This document is based on approximately 3 hours and 44 minutes of meeting minutes containing conversations between Deepseek founder and CEO Liang Wenpeng and investors, and summarizes Deepseek's vision, open source philosophy, AGI roadmap, organizational culture, gap between China and the United States, and hardware strategy. Wenfeng Liang explained that Deepseek is a research-driven company that prioritizes achieving AGI over profit maximization or market monopoly.

The key point he repeatedly emphasized was "moderation". The argument is that rather than trying to occupy all markets and profits, only reasonable profits should be taken and the remaining opportunities and ecosystem should be shared with other companies and society to increase the possibility of reaching AGI in the long term. However, since these meeting minutes are data that has gone through AI voice recognition and translation, there may be errors in some proper nouns and numbers.

1. A company that started with a vision for humanity rather than money

Wenfeng Liang said that Deepseek was not a company created from the beginning with the goal of listing, entering the capital market, and maximizing corporate value. The explanation is that the dozens of members in the early days of the business also did not come together to make a lot of money, but joined with the belief that AI could be something useful for humanity.

Of course, as the company grows and the scale of the AI ​​industry grows, commercial temptations and new possibilities arise, but that is a different problem from the starting point. DeepSeek's fundamental goal is still to create technologies that are useful to humanity, not to maximize commercial profits.

"When we first started this company, our starting point wasn't how much money we were going to make, whether we were going to enter the capital market, or whether we were going to go public. We believed that this was something that would be useful to humanity, and that it was something that went beyond money."

Wenfeng Liang selected vision and mission as the most important assets in corporate management. He evaluated that although many of the management theories of Jack Welch, former CEO of GE, may have become less effective over time, the idea that "a company's most important asset is its vision" is still correct.

The vision he speaks of is not a slogan written on a company wall or an official document. In reality, the vision is revealed in the decisions made, the attitude toward the world, and how members act.

"Vision is not a slogan on a wall. Vision is what you actually do and how you operate."

2. An organization driven by vision, not KPIs and organizational charts

Deepsee explained that it does not have an organizational structure in the traditional sense. Rather than managing members based on a clear departmental system, KPIs, and evaluation tables, a shared vision is more of a way to bring people together.

Wenpeng Liang said that Deepseek's vision is not written down and may be understood slightly differently by each member. However, he explained that the members are generally in agreement on the larger direction of contributing to humanity and the development of AI.

"We have no organizational structure. We are driven by vision and organized around vision."

This method has both advantages and disadvantages. A weak formal management system can lead to inefficiency or confusion, but on the contrary, it increases the freedom of members to discover and research problems autonomously. Wenfeng Liang said that as the organization grows in the future, hierarchies and departments may become necessary in some areas, but that Deepseek intends to maintain its essential autonomy.

Deepseek's organizational operations are divided into two streams.

1. Top-down work It is an official task that requires company-wide collaboration and role division, such as V4 launch or large-scale learning. However, the rule is not to exceed half of the total time.

2. Bottom-up exploration It is a time for members to freely research and experiment on issues they each consider important. It can be carried out without separate KPI or prior permission.

Liang Wenfeng emphasized that relaxation and concentration are more important than tension and pressure in research. Deepseek explained that they generally don't work overtime much, and because they don't do a lot of work, the work itself is relatively small.

"Research requires a comfortable environment. If you push too hard, it will actually hinder the research."

3. Open source is not a compromise, but a core belief

Wenfeng Liang said that Deepseek's open source strategy was not a choice forced by competition, but rather stemmed from the vision the company had from the beginning.

He said that open source for other Chinese AI companies may seem like a choice based on market conditions, but for Deepseek, open source is an action that is consistent with the original purpose. The argument is that if AI is a technology so large that it affects the productivity and knowledge of all mankind, it is neither realistically possible nor desirable for one company to monopolize all value.

He explained that if AI is an industry that can account for a significant portion of global GDP in the long term, the more a company tries to monopolize that market, the more likely it is that it will be ostracized by society and the market.

"AI is too big a field. You can't have it all to yourself. You have to share it with others."

In the past, the open source software market was relatively small, so the profits a company could earn by disclosing source code could be greatly reduced. However, because AI has a much larger market size and influence, the logic is that it can create a sufficiently large business even if it earns only a reasonable level of profit.

Wenfeng Liang believed that Deepseek's giving up some of its profits through open source actually increases the possibility of reaching AGI in the long term.

"It requires moderation. The more you think, 'What percentage of human GDP is mine?', the less likely you are to succeed."

DeepSeek also stated that it does not operate the public model differently from the model used internally. In other words, rather than exposing a low-performance model to the outside and hiding a more powerful model internally, the model that the company distributes is actually the same as the open source model.

"We don't publish weak models and only use better models ourselves. It's the same model."

He also said he is not too worried about other companies deploying DeepSeek's model and becoming competitors. The position is that the overall AI market is large enough, and that it is helpful for Deepseek if other companies successfully deploy models and expand the ecosystem.

4. Business model that emphasizes reasonable profits and low prices

Deepseek explained that even in pricing, accessibility is prioritized over profit maximization. The API price is set based on the level at which equipment costs can be recovered within about 10 months, and the goal is not to increase profits beyond that.

For accounting reasons, servers can be depreciated over 3 to 5 years, but in actual business operations, recovering the equipment investment within 10 months is considered a reasonable profit.

"A reasonable standard for us is to pay for the equipment within about 10 months. We are not trying to maximize profits."

Wenfeng Liang told an anecdote about how his team members were happy when he lowered the price of one model to a quarter of the original price. A typical company would be concerned about a decrease in sales and ARR, but Deepseek internally accepted the fact that more people could use the model at a lower cost.

"We lowered the price by a quarter and the whole team was happy. I think that shows what we really think."

However, he explained that this does not mean that prices will continue to be lowered unconditionally. This is because the current price is already sufficiently accessible to most users, and if the usage or social utility does not increase significantly even if the price is lowered further, the meaning of further reductions is limited.

The key reason why DeepSeek can maintain low prices is model efficiency. Low cost is not just a marketing slogan, but a technical foundation for learning larger models with limited computing resources.

• Lowering the cost allows more users to use the model.
• A larger model can be trained with the same resources.
• Competitiveness can be secured in environments where GPUs are lacking, such as China.
• In the long run, you can gain an advantage in price competition.

Wenfeng Liang believed that competitors had relatively little incentive to lower costs. For commercial companies, lower costs can lead to lower prices and sales. On the other hand, at DeepSeek, members sympathize with the cost burden of users and connect the creation of affordable AI with the company's mission.

5. Moderation is not giving up, but a long-term strategy

The moderation that Liang Wen-feng talks about is not simply the attitude of giving up greed. It is a strategic choice that involves giving up some immediate benefits and opportunities for a more important goal.

Even when Deepseek's service suddenly spread globally around the Lunar New Year last year, it did not invest a huge amount of money or try to create a super app to retain users. The explanation is that they did not aim to become a huge Internet platform like ByteDance or Tencent, and did not try to monetize the traffic they secured as much as possible.

"We weren't thinking about becoming the next ByteDance or the next Tencent."

Liang Wen-Feng likened the idea that rather than trying to seize all the small opportunities in front of us, it is better to appropriately utilize the current opportunity for a bigger opportunity called AGI.

"You don't have to pick up all the sesame seeds in front of you. There may be a bigger watermelon ahead."

C-side users and B-side sales are not something that Deepsee ignores, but are seen as a natural by-product of the AGI research process. This means that although we will maintain and grow user and API revenue, we will not make that the company's top goal.

This is why DeepSeek chooses open source and moderation. We believe that giving up some sales and exclusive rights will increase the pride and solidarity of internal members, gain support from society and the ecosystem, and ultimately increase the probability of success in AGI research.

6. AGI is the priority, commercialization is a by-product

DeepSeek's long-term goal is AGI, or artificial general intelligence. Wenfeng Liang explained that the current C-side products, B-side services, and API businesses are all results derived from the process of moving toward AGI.

Unlike other companies that create models and use them for user acquisition or corporate customer service, Deepseek conducts AGI research and provides the technology created in the process in the form of APIs or products.

"We are not doing AGI for the C-side or B-side. By doing AGI, these results were able to be commercialized."

DeepSeek is already in the process of commercialization, but we believe that it is not yet time to completely transition to commercialization. Currently, it is believed that raising the limit of intelligence leads to higher returns than expanding product lines or designing various profit models such as advertising, e-commerce, and life services.

Wenpeng Liang pointed out that the speed of change in the AI ​​industry is so fast that commercialization strategies from a few months ago may become meaningless today. If you invest too much time in your product and business model, you may become buried in a business with a short life cycle.

However, the API business was evaluated as a safety measure that can generate significant cash flow even with minimal operation. He explained that only a small number of people are needed to maintain the API, and that users can naturally flow in without a separate large-scale sales organization or customer support system.

7. AGI Roadmap: From CoT to Implementation Intelligence

Wenfeng Liang believed that AI development does not proceed randomly, but develops along certain stages and bottlenecks. The rough roadmap he presented is as follows.

1. CoT, Expanding Your Thinking Process
2. Agent, an agent that performs tasks autonomously
3. Continuous learning
4. Singularity of self-improvement
5. Implementation Intelligence

CoT and Agent

He explained that an important development last year was CoT, or chain of thought. By allowing AI to think through problems step by step and review them on its own, it was able to show performance that surpasses that of humans in areas such as mathematics and programming.

The next step this year is Agent. Agents go beyond simply answering questions, but also plan and execute multi-step tasks. This increases the range of problems that AI can handle and the upper limit of its intelligence.

However, both CoT and Agent eventually reach their limits. Although AI's ability to solve problems increases, it cannot learn the organization and tasks over a long period of time and grow continuously like human employees.

Continuous learning

What Liang Wenfeng selected as a key condition for the next-generation model is continuous learning. Current AI requires a lengthy learning process or provides detailed context to learn new information. On the other hand, after joining a company, humans learn the environment, people, and work methods for a few months, and then perform tasks with only brief instructions.

"AI doesn't learn continuously yet. This is the next big breakthrough."

For example, if you tell a human employee, "Bring Xiao Wang," you already know who that person is, what his role is, and where he is. However, to current AI, all of this information must be explained one by one.

When continuous learning becomes possible, AI can accumulate experience over a long period of time, understand the context of users and organizations, and perform new tasks increasingly better. Wenpeng Liang said Only models capable of continuous learning can be truly called next-generation models.

Self-improvement and progressive singularity

After continuous learning, the AI can move on to developing the next version of itself and conducting research. Although this is often called a singularity, Wenfeng Liang emphasized that a singularity is not a single event that suddenly occurs one day.

"The singularity is not a single point, but a gradual process. It is not a sudden leap, but a long, continuous change."

As AI accelerates AI research, the pace of development may not be linear but gradually accelerate. Therefore, rather than appearing suddenly, the singularity is likely to be reached gradually through the cumulative process of AI helping research and creating better AI as a result.

Implemented Intelligence

Implementation Intelligence was presented as the final stage of AGI. This is because human needs do not exist only in the digital space, but also exist in the physical world, such as cleaning, care, and labor.

"People's needs lie in the physical world, not in computers. Ultimately, embodied intelligence is inevitable."

Once embodied intelligence is realized, AI can perform physical tasks through robots and even play a role in improving the next generation of robots on their own.

8. The model that Deep Chic wants to make first

Wenfeng Liang said that the first goal of the next model is not to serve general users or corporate customers, but to help DeepSeek's internal research and development.

"The first goal of our model is to make it useful to us."

The logic is that if DeepSeek's model can make internal researchers experiment and develop faster, the model is likely to be useful to other users as well. In particular, once continuous learning is solved, AI can participate in DeepSeek's research process and accelerate the development of the next version of the model.

Current agents do not continuously learn, so their scope of use is limited. However, it is predicted that if continuous learning becomes possible, AI will greatly increase the research and development efficiency of DeepSeek, and achieving AGI itself will become easier.

9. Areas to focus on and areas not to focus on

Deepseek does not intend to enter all areas of AI. The policy is to focus on main technology paths that are directly connected to raising the upper limit of intelligence.

High priority areas

• Large language model -CoT -Agent
• Coding Agent
• Continuous learning
• Model learning efficiency
• Expand model scale

Currently, coding agents are a particular priority. They announced that they plan to review vertical agents such as finance and medical services after coding agents.

Low priority areas

• 3D
• Create video
• Some world models
• Product areas not directly related to the nature of intelligence

Video creation can be a good commercial business, but it is far from the core path to raising the upper limit of AI's intelligence. He pointed out that the fact that many companies jumped into video creation after the launch of Sora was close to catching up with the market atmosphere.

I thought that multimodality was necessary, but not intelligence itself. The explanation is that the ability to process various inputs such as text, images, voice, and video is an important component of product functionality, but it must be distinguished from the main path to increasing the core intelligence of the model. DeepSeek plans to support native multimodal in a future version, but does not see this as the essence of intelligence.

Hallucinations are also a problem that can be improved, but Wenfeng Liang classified this as a product quality issue. The position is that the top priority for research is continuous learning and raising the upper limit of intelligence, and improving hallucinations is the next issue.

10. The gap between China and the United States is more about computing resources than talent

Wenfeng Liang argued that the biggest difference between the Chinese and American AI industries is computing resources, not talent. He explained that both countries share a significant pool of Chinese AI talent, and that not all outstanding talent is leaving the country.

"The gap with the United States is mainly in resources. The talent gap is not large."

The reason why China seems to lack talent is believed to be due to a limited number of experiments due to a lack of GPUs, rather than a difference in actual capabilities. If computing resources are insufficient, it is impossible to sufficiently learn models and conduct various experiments, and as a result, the growth rate of talent is slowed down.

DeepSeek announced that it had approximately 20,000 H100-class computing resources at the time and that it plans to purchase as many GPUs as possible in the future. It was evaluated that it would be much better to replace the secured funds with reasonably priced NVIDIA GPUs rather than storing them in the bank.

"If I could convert all my money into GPUs, I wouldn't hesitate to do so."

However, he acknowledged that it would be difficult to immediately compete with models with hundreds of billions of active parameters learned by top U.S. companies. The explanation is that while American companies can learn about 800B active parameter models, Chinese companies are still in the stage of experimenting on a scale of billions.

Deepseek's realistic strategy is to learn models with tens of billions of active parameters as efficiently as possible within limited resources, and to expand to larger scales such as 150B or 250B when resources increase.

11. Chinese AI chips and changes in the NVIDIA ecosystem

Wenfeng Liang did not mean that Chinese-made AI chips have completely caught up with NVIDIA right away, but he assessed that they are at a historic turning point.

NVIDIA's strength was its CUDA ecosystem, but as AI became able to write code, the cost and time to build the ecosystem were greatly reduced, he explained. Additionally, by using TileLang, a high-level language developed by DeepSeek, CUDA operators can be rewritten more quickly and dependence on the NVIDIA ecosystem can be reduced.

Wenpeng Liang believes that the entry barrier to CUDA can be lowered quickly for the following three reasons.

1. Using AI, you can quickly build a software ecosystem.
2. CUDA operations can be reimplemented through high-level languages ​​such as TileLang.
3. As the AI ​​computing market grows larger than the gaming GPU market, there is a greater possibility that dedicated AI chips will create a separate ecosystem.

DeepSeek announced that it was working with Huawei to apply the model to Chinese chips, and that it would receive approximately 16,000 Huawei 950 series cards at the time. He explained that although this is less than the hundreds of thousands secured by Internet giants, it can play an important role in helping the development of the Chinese chip ecosystem.

Wenfeng Liang assessed that although the Huawei 950 is roughly two years behind NVIDIA's latest product and may require multiple units to perform the same task, it is a practical replacement in terms of performance and price. The biggest problem currently is not technology or ecosystem, but production capacity and supply.

12. The final criteria for model competition are cost, speed, and user experience.

Wenfeng Liang predicted that in the long run, the performance gap between models will not widen significantly. Ultimately, he explained that competition is likely to be determined by the following three factors.

1. Cost The most important thing is to provide the same quality service at a lower price.

2. Time Competitiveness depends on who launches first and who provides new features and models faster.

3. User Experience The convenience and usability of the product and user loyalty to the service can make a difference.

Among them, the most important factors were considered to be cost and time. Although some differentiation in user experience is possible, fundamentally, companies that provide the same level of model more cheaply and quickly are at an advantage.

"The basic difference is cost and time. It matters who does it first and who does it faster."

Wenfeng Liang assessed that Anthropic's current situation ahead of OpenAI is not a permanent advantage but a temporary phase. He said that OpenAI and Google may continue to exchange leadership, and that Anthropic's code agent advantage is not overwhelming.

He emphasized that the differentiation of DeepSeek is not in the highest performance, but in low cost and low consumption rate.

13. China's basic model companies will eventually decline

Wenfeng Liang diagnosed that there are too many basic model companies in China. In the United States, resources are concentrated around three major companies, but in China, resources are being dispersed as many companies are developing similar models.

However, it was thought that it would be difficult for the profit margin of the AI ​​model business to remain very high in the long term. If each company tries to make huge profits, competitors willing to accept lower profits may emerge and lower prices and take over the market.

"Those who try to take more will be defeated by those who try to take less."

Therefore, it is predicted that China's basic model market will consolidate over time, and that ultimately only three or four major companies and a few small companies will remain. The analysis is that if the performance difference between models is not overwhelmingly large, competition over cost and speed of release will ultimately lower profitability and reduce the number of companies.

14. Deepseek's only core benefit is team stability

Wenfeng Liang said that Deepseek's core benefit is not sales, market share, or company size, but team stability.

"We have only one core interest: maintaining the stability of the team."

If members remain and can continue research, achieving AGI is ultimately a matter of time and trial and error. Money and resources can be procured from outside, but team stability cannot be compromised because the organization's knowledge and collaboration structure may collapse if key researchers leave.

He explained that the recent expansion of investment and compensation is a measure to support this goal. It was believed that providing sufficient stock options and compensation to key personnel would increase the likelihood that they would remain in the organization, and that if core personnel were stable, other members would not easily leave.

Deepseek's choice of open source, low price, and competitor support ultimately helps the team's unity and vision. The explanation is that organizational commitment increases when members feel that their work is socially meaningful and that they are contributing to the goal of AGI.

15. Business scope that does not try to win everyone

DeepSearch does not attempt to vertically integrate all areas of the AI industry. Rather than taking full control of applications, hardware, chips, and customer service, the company said it wants to focus on the core areas it does best and leave the rest to its partners and other companies.

"AI is a big enough field. We don't need to do everything."

Although we build large-scale data centers and clusters ourselves, we are cautious about developing our own chips. We believe that if chips can be purchased at a reasonable price in the market, there is no need to produce them ourselves. I heard the analogy that just because you run a power plant, you don't have to build the generator yourself.

Deepseek said it is happy to be one of several trillion-dollar companies in the AI ​​era and has no ambitions to dominate the entire industry or monopolize all businesses.

"If there are multiple trillion-dollar companies in the AI ​​era, just becoming one of them is enough."

16. Bottleneck of data and high-quality labeling

Wenpeng Liang estimated that data accounts for about half of the model's performance. Open data is difficult to use as a long-term differentiator because it can be used by multiple companies, but high-quality data and labeling are still important bottlenecks, he explained.

American companies can invest in high-cost data labeling based on enormous capital, but it is difficult for Chinese companies to be cost-competitive in the same way. Accordingly, DeepSeek announced that it is labeling relatively inexpensive data first and devoting a significant portion of its research staff to data-related work.

We believed that building high-quality data is not a problem that can be quickly solved with funds alone, but is a problem that requires time. Since Chinese companies began to develop AI in earnest later than American companies, it was predicted that the problem of high-quality data could gradually improve after about a year.

Regarding whether synthetic data or real data is more important, it was evaluated that synthetic data can sufficiently surpass human knowledge. Just as AlphaGo discovered a number that humans have never tried before, AI can create new possibilities based on existing knowledge.

17. Balance between commercialization and research institutes

As DeepSeek entered the capital market, questions arose about how to maintain a balance between pure research institutions and commercial enterprises.

Wenfeng Liang responded that Deepseek has already secured C-side users and B-side sales, and that the API business alone can sustain the company in the worst case. The explanation is that if B-side sales reach hundreds of millions of dollars and demand continues to increase, the company will be able to cover research and development costs and operating costs on its own.

Therefore, even if technological development is temporarily stagnated, we believe that there is a realistic safety measure that can survive with only API services. However, he explained that that is not the final goal of DeepSeek, and that the ideal direction is to continue AGI research while ensuring survival.

Deepseek's decisions are not made solely by the founder. Wenfeng Liang said that he considers the consensus of members important in making major technology, business, and strategy decisions, and that his influence is also based on consensus.

18. Organizations are not imitations but products of the times

When asked whether DeepSeek's organizational model was an imitation of Bell Labs or other famous research institutes, Wenpeng Liang responded that there was no model in particular as a reference.

"We are not imitating anyone. We are a product of the times and reality."

Unlike research institutes like Bell Labs that do not have to worry about commercial survival, Deepseek is a company that must make its own profit and survive. Therefore, we acknowledged that even if we pursue a grand mission, we need realistic revenue sources such as B-side sales and API business.

Deepseek's organizational structure is currently very flat, but as headcount grows, hierarchies and departments may be needed in some areas. He explained that these changes are already starting to occur as the organization grows, and that all areas cannot remain unstructured forever.

19. Future models and release plans

Wenfeng Liang said that he is comfortable with Deepchic's model release cycle of once every 2-3 months. Each version should be improved over the previous version, and V4 and later models are planned to support multimodal features.

He explained that it is difficult for current models with billions of active parameters to make a significant difference in inference speed and performance compared to open source competing models, but there is still a gap compared to larger private models. The next step was to start learning a 150B model at or after the end of the year.

However, he reiterated that the true next-generation model should not be simply a faster or cheaper model, but a model with continuous learning capabilities.

20. Remaining questions and Liang Wenfeng's final perspective

In the second half of the meeting, questions were asked about the technical challenges of continuous learning, the timing of AGI, the pace of development of Chinese hardware, embodied intelligence, coding agents, the hallucination problem, and vertical applications.

Wenpeng Liang explained that continuous learning is not a single technology, but a research problem that combines multiple algorithms, systems, and engineering elements. Although no one has found a sufficiently effective method yet, he said there are ideas within DeepSeek that look promising.

No specific year was provided for when AGI was reached. Although China can narrow the gap with the United States within 1-2 years within the current model paradigm, it is difficult to call it AGI unless continuous learning is solved.

Chinese-made hardware may take several years to restore confidence in the ecosystem and increase production capacity, but the company is optimistic about the long term, he said. It is also predicted that although it will be difficult for China to reach the same level as the United States overall, it may be able to lead in certain fields with fewer computing resources.

In the end, the direction of deep chic that Wenfeng Liang envisions can be summarized as follows.

• AGI research is given top priority.
• Technology is widely shared through open source.
• Take only reasonable profits and lower prices.
• We consider cost efficiency as our core competitiveness.
• We view team stability as the most important asset.
• We do not try to monopolize all markets and businesses.
• Continuous learning is considered a key breakthrough in next-generation AI.
• Ultimately extends AI to the physical world.

Conclusion

The biggest characteristic of Deep Seek revealed in the minutes of this investor meeting is extreme concentration and restraint. Deepseek prioritizes the long-term goal of AGI over immediate goals such as user traffic, API sales, product expansion, and market exclusivity, and seeks to encourage growth of the entire ecosystem through open source and low prices.

Wenfeng Liang emphasized that Deepseek is not a company that only attracts richer or more talented people. Rather, he explained that they compete using vision, organizational style, cost efficiency, and team cohesion within limited resources.

"The moment you have a vision to have more, you have already lost. The world rewards those who want to have less."

Whether Deepseek can actually maintain this strategy to the end depends on future competition and market changes. However, at least in these meeting minutes, Wenfeng Liang clearly showed that he views Deepseek not as just another AI product company, but as a long-term research organization toward AGI and the center of the open source ecosystem.