DeepSeek Founder’s 4-hour investor meeting: DeepSeek is prioritizing AGI over user growth and commercialisation

SC_OFF

A Chinese article compiled 52 remarks from Liang Wenfeng’s four-hour investor meeting. I’ve summarised the most important ones below.

DeepSeek has one central objective: AGI. This is not the time to maximize returns through products. Products are one rung on the path to AGI, but we do not need to devote too much thought or energy to building consumer or enterprise products.

We have always been commercializing, but commercialization is not our objective. The point at which DeepSeek fully pivots toward commercialization is probably still very far away.

Restraint is a strategy: you give up certain things in exchange for more of something else. Open source is a form of giving up value. Internally, it gives employees a sense of accomplishment and strengthens organizational cohesion. It also benefits society. Other companies and ordinary people are happy about it.

I have no doubt that AGI will have enormous commercial value. Given that, my priority is not to capture a larger share of the value, but to increase our probability of succeeding.

Open source is beneficial if you want to make AI commercially successful. That may sound counterintuitive. Historically, a software company’s entire market might have been worth only a few billion dollars a year, so open-sourcing the software meant giving that market away. But AI is large enough that it may ultimately account for 10 percent of global GDP. If we try to monopolize that value, history will inevitably leave us behind. That is an objective law. It is a historical perspective.

The models we release as open source are the same models we deploy ourselves. We will not open-source an inferior model while privately deploying a better one.

The gap between Chinese and American AI is primarily a gap in resources. We believe in scaling: larger scale undoubtedly produces better results. We do not train models of this size because we believe this size is sufficient. We train them at this size because these are all the resources we have.

Anthropic’s current lead over OpenAI is temporary, not permanent. OpenAI and Google will most likely take turns pulling ahead in the future.

We do not want to build the next super-app. Become the next ByteDance? The next Tencent? We have absolutely no such ambition.

There is only one thing on which we cannot compromise: we must maintain the stability of the team. This is also one of the greatest risks we face. Of course, that risk has been substantially reduced by this financing round.

AGI offers the greatest return. As for everything else, we will do it if we have the capacity, and we will not do it if we do not. Restraint is part of our vision.

Full Article, translated to English

Full Transcript of Liang Wenfeng’s Four-Hour Investor Meeting
Original by elsewhere

July 22, 2026, 11:33 p.m. · Beijing ·

elsewhere
@elsewhere

Last month, elsewhere reported on DeepSeek’s fundraising story. The part that drew the most discussion was undoubtedly the rumored four-hour investor meeting.

Over the past month, various remarks attributed to Liang Wenfeng have circulated widely. We have also gathered some of what was reportedly said at the meeting from multiple sources.

During the meeting, Liang repeatedly said “no”: DeepSeek does not see itself as a company of geniuses; it does not seek excessive profits; it does not pursue user growth for its own sake; it will not become closed-source; it will not work on 3D generation, video generation, or world models; and it does not intend to build the next super-app. In his words, restraint is a strategy—one that improves the odds of achieving AGI.

Among the limited materials available to us, several terms appeared frequently: models, cost, AGI, time, open source, and so on.

Most of the time, Liang spoke cautiously and in plain, unadorned language. Only when discussing a handful of issues he cared deeply about did he reveal a sharper edge:

“As long as I can keep the team stable, I will be able to achieve AGI. It is that simple.”

Below are 52 remarks we collected. Some wording may differ slightly from the original, though we have preserved the intended meaning.

DeepSeek Has Only One Main Objective

This is not the time to maximize returns through products. Products are one rung on the path to AGI, but we do not need to devote too much thought or energy to building consumer or enterprise products. When you occupy a technological high ground and then apply it to lower-level technology, you have an overwhelming advantage. Products are a by-product of the journey toward AGI.

Many things do not belong on our main path—for example, 3D generation and video generation. The same is true of world models, which do not have much bearing on the upper limit of intelligence.

Multimodality is very important for products and for consumer users. But it is only a component. It is neither the main objective nor intelligence itself.

There are, of course, ways to address hallucinations in large models, but it is a long-term problem. Internally, we categorize hallucination as a product issue. We will work on it, but it is not our central priority.

At this stage, the most important thing is still coding agents. Given the situation in China, the most sensible approach is probably to focus fully on general-purpose agents. Agents for finance, healthcare, and other verticals should have lower priority.

If the AI era produces many trillion-dollar companies, it would be good enough for DeepSeek to be one of them.

First Continual Learning, Then AI Self-Iteration, and Ultimately Embodied Intelligence

AI today does not lack taste or intuition. What it lacks is the ability to learn continuously.

Humans can keep learning over time, whereas with AI, you have to provide all the relevant context again for the same task. That is almost impossible, which is why AI still cannot replace an employee. The next generation of models must be capable of continual learning before they can truly be called next-generation models.

We hope our next model will help us with our own development work. Put simply, the primary goal of the models we build is not for everyone else to find them useful, but for us to find them useful ourselves. That is the fastest path to AGI.

No one in the world has yet found a good solution, because “learning” consists of many different things.

DeepSeek’s long-term vision is AGI. If the route toward it is like climbing a staircase, last year’s step was chain-of-thought reasoning. This year’s step is agents. After agents, the next problem to solve is continual learning.

Once continual learning is achieved, we may reach a gradual singularity: models could perform everything humans can do, including developing more advanced AI models themselves. In other words, AI could accelerate AI research. Only after completing that step do we arrive at embodied intelligence.

The ultimate form of intelligence may be embodied. For an ordinary person, what they need is not a computer; they need labor.

A Full Shift Toward Commercialization Is Still a Long Way Off

We only seek a reasonable profit. We do not price our services to maximize profit.

With one of our models, we initially worried that demand would be too high, so we priced it relatively expensively. Later, when we cut the price to one-quarter of the original level, many people in the company chat celebrated. That was the whole point of putting so much care into making the model good: enabling everyone to use it as fully as possible.

Low cost is an outcome. We have continuously designed our model architectures to reduce cost. We also want the cost to be affordable, especially in an environment where compute is scarce.

There is another reason: the lower the cost, the larger the model you can support. When compute is limited, greater computational efficiency allows you to train larger models. Large companies can solve the problem simply by adding more resources. We prioritize cost efficiency.

From t…