What Liang Wenfeng DeepSeek Thinking Means for AI App Developers

This article starts from a closed-door investor communication held after DeepSeek completed a first external financing round reportedly exceeding RMB 50 billion. Liang Wenfeng spoke with all institutional investors in an in-depth Q&A format. The session was not publicly released, but a transcript of the recording went viral across the Chinese internet on July 23, giving the market a rare look at how DeepSeek thinks about AGI, open source, model capability, organizational boundaries, and the application ecosystem.

Note: The following content is compiled based on the transcript of the recording of the “Liang Wenfeng Investor Meeting.” The original transcript shows clear signs of speech-to-text transcription, including colloquial repetition, disordered word order, and localized recognition errors. This article is not a verbatim transcript; rather, without changing the core views, it streamlines, consolidates, and rewrites the remarks in a more formal written style. The quotations in the article are polished excerpts based on the identifiable original text, intended to preserve the down-to-earth expressions and examples in Liang Wenfeng’s remarks.

Edited Transcript of Liang Wenfeng’s Remarks

The Main Thread Is Not Commercialization, but AGI

DeepSeek’s most important goal today remains advancing AGI.

In conversations, Liang Wenfeng repeatedly emphasized that the company was not founded for the purpose of “how much money it would ultimately make, going public, or entering the capital markets.” The earliest group of people started doing this with a simple desire to be useful to AI and to the development of human intelligence.

*“At the very beginning, we did not think about how much money we would ultimately make, whether we would go public, or what we would do in the capital markets. At first, it was just a few people, then dozens of people, who held great goodwill toward the world and felt that this work would be useful to humanity, so they came to do it.”*

This way of putting it is not ornate; it even runs somewhat counter to the usual business narrative. But it explains many of DeepSeek’s later choices: why it insists on open source, why it is not in a hurry to monetize users, why it does not push API pricing toward profit maximization, and why it continues to emphasize restraint in the face of many temptations.

*“A vision is not a management slogan hung on the wall. A company does not run on what you say or how its rules and systems are written, but on what you actually do. We don’t really have a particularly complex organizational approach; we are simply vision-driven.”*

DeepSeek will pursue commercialization, and it has been doing so all along. Consumer users, enterprise revenue, and API services can all provide support for the company. But these are not the ultimate goal. They are more like byproducts and support systems along the main path toward AGI, rather than the entire reason the company exists.

*“Consumer products, enterprise services, and API are all products of our journey toward AGI. We are not building models in order to serve consumer users or enterprise users; our original intention is still to pursue AGI.”*

That is why he does not believe much time should be spent now discussing product lines and commercialization paths. AI is changing too quickly; a business path that looks clear today may no longer hold after some time.

*“If we spend a lot of time now discussing product lines and commercialization paths, I think it is all a waste of time. Because you cannot foresee the future, and even the experience so far supports this judgment: it is still far too early to talk about commercialization paths and product lines.”*

This is not a rejection of commercialization, but a matter of putting commercialization back in its proper place: commercialization should serve the main thread, not pull it in the opposite direction.

Open Source Is Not Anti-Commercialization; It Just Doesn’t Take All the Profit

Open source and commercial interests are not inherently in conflict.

In discussions, Liang Wenfeng put the issue very plainly: if a company wants to capture all the profit, then open source will certainly weaken its commercial upside; but if it only seeks reasonable returns, open source may instead expand the ecosystem, strengthen collaboration, and lower the barrier to entry for users and developers.

*“If I want to make a hundredfold profit, open source would indeed affect my ability to make that hundredfold profit. Because a third party can deploy it, and maybe their cost is twenty times, which is lower than mine. But if we only make a reasonable profit, there is no conflict between open source and commercial payment.”*

Behind this statement is a broader judgment: what open source truly changes is not the technical path, but profit distribution. Closed-source models make it easier to concentrate profits in the hands of the platform; open-source models leave more profit to developers, deployment providers, application teams, and end users.

DeepSeek’s API pricing logic is also not “how much users can still bear,” but rather “how long it takes to recover hardware costs, with a reasonable profit left over.”

*“Our API pricing logic is roughly based on recovering equipment costs in ten months. It is not profit-maximizing pricing. If the price doubled again, token consumption might not differ much and revenue would nearly double, but that is not our starting point.”*

He also mentioned a very down-to-earth detail: at first, the team worried that demand would be too high, so they set the price relatively high; later, when the price came down, the team was actually very happy. Because everyone felt that this thing could finally become affordable for more people.

*“At first, we were worried demand would be too high, so we set the price relatively high. Later, when the price dropped to one quarter, everyone was very happy. Not because we were making more money, but because we felt this thing was useful to people and could finally be affordable for everyone.”*

So DeepSeek’s low pricing is not merely a price war; it is an ecosystem strategy. Low prices reduce the cost of trial and error, giving more developers the confidence to use, experiment with, and deploy it. The lower the cost of calls, the more easily AI can enter real workflows; the larger the ecosystem, the more opportunities the model itself has to be used, receive feedback, and improve.

*“Restraint is also a strategy. Sometimes you can give up certain things in exchange for more of other things. Conceding profits is beneficial internally for the company, for society, and for peers. In the long run, it can increase our probability of achieving AGI.”*

This is also the “goodwill” Liang Wenfeng has repeatedly emphasized. He does not see competitors reproducing the model as an absolute threat, and even says that if others are willing to deploy and reproduce it, DeepSeek will help as much as possible.

*“We hope others can deploy our model, and we are not worried that they will take business away from us. We only worry that they may fail to deploy it, or that certain details may not be done correctly, causing the results to degrade. We will do our best to provide help to the open-source community.”*

Competition Will Come Down to Cost, Time, and User Experience

When discussing the AI gap between China and the United States, Liang Wenfeng did not say that talent is the biggest issue. He emphasized resources more, especially computing power.

In his view, the most practical gap between Chinese teams and leading U.S. teams lies in training resources, the number of chips, experimental conditions, and iteration density. U.S. companies have more computing power, allowing them to build larger models and run more comprehensive experiments; Chinese companies must acknowledge the real gap in this regard.

*“The biggest gap between us and the United States is in resources. Talent is not the bottleneck; resources are the biggest bottleneck. Resources first affect talent development, and they also affect how many experiments we can run and how large a model we can train.”*

But he did not mythologize U.S. models as impossible to catch up with. On the contrary, he believes that the ultimate competitive gap in large models may not lie in an ability divide that can never be crossed, but in three areas: cost, time, and user experience.

*“Where will the ultimate gap in large models show up? I think it comes down to three areas: cost, time, and user experience. Beyond that, there may not be much of a gap.”*

This statement is important. It brings model competition back from “absolute intelligence” to real-world business systems: Can you provide the same quality at a lower cost? Did you build it a few months earlier, or a few months later? Is the user experience a bit faster, or a bit slower?

*“Cost is definitely one differentiator. The second is time: whether you build it a few months earlier or a few months later makes a difference. The third is experience: with the same thing, whether the user experience is a little faster or slower will also make a difference.”*

This also means that model capability determines whether something can be demoed, while cost structure determines whether it can become a product. If capability gaps gradually converge, the truly long-term competition will shift toward engineering efficiency, inference cost, product experience, and ecosystem building.

Model Efficiency Is Not Just Technical Optimization; It Determines Whether a Product Is Viable

Liang Wenfeng places great importance on model efficiency and inference costs.

But his understanding of low cost is not simply “selling it a bit cheaper.” Low cost is first about whether more users can afford to use it, and second about whether larger models can be trained and run when compute is limited.

*“Many of our colleagues want to drive costs even lower, because they know this costs us money, and they also hope it will be less burdensome for others to use. If it is a bit cheaper, people will be more willing to accept it.”*

This is a very straightforward product philosophy: if AI is expensive, it can only be used by a small number of people; if AI is cheap enough, it can enter more production environments. Cost reduction is not a financial detail, but a prerequisite for the widespread adoption of AI.

*“The lower the cost, the higher my computational efficiency when compute is limited, and the more capable I am of supporting larger models. For large companies, they can solve the problem by adding resources; but we prioritize cost efficiency.”*

For application teams, this judgment is especially important. Many AI applications do not fail because the demand is invalid, but because the cost model does not work. A demo may look impressive, but once it enters high-frequency use, model costs, latency, and failure rates immediately become business issues.

Continuous Learning Is the Key Challenge of the Next Stage

In the section on the technical roadmap, Liang Wenfeng repeatedly mentioned CoT, Agent, and continuous learning.

He believes the development of AI is like climbing steps one level at a time. Language models are one step, CoT is another, and Agent is yet another. But once Agent progresses to a certain point, it will encounter the next bottleneck: continuous learning.

*“The step we are taking this year is Agent. With an Agent-based approach, many things can be done, and the scope of capabilities will become broader. But Agent still cannot replace employees, because it still lacks continuous learning.”*

He used a very down-to-earth example to explain the issue:

*“When you hire an employee, they may spend two months getting familiar with the company environment and the work. Later, when you say, ‘Ask Xiao Wang to come over,’ they know who Xiao Wang is. But AI cannot do that. You have to tell it who Xiao Wang is, what role he has, where he is, and what to pay attention to when approaching him. You cannot provide all the context to AI every single time.”*

This accurately describes a pain point in many AI products today. A model may be very capable in a single interaction, but it has weak long-term organizational memory. It can answer questions, but struggles to accumulate experience like an employee; it can execute tasks, but struggles to understand the company better over time.

*“If AI had continuous learning capabilities and could spend two months learning at a company like an employee, then it could replace many people. Right now, the next step still requires ‘learning how to learn.’”*

Therefore, continuous learning is not a minor feature, but a critical capability on the path toward more powerful Agent. For the application layer, it does not necessarily mean continuously training model parameters; it can also be achieved through memory, knowledge bases, workflow records, feedback loops, and the accumulation of organizational data.

The future moat of AI applications will not be how impressive the first answer is, but whether, by the twentieth use, it understands the user, the business, and the company better.

The Challenge for Domestic Hardware Is Not Just Chips, but the Ecosystem

When discussing domestic AI chips, Liang Wenfeng’s assessment is very pragmatic: in the short term, the issue is not only whether the hardware exists, but whether an ecosystem exists.

NVIDIA’s advantage lies not only in the cards themselves, but also in CUDA, the toolchain, developer habits, and the ecosystem built up over many years. For domestic cards to become truly usable, issues around adaptation, compilers, training stability, inference efficiency, and the software ecosystem all need to be addressed.

*“Building an ecosystem for domestic cards that matches NVIDIA’s, or even surpasses it, is not without obstacles, but it will take time.”*

He also mentioned TileLang, which uses a higher-level language and compilation approach to reduce dependence on the CUDA ecosystem. There is no need to get bogged down in specific technical details here; what is truly worth noting is another judgment: AI will, in turn, rebuild AI infrastructure.

*“TileLang is a high-level language. In the past, we could not do without the CUDA ecosystem. Now we can rewrite everything with TileLang, with less code and much higher efficiency. A 1% to 2% loss in execution efficiency is acceptable to me.”*

In the past, the CUDA ecosystem was difficult to work around. But if AI can help write code, build compilers, and optimize operators, the cost of developing a domestic hardware ecosystem may decline. Competition in domestic AI hardware is not only about hardware performance, but also about the software ecosystem and development efficiency.

The Keyword in Organization Is Focus, Not Looseness

DeepSeek’s organizational approach has two lines: one top-down, and one bottom-up.

Top-down means that key releases and key goals require company-wide collaboration. For example, when an important version is to be released, work needs to be divided, with each person responsible for one part.

Bottom-up means researchers need to have room for their own exploration. Not everything is driven by KPI, and not all research is assigned.

*“Our company’s management has two lines: one top-down, and one bottom-up. Bottom-up means everyone does what they want to do; no one manages them, and there are no KPI.”*

But this should not be simply understood as “loose management.” The premise of having no KPI is that the company knows what not to do. If a company lacks a clear direction, has a chaotic product line, and takes on every customer request, then trying to learn “no KPI” will only end in loss of control.

Liang Wenfeng has also said that formal releases and critical tasks still require top-down collaboration. It is just that such formal arrangements should not take up all the time; room must be preserved for research and exploration.

*“Top-down means that when we formally need to release a version, everyone works together and each person does one part. But we hope formal arrangements will not take up all of employees’ time—ideally no more than half—so they still have time to explore on their own.”*

On overtime, his view is also very direct: DeepSeek generally does not work much overtime. One reason is that research requires a relatively relaxed environment; the other is that the company is extremely focused, so there are fewer things to do.

*“We generally do not work much overtime. First, doing research requires a relatively relaxed environment; second, we are highly focused, which means there are very few things we need to do. When there are fewer things, each person has less work assigned, so there is no need for overtime.”*

Behind this statement is still restraint. The company does not do everything, does not seize every opportunity, and does not enter every commercial direction. Management is not about filling up the team’s schedule, but about helping the team reduce ineffective actions and preserve space for truly important exploration.

The Ecosystem Grows Larger When You Don’t Vertically Integrate

The AI era will create many business opportunities, but DeepSeek does not intend to do everything itself.

In the conversation, Liang Wenfeng used a very down-to-earth example: if you operate a power plant, you don’t necessarily have to build the generators and equipment yourself; as long as the price is reasonable, others can build them. The same is true for AI companies: they don’t necessarily need to swallow up the entire upstream, downstream, applications, and industry scenarios.

*“For example, if you operate a power plant, why must you build the generators yourself? As long as the price is reasonable, the power generation equipment can be built by someone else. We hope we don’t have to make chips ourselves, as long as we can buy them at a reasonable price.”*

For DeepSeek, AI is already big enough; there is no need to consume the entire value chain. A company should focus only on the most critical part and the part it is best at, while leaving other industries, applications, and scenarios to ecosystem partners.

*“The AI era will produce many trillion-dollar companies. Doing just a small part of it is already big enough for me; there is no need to do everything ourselves. We hope others will do the other things.”*

The more an AI company tries to capture all the value, the smaller the ecosystem becomes; the more willing it is to focus only on what it does best, the larger the ecosystem becomes.

*“We hope these AI technologies can be used in all kinds of production environments to improve social productivity. We have the motivation to do this, but whether we have the time, the people, and whether our colleagues themselves are interested is another matter. There is no conflict of interest.”*

Insights for the AI Industry and Application Teams

Closed-source and open-source models are not a matter of belief, but a system-level choice

Liang Wenfeng’s view on AI competition between China and the United States offers application developers an insight that is not about “betting on who will win,” but about rethinking model selection.

Closed-source and open-source models are not a matter of belief; they are a combined choice involving capability, cost, controllability, delivery speed, data security, and ecosystem space.

Leading closed-source models in the United States still have advantages: ample computing power, intensive research, mature user experience, and stable products. But if closed-source models remain expensive over the long term, with high invocation costs and limited deployment options, they will push some application innovation toward cheaper, more open, and more flexible model systems.

For developers, the future question should not simply be “which model is the strongest,” but rather:

• Does this task truly require the strongest model?
• Can the cost support high-frequency usage?
• Is the latency low enough to fit into real workflows?
• Are the data and deployment methods controllable?
• Are the ecosystem tools mature enough?
• Does the product have a fallback mechanism when errors occur?

The strongest model is not necessarily suitable for every task. Truly mature AI applications often place different models in the right positions: powerful models handle complex judgments, lower-cost models handle high-frequency tasks, and rules and caching handle stable workflows.

Small and Mid-Sized Developers Should Build on Top of Models and Solve the Problems Closest to Users

Liang Wenfeng has made it clear that DeepSeek does not need to do everything itself; many applications and industry scenarios should be completed by ecosystem partners.

This is good news for small and mid-sized developers. The future AI ecosystem will not belong only to foundation model companies. Small and mid-sized developers should focus on narrow, practical entry points rather than trying to replicate large model companies.

More realistic directions include:

• Vertical process automation: Instead of building a general-purpose chatbot, solve a specific process within a particular industry.
• Packaging model capabilities: Turn complex model capabilities into tools that users can use directly.
• Data connectors: Help enterprises connect internal data, documents, spreadsheets, and systems to AI.
• Evaluation and monitoring: Help enterprises assess AI output quality, cost, latency, and risk.
• Localization and deployment services: Provide private deployment, fine-tuning, and operations around open-source models.
• Industry knowledge layers: Transform industry rules, terminology, cases, and workflows into knowledge systems that AI can call upon.

Small and mid-sized developers do not need to compete with model companies on the main track. The smarter approach is to build on top of model capabilities and create the layer that is closer to users.

AI Applications Should Follow User Workflows, Not Model Hype

The biggest problem for many application teams today is that they treat AI as packaging rather than as a capability.

A truly valuable AI application is not one that simply adds another input box to the page, nor one that turns the original search box into a chat box. It is one that helps users take one fewer step, make one fewer judgment call, organize one less round of information, or wait through one less stage.

Users are not using AI for the sake of using AI. They want to complete tasks faster, make fewer mistakes, reduce costs, make better decisions, and achieve higher conversion.

So AI application teams should focus on several metrics in real-world scenarios:

• Whether the time users need to complete tasks is reduced;
• Whether the cost of manual review is reduced;
• Whether output quality is consistent;
• Whether failures are controllable;
• Whether it can integrate into existing systems;
• Whether it can continuously learn business context.

If these questions cannot be answered, the product is most likely just an AI wrapper.

AI Products Must Evolve from Tools into Work Systems

Many AI applications today are still at the tool stage: the user provides input, and the AI produces output.

But products with real long-term value will become work systems:

• They can understand the task context;
• They can call multiple tools;
• They can read business data;
• They can execute actions;
• They can record results;
• They can reuse experience in the next task.

For example, for marketing teams, AI is not just about writing a piece of copy; it can form a closed loop from data analysis, competitor monitoring, asset generation, and campaign recommendations to performance reviews and the next round of optimization.

For development teams, AI is not just about code completion; it can understand the codebase, reproduce issues, identify root causes, write tests, submit fixes, and record lessons learned.

This is the real opportunity at the application layer.

The Final Keyword Is Restraint

The keyword that runs most consistently through Liang Wenfeng’s remarks is restraint.

Restraint is not conservatism; it is a strategic choice. What makes DeepSeek most distinctive is not only what it has done, but also the many things it clearly could do, profit from, and scale, yet chooses not to do.

Do not build image features immediately just because a model can generate images; do not build Agent features immediately just because Agent is hot; do not rebuild a product immediately just because an open-source project has gone viral; do not put a requirement into the roadmap just because a customer asked for it.

There are many opportunities in the AI era, but the more opportunities there are, the more important it is to make trade-offs.

Directions truly worth pursuing usually meet three criteria:

• Users already have a clear pain point;
• AI can significantly reduce costs or improve efficiency;
• The product can accumulate data, workflows, or industry knowledge, becoming stronger with use.

If a direction does not meet these three criteria, then even if it looks hot, it may just be noise.

Summary

On the surface, this transcript of Liang Wenfeng’s investor meeting discusses DeepSeek’s open source strategy, commercialization, computing power, model efficiency, and organizational approach. At a deeper level, it is about how AI companies can stay focused on their main direction amid enormous opportunities.

For foundation model companies, beyond leading in capability, cost, openness, and room for ecosystem development will become increasingly important.

For managers of AI companies, management is not about filling up the team, but about helping the team reduce ineffective actions and preserve the space needed for truly important exploration.

For small and midsize developers, do not compete with foundation model companies on their main track. Instead, find your own focused entry point around the open source ecosystem, vertical workflows, data connectivity, and user experience.

For AI application teams, the model is only the beginning. Real product value comes from cost control, closed-loop task execution, continuous learning, and business outcomes.

Ultimately, what is most worth remembering from DeepSeek’s message is not any specific technical detail, but a way of judging: in a phase when AI opportunities are exploding, what is truly scarce is not the ability to act, but restraint; not doing more, but consistently doing the right things.