Leung Man-Fung Investor Seminar - Audio Transcript (PDF file)

At the investor conference, Liang Wenfeng systematically explained DeepSeek's underlying logic: the company takes AGI as its sole objective, driving management through vision rather than KPIs; and adheres to...open sourceThe concept of "recovering investment in ten months" and the belief that restraint leads to success; depicting the process from...AgentFrom continuous learning, to self-iterative singularity and embodimentintelligentThe technology roadmap; admitting that the gap between China and the US lies only in computing power resources rather than talent, and that they are making every effort to expand card availability and are optimistic about domestic chip substitution; emphasizing that team stability is the only core interest, and that the C-end and B-end are merely byproducts on the road to AGI.

freeGet theLeung Man-Fung Investor Conference - Audio TranscriptThe original PDF report is available by scanning the QR code and replying with the following code: 20260723

• timeRecorded on May 20, 2026, approximately 3 hours and 44 minutes.AIEdited and compiled into a draft.
• SpeakerLiang Wenfeng (Founder of DeepSeek)
• Audience:investor
• natureInternal communication meeting, containing sensitive information (card volume, finance, etc.), requires not to be disclosed.

Company Vision and Core Values

DeepSeek was founded with the intention of "having a great deal of goodwill towards the world," believing thatAIIt's about things that are useful to humanity, things beyond money. The company's management isn't based on rules and KPIs, but on a vision-driven approach. This vision has never even been written down, but it permeates everyone's way of doing things and their attitude towards the world. Everyone's understanding of the vision may differ, but the general direction is the same.

Liang Wenfeng's most admired manager is former GE CEO Jack Welch, who believes he was right about one thing: managing a large company relies not on rules and regulations, but on vision. Vision is not a slogan on the wall, but how you do it, not how you say it.

open sourceThis is DeepSeek's core strategy for two reasons:

• First, the vision requirementsWithout this vision, it's impossible to organize people. (This is also related to Zhipu.)open sourceBut there's a sense of being forced; for DeepSeek,open sourceThat's the original intention.
• Second, business judgment。AIUltimately, it will account for 10% of the GDP of human society. The market size is too large for any single company to monopolize. The more a company tries to monopolize, the greater the resistance it will face, and the easier it will be to be abandoned by history.

open sourceThis doesn't conflict with commercial viability, provided that only a reasonable profit is made—pricing based on a ten-month payback period for equipment costs (approximately 6 times the profit). At this profit level, third parties cannot achieve the same deployment costs.open sourceIt won't affect income. He even stated that the strongest model will...open source"I can't see any benefits from closed-source systems."

• Price restraintThe goal is to break even in ten months; maximizing profits isn't the primary objective. Once, when the model was first launched, the price was too high, which upset the team. Later, it was reduced to a quarter of the original price, which caused a stir in the group chat. This is because we invested so much effort in developing the model so that everyone could afford it.
• User restraintLast year, during the Spring Festival, the number of C-end users surged, but we didn't capitalize on this opportunity to monetize, grab traffic, or become a super app. We didn't have the ambition to "become the next ByteDance or the next Tencent." The opportunities in AGI are too great; what's ahead of us are just trivial matters.
• Business restraint3D, video generation, world models, etc. are not included.intelligentWe won't do it on the main storyline. Not because it's worthless, but because it's not a key bottleneck currently limiting AGI.MultimodalI can do it, but only as a component, not as a special function.intelligentitself.
• Competition and restraintThey don't antagonize any large or small internet company, and are even willing to help competitors. Because...open sourceThere are no boundaries to begin with, and I hope to help everyone reproduce it.

AIDevelopment is a step-by-step evolution, with each step building upon the previous one, and no step is wasted.

• Language models are a fundamental capabilityThe CoT mindset chain is a ladder I climbed last year, improved through self-reflection.intelligentUpper limit.AgentThis is the stepping stone we are taking this year, with a wider range of capabilities.AgentIt will reach its capability limit, but "still cannot replace your employees" because it cannot continuously learn.
• The next bottleneckContinuous learning.AISpend two months familiarizing yourself with the environment and understanding the context, just like a new employee. A truly effective solution has yet to be found worldwide.
• After solving the problem of continuous learning, we arrive at the singularity of self-iteration.The model can develop its own next version and research more advanced technologies on its own.artificialintelligentThis is a gradual process, not a sudden change.
• BodyintelligentThe model enters the real world, doing housework and caring for the elderly. At this stage, no human intervention is needed; the model can emerge on its own.

• Two lines in parallelThe top-down approach involves formal projects requiring division of labor and collaboration, such as developing the V4 model, and takes no more than half the time. The bottom-up approach is free exploration, with no KPIs or supervision, and takes at least half the time, allowing exploration in directions that one deems important.
• No overtimeFirst, research requires a relaxed environment; too much pressure makes research impossible. Second, the company is very focused and restrained; there aren't many things to do. Many products are imperfect, but there's no effort to fix them.
• There is only one core interest.Team stability is the only non-negotiable factor; everything else can be mitigated. As long as the core personnel stay, AGI will definitely be achieved; everything else is just a matter of time, at most six months to a year later. Money and resources are not issues.

The gap lies primarily in computing power resources, not in talent. There's almost no difference in talent; it's essentially the same group of people, but China has a large talent pool and recruits many more each year. Currently, China's computing power is approximately 20,000 H-frames equivalent, far lower than that of the United States. The most...Large ModelActivation speed is only a few tens of kilobytes (B), while the US has reached 800 B, a difference of an order of magnitude. Training is equivalent.Large ModelWe need 50,000 GB300 cards or 200,000 Huawei cards.

The time lag is approximately 12 to 18 months. The strategy is to "buy as many cards as you can," ideally spending all the funds within six months. Buying cards is more profitable than putting money in a bank; the social cost can be recovered in ten months.

Domestic chipsHuawei's 950 processor (four chips) is equivalent to one Nvidia GB300, but it's two years behind. However, there won't be a gap in the ecosystem in the future because of the TileLang advanced compiler +AIWriting code is possiblefastRebuilding the ecosystem. V3 training is now almost entirely independent of the NVIDIA ecosystem. The only bottleneck is insufficient production capacity; Huawei can only provide approximately 16,000 units.

Both C-end and B-end are byproducts of AGI (Automatic Generative Technology), not the goal itself.

• C-endUsers are "impossible to get rid of," so it's low-cost to maintain them. They don't know what use they'll be in the future, so they just grab whatever's readily available.
• B-endWith optimistic growth prospects, it's possible to achieve an ARR of several hundred million US dollars. If...AIIf the company can reach a billion US dollars, its cash flow will turn positive and cover R&D expenses.

The final difference lies in three aspects: cost first, time second, and user experience third. There are no other fundamental differences. There are "too many" companies doing basic modeling in China, compared to perhaps only three in the US. Resources are scattered, leading to some waste. Eventually, the number will converge to three or four, resulting in sufficient competition. There won't be exorbitant profits; only reasonable profits will be made. Because if profit margins are extremely high, "it certainly doesn't conform to objective laws." China is very likely to be the global leader in this field.AI"One of the Three-Body Problems," its low systemic cost is a structural advantage, "Made in China."AIThe price might be cheaper.

freeGet theLiang Wenfeng Investor Conference - TranscriptThe original PDF report is available by scanning the QR code and replying with the following code: 20260723