVIDIA card.
69. Our gap with the US in chips, I think in ecosystem there will be no gap in the future, but in chips it's four times plus two years.
70.We mainly cooperate with Huawei now. Huawei adapts themselves, but we will participate in this ecosystem, deeply involved in Huawei's. Huawei's problem is still insufficient production capacity.
71. I don't quite believe that five years later, we will still be stuck on production capacity issues. Now definitely stuck on production capacity issues, this year, next year, the year after, I think still stuck on production capacity issues. But five years later, I think maybe not necessarily; I'm still relatively optimistic.
06 Competitive Landscape and Industry Judgment
72. The gap in final model performance between various companies should be comprehensive. Comparing model effectiveness must be at the same cost; that's meaningful. Because comparing two cars, you compare cars of the same price range.
73. Anthropic surpassing OpenAI now, is this long-term? I think it's not long-term; it's definitely temporary. OpenAI and Google will likely continue to alternate rising in the future.
74. In the global AI division of labor, Chinese companies are very likely to play a role of having the largest production volume. Logically, our production capacity is largest, including chips; chips maybe our production capacity is largest, our electricity is most abundant.
75. Chinese people will make the product cheapest, then in terms of effect, after all, foreign products, many products now have little difference between Chinese-made and US-made. In the future, AI might be like this too, but Chinese-made AI might be cheaper. This cheapness might be systematically lower, just like other industries where Chinese-provided services are cheaper.
76. The final gap should be three aspects: one is cost, one is time, one is user experience. Beyond that, probably no gap.
77. Cost is definitely a difference; I think cost might be the primary difference. Then second is time, when you can achieve it. Being a few months earlier or later makes a difference.
78. OpenAI initially thought they could truly monopolize the world, but actually they will encounter many, many challengers. They will face challenges, then they won't be so relaxed. The US will face challenges, then in the future they might also face challenges from China, because Chinese people are willing to provide this service for less.
79. Those who take more will be defeated by those who take less. You don't even need to actually take more; if the vision is to take more, you lose first; you will face greater difficulties.
80. For us, we don't aim to take the most money in profit, or price for profit maximization; we only earn a reasonable return. This is one explanation. I believe this; I'm not finding reasons for this, because no need to find reasons.
81. I think in many user experience aspects, we might be able to do better than the US. In product aspects, product capability might not necessarily be worse than the US. Cost should also be lower than the US, so China will still have competitiveness.
82. Cost is easy to understand because they don't need to do it, so they don't develop this capability. They definitely don't value this as much as we do. We can treat it as very important, but for them, this is unimportant.
83. For large models, maybe not two big companies, two small companies; maybe that's enough. The gap only has two things: one is time, one is cost. So it won't lead to huge profits for any one; I don't think huge profits will happen. Those with good cost control earn a bit more; those with poor cost control earn a bit less, just that.
07 Model R&D and Technology
84. In our company, maybe half the people usually think OpenAI is better. Actually, Anthropic has first-mover advantage, but this first-mover advantage should disappear soon; it's not a long-term sustainable advantage. All three of these are very strong; among these three, efficiency is highest; the cost it spends, the money it burns should be the least.
85. Multimodal layout, we have been doing it. For products, it's important; for C-end user products, it's important. But for the upper limit of intelligence, it's a component, not the main line itself.
86. We should launch related models; our V4, subsequent versions of V4 will support native multimodal. But for us, multimodal, for intelligence, it's a component; we don't treat it as intelligence itself.
87. I can only say language model Scaling, I haven't seen an upper limit yet. Our current intelligence level, or the US's achieved intelligence level, haven't seen an upper limit.
88. Many people internally think like this: first, it must be useful to ourselves, first used by ourselves. Then this is the fastest way to achieve AGI. When it's good for us, that means it might be good for others too, but first must ensure it's good for us.
89. The first goal of the models we make is not that others find them easy to use, but that we find them easy to use. First, useful to ourselves. After useful to ourselves, when I develop the next version of the model, it will be faster.
90. We call this "scratch-off lottery." The threshold is low; anyone can scratch, but who can scratch out what, maybe I don't know if it's based on talent or something. So here we don't need to allocate resources. It's just that, the difference between us and other companies is we spend time discussing this problem, think about this problem, then treat it as an important matter.
08 Commercialization and Pricing
91. Our API pricing is a reasonable profit, roughly we buy a batch of equipment from the market, recover costs in ten months. I think this is a reasonable profit.
92. If maximizing profit, should set price higher. Because in this price range, user demand is inelastic; that is, if I double the price, or raise price by another half, token consumption difference is not significant.
93. For one of our models, initially we worried demand too high, so initially set price relatively high; the team wasn't very happy. Later I lowered the price to one-fourth; everyone was happy.
94. The ceiling for To B business should still be demand; under the current generation AGI, AI technology background, To B demand should be limited. It will grow rapidly, but not an infinitely large matter; ultimately constrained by demand, not computing power.
95. Now I think should be achievable, to have both. Suppose this year we can have several hundred million USD in B-end revenue, plus we have C-end users, then this itself already has a certain commercial foundation. Next year we have B-end revenue; if this demand can increase further, the company is not far from net profit; might already be net profit.
96. Worst case selling API might support a listed company. If technology later has no new progress, our technology freezes here, then we go all out selling API, do these services well, I think also enough.
97. Looking at our current situation, I think the most reasonable approach should be fully focusing on general Agent; other Agents should have lower priority, including finance, doctor Agents. Do Coding first, because Coding Agent can do many things, also many vertical Agents. At this stage, we think most important should still be Coding Agent.
98. I think low cost is first a result. Our model indeed has been moving towards a lower-cost direction in model architecture, related to our vision. We also have many algorithmic methods; cost can go further down.
99. Another reason cost goes down is, the lower the cost, the more I can train larger models, the more I can afford larger models. With limited computing power, if my computational efficiency is higher, I can afford larger models.
09 Open Source Strategy
100. I think we will open source, and our strongest model will likely also open source. Because I see no benefits in closed source; see no necessary benefits. ByteDance's model is closed source, what benefits does it have? I see no benefits.
101. Even if the model is open source, you tell everyone everything, the barrier is still very high. For others to use it, the barrier is also very high. For them to use it is difficult; secondly, to use it, also keep costs low, also very difficult, not that easy.
102. Open source will not affect revenue. Open source, I think has no impact on our business model.
103. I'm not worried about others deploying our model to compete with us, not worried at all. We hope they can deploy it. We try to provide help to the open-source community, assist everyone to deploy our model.
104. When dealing externally, our attitude is: we only focus on the AGI main line. When dealing externally, we are very willing to assist, help anyone, even our competitors, including Alibaba, Zhipu, MoonShot AI, to do better. Because we don't lose anything; we are open source anyway.
105. The open-source model we provide, is it the same as the model we deploy ourselves? Yes, same. We won't open source a worse model, then deploy a better model ourselves; no, same.
10 Data and Post-Training
106. Data should almost equal half the model. There's also the issue of labeling data before that. In data labeling, this relates to our capital investment structure. With our capital investment structure, we can't support that high cost of high-quality data labeling, because cost is high.
107. US data labeling cost and Chinese data labeling cost have little difference. China labeling data has no cost advantage, especially labeling high-end data, no cost advantage, making it difficult for us to invest like the US in labeling data. This path is difficult in China because labeling data is too expensive, whether outsourcing or labeling ourselves, both uncomfortable.
108. Now basically two legs walking. Not that we completely can't label, but because labeling data, some costs low, some high. We label low-cost first.
109. You could also say, now half the people in our company label data. Half the core researchers, the most important people, half label data. We concentrate on labeling data. Solving the AI problem, at this stage, relies on labeling data.
110. The bottleneck for high-quality data labeling, I think is time; needs time. Because for OpenAI, for abroad, for Anthropic, they started earlier, have more capital, more GPUs.
111. The hallucination problem in large models relatively affects user experience. The hallucination problem also has a method to solve, but this is a long-term proposition. The hallucination problem can be considered solvable through better Post-training, is a problem that can be solved, improved.
11 Organization and Company Positioning
112. First, we have no model to imitate. Every step is from actual situation, seeking truth from facts, making decisions based on actual situation, finding how we should do. So it's a product of the times, or a reflection of reality; it's not an imitation result.
113. We clearly need commercialization. We ultimately need to survive; after all, we are a company; the government won't give me a cent.
114. We are essentially still a company, just when considering which money to earn, when to earn, how much to earn, what to earn from, we make choices. Many companies become great because they have a pursuit beyond profit. That pursuit not only doesn't affect its commercialization but allows it to commercialize better.
115. Regarding partners, actually this fundraising was carefully selected. First, I think interests are relatively aligned, those most aligned with our interests, least hostile to us, or most hoping we succeed. Not everyone hopes we succeed, because we still harm many others' interests.
116. AI now lacks not taste and intuition; it lacks continuous learning ability. AI's taste and intuition are fine. Let it write an article, its taste and intuition, I think fine.
117. We hope to only do one part. I think AI is big, doesn't need me... I only do one part. If focused, and I think the business interest here is already large enough, if the AI era will produce many trillion-dollar companies, I think we are one of them.
118. We hope to support more people, but we don't have that much energy. We have this willingness, and no conflict of interest, but whether we do is another matter. But at least here no conflict of interest; we hope for win-win cooperation.
