Liang Wenfeng Investor Meeting: DeepSeek’s Vision, Open-Source Strategy, and AGI Roadmap
目录
TL;DR
I. A Vision-Driven Organization
II. Why We Remain Committed to Open Source
III. API Pricing and Reasonable Profits
IV. Not Pursuing Every Commercial Opportunity
V. Open Source and Commercial Monetization Are Not in Conflict
VI. From Agents to Continual Learning
VII. Consumer and Enterprise Businesses Are Both By-Products of AGI
VIII. The One Core Interest We Cannot Compromise
IX. Focusing Exclusively on the Path to AGI
X. World Models, Multimodality, and the Upper Bound of Intelligence
11. Compute Is the Primary Gap with the United States
12. The Endgame of Large-Model Competition
13. Why We Are in No Rush to Expand the Product Portfolio
14. Consensus-Based Decision-Making
15. Global Model Competition and the Role of Chinese Companies
16. Domestic Chips and the Changing CUDA Ecosystem
17. Collaboration with Huawei and TileLang
18. Whether to Integrate Vertically Upstream and Downstream
19. B2B Demand, Multimodality, and Scaling
20. Next-Generation Models and the Low-Cost Approach
21. Why Restraint May Win Out
22. Top-Down and Bottom-Up Management
23. Q&A: How Long Will Continual Learning Take?
24. Q&A: The Open-Source Ecosystem and Industry Consolidation
25. Q&A: Real-World Data, Simulated Data, and Surpassing Humans
26. Q&A: Compute Depreciation and the Path to Catching Up
27. Q&A: Technical Challenges in Continual Learning
28. Q&A: Huawei 950, AGI Timeline, and Organizational Changes
29. Q&A: Release Cadence and Larger Models
30. Q&A: Whether AGI Has a Critical Threshold
31. Q&A: Allocating Resources Between Defined Objectives and Exploratory Research
32. Q&A: Hallucinations and High-Quality Data
33. Q&A: Structural Advantages of Chinese Models
34. Q&A: Post-Training Investment and Data Bottlenecks
35. Q&A: Vertical Agents and Coding Agents
36. Q&A: Balancing Research Ambitions and the Capital Markets
37. Q&A: Is There a Reference Model for the Organizational Structure?
38. Q&A: Will TileLang Sacrifice Execution Efficiency?
本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读
Liang Wenfeng systematically outlined DeepSeek’s vision, its rationale for open sourcing and pricing, its technical roadmap toward AGI, and its compute plans. The company sees continual learning as the next major breakthrough. Commercialization provides a safety net, but team stability, compute availability, and high-quality data remain the key variables.
TL;DR
DeepSeek does not seek to maximize near-term commercial returns. Its core vision is to advance AGI and make AI useful to society. The company unites its team around a shared vision and consensus, operates without KPIs, and views “restraint” as a strategy for increasing its long-term probability of success.
Open sourcing reflects both DeepSeek’s vision and a commercial choice. Liang believes the AI market is large enough that open sourcing will not undermine reasonable profits. The open-source model is identical to the one used in DeepSeek’s proprietary services, while API pricing is benchmarked to recovering equipment costs in roughly ten months.
Consumer users, enterprise revenue, and API services are by-products of the path toward AGI rather than primary objectives. If demand continues to grow, API ARR could reach several hundred million dollars or even $1 billion, sufficient to cover R&D; and operating costs.
DeepSeek sees the technical progression toward AGI as language models, chain-of-thought reasoning, agents, continual learning, self-iteration, and ultimately embodied intelligence. Continual learning is the next clearly defined bottleneck for agents. Once overcome, AI will materially accelerate the development of subsequent model generations.
The company’s only non-negotiable core interest is team stability, while its largest external bottleneck is compute. Current compute capacity is equivalent to approximately 20,000 H-series GPUs. DeepSeek will procure as much compute as possible at reasonable prices; spending RMB20 billion on compute in a single year would indicate highly successful procurement execution.
The primary gap between China and the US lies in resources rather than talent. Chinese models are mainly exploring architectures with tens of billions of active parameters, while the largest US models have approximately 800 billion active parameters. Comprehensive leadership is unrealistic, but China can lead in selected areas through efficiency and strategic trade-offs.
Domestic AI chips have an opportunity to reshape the ecosystem. High-level languages such as TileLang can reduce dependence on the CUDA ecosystem. Huawei’s 950 SuperPoD can handle workloads supported by the GB200 and GB300, but still faces a gap of roughly “four times the hardware plus two years,” as well as insufficient production capacity.
Long-term competition will primarily center on cost, time, and user experience, with cost and time being the most critical. DeepSeek aims to build an advantage through lower costs, higher efficiency, and faster iteration. Commercialization provides the foundation for survival and a downside-protection path, but the company currently continues to prioritize its core AGI agenda.
I. A Vision-Driven Organization
Liang Wenfeng:
We did not start this company to maximize how much money we could ultimately make or to pursue a public listing. If that had been the objective, the first few dozen people would not have joined. We approached the world with goodwill and believed this was something useful to humanity—something beyond money. As the stakes have grown, other temptations have emerged, but neither our original intent nor the vision we have maintained to this day is to maximize commercial returns.
About 20 years ago, the management figure I admired most was former GE CEO Jack Welch. In retrospect, most of what he said may have been wrong, but he was right about one thing: large companies are managed through vision, not rules and systems. A vision is not a slogan on the wall. It is reflected in how a company acts and operates, not in what it says. I no longer remember his exact words, but that was the essence.
We do not have an organization in the conventional sense. Instead, we organize people around a vision, without KPIs or performance assessments. This has both advantages and disadvantages, and we will build on the strengths while addressing the weaknesses over time. The vision is not codified either. Everyone interprets it differently, but we are aligned on the broad direction, as reflected in how we work and how we view the world. This vision is genuine; otherwise, many of our choices would be impossible to explain.
Each person at the company has a somewhat different personal vision, and interpretations of the shared vision are not identical. Even so, everyone remains aligned on the same broad direction. We continue to act with goodwill toward the world. This is the foundation that holds the team together and the starting point for all my subsequent judgments.
II. Why We Remain Committed to Open Source
Our vision explains why we remain committed to open source: the vision itself requires it. Zhipu AI, for example, also open sources its models, but it feels somewhat compelled to do so. For us, open source was always the original intent. We were clear from the outset: first, it serves our vision; second, open source also has commercial benefits if the objective is to make AI successful.
Historically, open source has conflicted with commercialization. A software product may have an annual addressable market of only several billion dollars, which could shrink to tens or hundreds of millions once the software is open sourced. AI is different. It could ultimately account for 10% of global GDP. Benefits on that scale cannot be monopolized by any one party and must be shared with others; otherwise, resistance is inevitable, and history will leave that party behind.
This reflects both an objective principle and a view of history. Keeping a model closed source does not mean one can monopolize the market; that is neither theoretically sound nor consistent with reality. When a company attempts to monopolize benefits on such an enormous scale, other forces and mechanisms will inevitably emerge to prevent it.
AI therefore cannot be approached entirely through conventional commercial thinking. A mechanism is needed to limit the share of benefits one can capture. The more aggressively one seeks to capture a large portion of China’s or the world’s GDP, the less likely one is to succeed. Greater restraint instead increases the probability of success. Open source is one element of that restraint.
When we founded the company two years ago, we did not have much capital, compute, recognition, or influence. We were simply a group of ordinary people. I prefer the idea of “a group of ordinary people accomplishing extraordinary things” to a narrative centered on a group of geniuses. We were no better funded than others, our people were not necessarily better, and I myself graduated from an ordinary university.
Nor was our team a carefully selected group of geniuses assembled in advance. It was closer to a group of ordinary people who came together by chance. We started from a very low base with very limited resources, yet we have made it this far. At a minimum, this shows that restraint has not weakened us and may instead have made the organization stronger.
The potential value of AI is large enough that, if we succeed, even a small share will be more than sufficient. There is therefore no need at this stage to calculate which portion to capture or how to capture it. We seek only reasonable profits, as determined by our own intent. To date, restraint and vision help explain how we have achieved what we have from such a low starting point.
The question of whether open source conflicts with commercialization therefore cannot be answered using conclusions drawn from traditional software markets. For us, increasing the probability of making AI successful first—and then earning a limited, reasonable return from an enormous market—is more commercially realistic than pursuing monopoly from the outset.
III. API Pricing and Reasonable Profits
Our API pricing follows a simple principle: when we purchase a batch of equipment, we aim to recover its cost in roughly ten months. Servers may be depreciated over three or five years for accounting purposes, but commercially, we consider a ten-month payback period to represent a reasonable profit. V3.2 Flash and our other models are all priced according to this principle.
This also accounts for risk and upfront investment; pricing is not based solely on the equipment depreciation period. Even after factoring in these considerations, however, we still see no reason to raise prices to the maximum level demand can bear.
This is not profit maximization. Demand is inelastic within the current price range. Even if prices rose by 50% or 100%, token consumption would likely change little, and total revenue could nearly double. We had a DDCP model that we initially priced relatively high because we were concerned about excessive demand, which disappointed the team. When we later cut the price to one-quarter of the original level, everyone cheered. We build models to make them affordable and effective so that people can use them extensively, while still earning a reasonable profit.
For competitors, price cuts reduce revenue and ARR and are therefore naturally unwelcome. For us, they generate a sense of accomplishment within the company while also benefiting society. A ten-month payback period is sufficient to satisfy the company and acceptable to users.
After the price cut, many people celebrated in the company group chat. The team had invested substantial effort in building a strong model and genuinely wanted it to be both affordable and effective so that more people could use it extensively. This reaction was not for publicity; it reflected the team’s genuine motivation and consensus. A company that prioritizes revenue would not typically respond to a price cut in the same way.
Some argue that a ten-month payback period still implies excessive profits. There is indeed further scope for model optimization and broader price reductions, but we can achieve these efficiencies while Alibaba and Tencent lack comparable optimization and may face costs several times higher. We are not cutting prices further because demand is already inelastic: additional reductions would not materially increase demand, revenue, or social value.
Prices are already affordable enough that users are not deterred by cost. Further cuts would generate no additional revenue for the company and would not meaningfully improve user satisfaction. There is therefore no need to reduce prices simply to signal affordability.
Our pricing does not seek to maximize revenue or profits. Price increases might generate more revenue in the short term, but not necessarily over the long term. Restraint is a strategy: sharing value strengthens team cohesion while benefiting peers and the public, thereby increasing our probability of achieving AGI. I have no doubt that AGI has enormous commercial value. The priority is therefore to increase the probability of achieving it, rather than capturing a slightly larger share.
IV. Not Pursuing Every Commercial Opportunity
When user numbers surged unexpectedly during last year’s Lunar New Year holiday, we did not focus on retaining and monetizing users or capturing commercial value. Instead, we concentrated on delivering the best possible service. We do not aspire to build the next super app, ByteDance, or Tencent. Investing heavily to compete for consumer users is commercially viable, but we chose restraint because the larger AGI opportunity lies ahead. Even a relatively large near-term opportunity should not distract us from moving forward.
Competing head-on with ByteDance for users at the time could also have constituted a coherent commercial strategy. We did not dispute its viability; we simply concluded that it was not the direction that would maximize the company’s returns at that stage, so we deliberately opted out.
In retrospect, not pushing aggressively into the consumer market may have been the right decision. We have simply maintained daily active users at low cost. Although this is currently a pure expense, it may prove useful later, so we retain what comes easily.
Had we invested heavily last year to scale the consumer business, it is unclear what we would actually have gained by now. User scale is not an end in itself, nor should we transform a research organization into a product company competing across the board for traffic merely to prove that we have a seat at the table.
We will still pick up opportunities along the way, but only opportunistically; we will not stop and make them our core business. I do not even need to decide today what position we will occupy in future AGI commercial opportunities or what business model we will adopt. As long as the opportunity is large enough, we will find a way when the time comes.
There is also an opportunity in API or AI ARR this year. If demand continues to expand and we can procure more GPUs, reaching several hundred million dollars is highly plausible. At $1 billion, the company could essentially become cash-flow positive, covering R&D; and all other expenses. We will pursue this and execute well, but it will not be our top priority. Last year’s consumer business and this year’s enterprise business are both important, but neither is comparable to AGI.
V. Open Source and Commercial Monetization Are Not in Conflict
We will open-source our models, potentially including our strongest model, because I see no inherent advantage to keeping them closed. Even if we disclose the models in full, deploying them effectively while also reducing costs remains a high-barrier undertaking. Making the underlying principles public does not mean every company has the willingness or capability to organize the people required for deployment and optimization.
This is also the sweet spot for a startup of our size: smaller companies lack the resources, while large companies face managerial and physical constraints. Even with open source, we will continue pricing for an approximately ten-month payback period. At this price, independent third-party deployments are already uneconomic because they cannot match our cost structure.
Smaller teams lack the resources to complete the full deployment and optimization stack; larger organizations may encounter internal resistance. We may face new problems as we grow, but our current scale is particularly well suited to executing this thoroughly.
If we sought a 100-fold profit margin, open source would affect revenue because third parties could offer the service at 20 times cost and still undercut us. But because we target only about six times cost and a ten-month payback period, open source has no impact. Over time, a reasonable margin may fall to four times or three times cost; much lower may not be realistic, but commercial monetization would remain sustainable.
Open source gives us more opportunities to operate at the technological frontier and increases our probability of achieving AGI. It does not conflict with commercial monetization, provided we capture only reasonable profits. Selling APIs itself is not especially compelling to me, but it at least demonstrates that the model works.
The outside world may view our commitment to research and open source as choosing the hardest possible mode. In reality, we have given up many things elsewhere and do not need to pursue every business simultaneously. This still allows us to operate comfortably, even without relying on overtime. That, too, is a result of restraint.
We do not worry about others deploying our models and competing with us. On the contrary, we hope they will deploy them, and we will do our best to help the open-source community address performance and cost issues. Last year, some worried that Tencent and other traffic-rich platforms would deploy our models and take away consumer users. Yet our open-source models are identical to those used in our own services, and we have seen no such conflict over the past year.
If others achieve poor performance and incur high costs because they mishandle a detail, our greater concern is that they will fail to put the models to genuine use—not that they will take our business. The market is large enough, and helping the community reproduce and deploy our models is consistent with our understanding of open source.
VI. From Agents to Continual Learning
The company’s long-term vision is AGI. With this generation of technology, AI can already outperform humans when a problem is clearly defined and complete context and instructions are provided. The difficulty is that these conditions are rarely met in the real world. A single meeting is backed by extensive context, and each participant may bring decades of experience that AI does not possess.
People can learn continuously. After spending two months learning how a company works, a new employee knows who “Xiao Wang” is, where to find him, and how to communicate with him. AI lacks those two months of learning and must be given the full background every time. It can therefore complete individual tasks but cannot yet truly replace an employee. The next missing capability is “learning how to learn.”
AI development resembles a staircase: first came language models; last year, we crossed the CoT—or chain-of-thought—threshold; this year is about agents. Each step builds on the previous one: agents require CoT, and CoT requires language models. Even after CoT surpassed the best humans in mathematical olympiads and programming, it still faced a ceiling. Agents expand the range of capabilities but will also reach a ceiling because they cannot learn continuously.
It is a staircase because none of the preceding steps is wasted. Agents can solve more tasks, but once every problem solvable at this stage has been addressed, they still cannot work over the long term like employees. That is the capability ceiling, and it cannot be crossed simply by making them somewhat better at individual tasks.
The challenge after agents, therefore, is continual learning: enabling models to learn over the long term rather than relying on a single intensive training cycle each time. There must be a way through this bottleneck, but it will take time. Once continual learning is achieved, models will be able to do what humans can do and develop the next generation of models, enabling self-iteration. This can conventionally be called the singularity, but it will not happen suddenly; it may instead be a gradual, continuous process unfolding over an extended period.
Continual learning and completing long-horizon tasks are fundamentally the same problem. We now stand at the agent stage and can already see the next bottleneck, but seeing the problem is not the same as knowing the answer. The obstacles ahead still require research and time to overcome.
Embodied intelligence comes only after that. Our roadmap is to first solve “learning how to learn,” then move into self-iteration, and finally bring AI into the physical world to perform household chores and care for the elderly. Different people may have different roadmaps. We chose this path because each step requires relatively few new elements, and later technologies can be developed with the help of capabilities built earlier. Starting with embodied intelligence would be arduous; if continual learning and self-iteration are achieved first, models can help develop embodied intelligence. This is our long-term AGI objective.
VII. Consumer and Enterprise Businesses Are Both By-Products of AGI
Last year, the industry competed for consumer chatbot traffic; this year, it is competing for enterprise revenue, as though companies that do not participate have no seat at the table. Internally, however, what truly matters to us is the AGI roadmap and the next technological breakthrough.
With AGI as our objective, applications and consumer and enterprise businesses require relatively little additional effort because we have the advantage of working downward from a higher technological position. Last year, we put little effort into the consumer business and at one point did not even want to maintain it, yet the users would not leave. This year, enterprise revenue growth also looks relatively encouraging.
Our enterprise figures this year should compare favorably with peers, but once again, we have not devoted substantial effort to them; they have largely emerged as a by-product. This does not mean the business need not be run well. It means the business can commercialize interim outputs from the core technology roadmap.
