In May 2026, DeepSeek completed the first external fundraising round: – RMB50 billion fully subscribed, with Founder Liang Wenfeng himself contributing 40% of it.
The online meeting with investors lasted four hours, as we mentioned in an earlier article. Last night, a PDF document claiming to be the transcript of the meeting started circulating on the Chinese Internet.
In the transcript – Liang discussed DeepSeek’s open-source and pricing logic, its technology roadmap to AGI, the organisation and the compute gap with the U.S. He also shared a bit about Huawei chips as well as his views on how the industry will evolve towards.
MW has read the transcript that highlights the thinking and culture of the company. It is more sensitive in some areas, and we are publishing selected Q&A which captures the core questions from the investment community:
Q&A Session
Q1: Can you share more on the timeline for when continual learning will bring a breakthrough – including what architectural and algorithmic innovations, and what other key elements are still needed to achieve it?
A1: Liang Wenfeng – Founder
There are few people on it; it needs research, and the whole world is researching this problem right now. For investors, what you see most of is agents; but for those of us doing the research, what we see more of is learning, and how to solve it. Learning may not be a single technology – it is a problem, and there may be many technologies to solve it, not one thing but many things. AGI is made up of many things: you need a model, and you need many other things too. It is both an engineering and an algorithmic problem. It’s fairly technical, but there are many approaches and a lot of research going on.
Q2: Two questions. First, on the ecosystem: now that you’ve open-sourced, how many partners, teams and people in the industry can reproduce your models and results well? Where do you need more high-quality talent to connect to your models, and will the future be a “model matrix” – you make the base model better and better, and partners build vertical-industry or application models on top? How do you see this ecosystem growing over the next two or three years? Second, on hardware: global AI leaders may each invest at the US$100-billion scale, while China looks to have shortcomings in compute. How long until that is solved so it doesn’t hold AGI back – is it something China is bound to achieve, just a matter of time and money?
A2: Liang Wenfeng – Founder
On the ecosystem: the problem every enterprise faces right now may be a shortage of talent, but I think that shortage is only a phase. In the early days of any industry, talent is always short – it was the same when people first built websites, and again when the internet needed server-side engineers. But that kind of shortage is resolved very fast, in two or three years, because large numbers of people get trained up. The AI talent shortage is also just a phase, and we’ve already seen it greatly eased. Training people is fast, so across the industry – ecosystem, model companies, whatever – talent isn’t scarce. A shortage is a short-term phenomenon; there has never been a long-term shortage of any one type of person. I remember over a decade ago people said pilots were scarce, and pilots take a long time to train, but that too was resolved quickly.
Also, there are a bit too many companies building models domestically. The US has maybe three; China has far too many building foundation models. In the end there won’t be a need for so many, so it will surely converge – but that takes a process. Resources are spread thin, which is somewhat wasteful: every player does the same thing, whereas in the US only three do it and the resources concentrate. Right now everyone thinks the margins are very high, so they all want to do it themselves; when they find it isn’t that profitable, they may stop. It certainly isn’t that profitable lately – I don’t believe it is, because that wouldn’t fit the objective laws. A very high profit margin would be abnormal; it should be a reasonable profit. I absolutely don’t believe large-model companies can take most of the profit – that’s impossible, because with so many of them the gaps needn’t be large, and the gap is only two things: time and cost. So there won’t be windfall profits: those who control costs well earn a bit more, those who don’t earn a bit less. That’s all. In the end China will have three or four in real competition, and the prices are already low enough for a price war.
Q3: On data – public data is now accessible to model companies, so that’s just a matter of time and cost. But when we truly reach AGI, the ideal is a model that iterates and trains itself. For that, is simulated/synthetic data usable, or is real data always the highest quality? If it still needs real data, does that cap AI’s intelligence at humanity’s past – or can it break that ceiling through simulated, synthetic or created data and surpass all prior human knowledge?
A3: Liang Wenfeng – Founder
I think it can surpass it. Take Go: *AlphaGo played a move no human had ever seen. So it can certainly surpass humans within a certain range. It may also have a ceiling and limitations, but we can’t see them now. Broadly, we believe it can go beyond the knowledge humans already have and can articulate. As for whether that runs on real or synthetic data – there will surely be many methods.
MW Note: *The reference is to move 37 in AlphaGo’s 2016 match against Lee Sedol, a play so unconventional that commentators initially thought it a mistake.
Q4: The core problem for this stage is continual learning, which abroad is called recursive improvement. First, what is the biggest technical difficulty today, and when do you think it can be solved? Second, you said you’d solve continual learning first and then move to general intelligence – what’s the technical reason for that order? Does it mean that once continual learning is solved, DeepSeek will also build general intelligence?
A4: Liang Wenfeng – Founder
The difficulty is that we haven’t yet found a method that really works – nobody in the world has, everyone is still groping for it. We have many ideas that look promising, but none has been made to work yet.
On the second point, internally we now put weight on this narrative: the next model we train should help our own development – it should raise DeepSeek’s own efficiency and help us build the version after it faster. Put simply, the first goal of the models we build is not that everyone finds them useful, but that we find them useful. Once they’re useful to us, we develop the next version faster, and that’s the fastest route to AGI. If it’s good for us, it’s probably good for others too – but first it has to be good for us. It’s an odd narrative, but many of us really think this way. We are now quite certain that we need AI to help us build AGI; it isn’t working autonomously yet, only in tandem with people, but it’s already very useful.
And the reason to solve continual learning first is that it greatly speeds up our R&D. Once it’s solved, general intelligence becomes easy – with AI able to learn continually, its capability is very strong, and today’s agents are limited precisely because they can’t. Solve continual learning first and general intelligence follows easily. Otherwise, building general intelligence by hand right now is tiring and bitter – data-intensive, labour-intensive, poor value for effort.
Q5: You often say AGI arrives gradually rather than through a sudden jump. Can I take it, then, that it’s a process with no critical point?
A5: Liang Wenfeng – Founder
It has no critical point, but it is non-linear. One narrative we fairly believe is that AI can accelerate AI research – because you can use AI to speed up your own research, so further down the line it may become non-linear. And on scaling: I currently see no ceiling to the scaling of language models. Neither our present level of intelligence, nor America’s, shows any ceiling.
Q6: I’m curious – for the US, an 800-billion-activated model can be trained but is hard to actually serve, because it’s so expensive. Humans acquired language only in the last hundred thousand years or so, after billions of years of evolution. In training AI the order can be reversed, but ultimately, after this ceiling, don’t you still have to enter something like a physical or world model – the embodied part?
A6: Liang Wenfeng – Founder
Yes, I think we will certainly move into embodiment in the end – for our company, naturally, the endpoint may be embodiment. Because an ordinary person doesn’t need a computer; in their daily life – food, clothing, shelter, travel and leisure – what they need is embodied intelligence to meet real, physical needs. If the goal is to meet human labour needs, embodiment is unavoidable. Put another way: suppose there were no embodiment – then what we hope AGI can do is iterate the next version of the model. And once we have embodiment, we hope it does the same there: iterate the next version of embodiment, build the next robot.
Q7: From DeepSeek’s earlier interviews, it seems taste and intuition matter a lot in choosing important directions and research – not just engineering optimisation. If AI can self-evolve later, will taste, intuition and the like still matter, or what will?
A7: Liang Wenfeng – Founder
AI doesn’t lack taste and intuition today; what it lacks is the ability to learn continually. Its taste and intuition are fine – ask it to write an essay, and its taste and intuition are perfectly fine.
Q8: Continual learning is still an unsolved research problem, while a coding agent catching up to the *MILES/office level is a fairly certain target. For an unsolved research target and a fairly certain scaling target, how should you allocate research resources – especially researchers – to strike the best balance?
A8: Liang Wenfeng – Founder
The “office” kind of target is fairly certain; the *”MIS” kind I wouldn’t yet call certain – it’s only a goal. That said, this sort of exploratory work doesn’t consume resources. CoT-style free research doesn’t burn cards – it needs very few, and what it needs is ideas. It doesn’t consume talent either, because it isn’t a project that needs someone sitting on it full-time; it needs many people all thinking about the problem. So there’s nothing to allocate. We call it “having a lucky dip” – the barrier is very low, anyone can reach in, but who pulls out what, I honestly can’t say is a matter of talent or something else. What sets us apart from other companies is simply that we spend time discussing it and treat it as important, even though it needs few resources.
On hallucination: it does hurt the user experience, and there’s a way to solve it, though it’s a long-term proposition – solvable through better post-training. People just haven’t put much effort in. For me it’s a problem, but we tend to treat it as a product problem; we’ll fix it, but it isn’t a priority.
On data labelling: this is tied to our capital structure. With our level of capital investment, we can’t support the cost of so much high-quality annotation. The cost of annotation is no different in the US than in China – China has no cost advantage, especially for high-end data – so it’s hard for us to invest in it the way the US does. Labelling data is simply too expensive, whether we outsource it or do it ourselves. So we walk on two legs: we label the cheaper data first. You could say half our company is now labelling data – half of our core researchers, our most important people, are on it. At this stage, solving the AI problem comes down to labelling data.
MW Notes: “MILES/MIS” in the exchange meeting refers to DeepSeek’s proprietary reinforcement learning training framework, Miles.
Q9: You said Chinese models are certainly stronger than US ones on efficiency, and that we may be stronger in other areas too. Which areas – intelligence, or others?
A9: Liang Wenfeng – Founder
On many aspects of experience, we may be able to do better than the US – I don’t think we’ll necessarily be worse on user experience, and on product capability we won’t necessarily be worse either. Our costs should also be lower. As for a broader structural advantage, there may not be one – but on cost and product I think there genuinely is. Cost is easy to understand: because they don’t have to do it, they don’t develop the capability, and they certainly don’t take it as seriously as we do. Product is similar. So on those two fronts we may have a structural advantage.
Q10: You mentioned post-training is relatively costly, and Anthropic and OpenAI invest huge sums. After this fundraise, will you increase your investment in post-training?
A10: Liang Wenfeng – Founder
The gap is mainly in high-quality data annotation, and mainly within AI research. We’ll certainly increase investment, but high-quality annotation isn’t typically a matter of capital – the bottleneck is time. OpenAI, Anthropic and others abroad started earlier, with more capital and more cards, whereas domestically we’ve really only been at it for the last half-year, so we need more time. It has little to do with capital investment: even without more money, the existing capital is enough to expand at the fastest pace, and that pace has a ceiling that isn’t set by money or cards. It’s a rapid-expansion process, and I think within a year the high-quality-data problem will be handled fairly well domestically. It does, though, take time.
Q11: We’ve seen Anthropic use its own model to build vertical products – finance, law, and heading towards healthcare. Will we consider such vertical applications at some stage?
A11: Liang Wenfeng – Founder
I haven’t thought through what our domestic business model will be, or the smoothest path – we’re not at that stage, and conditions at home and abroad aren’t necessarily the same. As things stand, the most sensible approach is to go all-in on a general-purpose agent; other agents – finance, doctor and so on – should be lower priority. We should do coding first, because a coding agent can do a great deal, and there are many vertical agents besides. At this stage, what matters most is still the coding agent.
Q12: We admire that you’ve built DeepSeek in a very pure, research-driven way, but the industry has now reached the capital markets. You’re responsible both to your team and to investors, and you’ll have public shareholders ahead. How will you balance pure AGI research against the capital markets?
A12: Liang Wenfeng – Founder
I now think we can do both – we want it all. If I can earn a few hundred million US dollars of B2B revenue this year, plus our consumer users, that alone is a certain commercial base. With B2B revenue next year, and if demand grows further, the company won’t be far from net profit – it may already be there, out of the pure cash-burning stage – so the moves we can make afterwards should be fairly large. In the worst case, selling APIs could support a listed company: if technical progress stopped and we froze where we are, we’d go all-in on selling APIs and doing those services well, and I think that would be enough. So I’m confident; it really isn’t that hard, because we’re in a high-leverage spot in a fast-moving field. We hope for a bigger dream, but we also have a fallback set of results we can put on the table.
Q13: DeepSeek’s biggest difference from other companies is its organisational form. Is there a good model for it to learn from – historically, something like *Bell Labs – or do you have to explore it yourselves?
A13: Liang Wenfeng – Founder
First, we have no model to imitate. Every step proceeds from our actual circumstances – seeking truth from facts and deciding according to reality. So it’s a product of the times, a response to real conditions, not the result of imitation; we chose each step by analysing the pros and cons, not by copying someone. I think we’re unlike Bell Labs, because Bell Labs explicitly didn’t need to commercialise, whereas we explicitly do. In the end we have to survive; we are, after all, a company – the government won’t give me a single cent. So we can have a very lofty mission, but at bottom we are a company and have to think about staying alive. That’s why the B2B matters to us – we may have to live off it later – even if it’s just a cost line now. Historically, many great companies have had a pursuit beyond profit, and that pursuit, far from hurting their commercialisation, made it better. At bottom we’re still a company; we just make trade-offs about which money to earn, when, how much, and how.
MW Note: Bell Labs was AT&T’s research laboratory, credited with foundational inventions from the transistor to information theory.
[THE END]
Disclaimer:
This translated transcript was compiled by Momentum Works from a circulated, unofficial Chinese-language recording of DeepSeek founder Liang Wenfeng’s investor meeting. It is intended for informational and analytical purposes only. The recording has not been confirmed by the speaker, and the translation may contain unintentional errors or omissions.
All original spoken content belongs to the respective speakers. Please refer to any official announcements or statements for authoritative reference.
Any analysis or commentary by Momentum Works is independent and does not represent the views of DeepSeek or any other organisations mentioned.
