Since last night, WeChat public accounts have been frantically reposting the recording of DeepSeek's Liang Wenfeng, using AI to summarize and write articles.
Download the original recording transcript PDF
📎
https://xiangyangqiaomu.feishu.cn/wiki/Ojsuw8gE6ieBZJkYOMtcOcaenwq
DeepSeek has a model called DDCP. On the day its price was cut to a quarter of the original, the company group chat was filled with cheers.
In a normal company's financial meeting, a price cut means a drop in revenue and is bad news that leads to a scolding.
DeepSeek treated it as a moment worth celebrating.
In this three-hour and forty-four-minute exchange meeting, Liang Wenfeng repeatedly said one thing: They have no money, no cards, no fame, no top-tier talent in the team, and even he himself did not graduate from the best school.
How did a group of ordinary people create something in two years that made the whole world nervous?
Liang Wenfeng's counter-intuitive judgment: In the matter of AI, those who want to take the most will lose first.
The More You Take, the Faster You Die
Liang Wenfeng did a calculation. Suppose AI can eventually consume 10% of global GDP, and OpenAI wants to take 5% of that; this calculation is theoretically sound.
But as long as a competitor appears who is willing to take only 1%, they can crush the former with a lower price.
Then someone else will come along who only wants 0.1%, crushing the previous one as well.
This chain will continue to roll downward until it reaches an equilibrium point: if you take too little, the business model won't hold up and you'll be naturally eliminated; if you take too much, you'll be squeezed out layer by layer by those who take less.
Liang Wenfeng put it bluntly: He has no doubt that AGI will have huge commercial value, but rather than how to get a larger share, he cares more about how to increase the probability of making it happen.
And the way to increase that probability is precisely to actively shrink the portion he wants to take.
Applying this logic to pricing, DeepSeek's API is designed to break even in ten months, corresponding to roughly six times profit.
This is not profit maximization. If they really wanted to maximize profit, the price should be set higher because user demand in the current range is inelastic; a price increase would not significantly reduce consumption.
But the six-fold profit is actively locked at this upper limit for a very practical reason: By keeping the profit margin low enough, it becomes unprofitable for third parties to steal business by deploying open-source models themselves.
Open source and low profit are two sides of the same coin.
Instead of relying on closed source to lock in profits, they rely on compressing profits to a level that others cannot afford to replicate, thereby preserving the market itself.
Liang Wenfeng calculated that if they really wanted to earn a hundred times profit, open source would become a loophole, as third-party deployment costs might only be one-twentieth of theirs.
But DeepSeek never intended to take that path from the start.
He is not worried about open source because he never thought about earning a hundred times profit through vertical monopoly.
This logic of "intentionally earning less" explains almost every seemingly unprofitable decision this company makes.
Sesame Seeds in Front, Watermelons Behind
Last Chinese New Year, DeepSeek's consumer-end (C-end) user base suddenly exploded.
This was not in the plan. The team even considered not maintaining those users, but the users "couldn't be driven away" and eventually stayed.
Liang Wenfeng's explanation for this phenomenon is straightforward: Treating the things you can grab right in front of you as the top priority will cause you to lose bigger opportunities.
He calls this "picking up sesame seeds." Last year's C-end traffic and this year's business-end (B-end) revenue are, in his view, just sesame seeds—picking them up in passing is enough; it's not worth stopping for.
The real watermelon is AGI, and the opportunity for AGI is always large enough that there's no need to precisely calculate how much share one can occupy right now.
This leads to an interesting result: DeepSeek makes products almost by accident.
Liang Wenfeng said that the starting point for the company's training and API was never to serve users well, but because this step must be passed on the road to AGI. Users and revenue are just by-products generated along the way.
This strategy is, in his eyes, a form of dimensionality reduction strike: an organization with AGI as its goal puts far less thought into C-end and B-end products than companies that treat the product itself as the goal, yet the results are not necessarily worse.
His explanation is that a larger vision can better unite top talent; the resulting talent advantage will naturally overflow into the product level, becoming a dimensionality reduction strike.
Conversely, a company that treats "making a good product" as the end goal will have advantages in product experience, traffic, and user service, but in terms of upstream technical capabilities, it may never catch up with a company that treats technology itself as the end goal.
The same restraint also determines the things DeepSeek chooses not to do.
No Chip Making, No Full Industry Chain
An investor asked Liang Wenfeng if he would move upstream to make chips or downstream into vertical applications like finance or healthcare.
His answer was no.
The reason is a power plant metaphor: the person operating the power plant doesn't need to build the generator themselves. As long as they can buy it at a reasonable price, why build it?
This is an extension of the same logic of "restraint."
Liang Wenfeng said the AI pie is large enough that many trillion-dollar companies will emerge in the future. DeepSeek only needs to be one of them, eating only the part they are best at, without needing to swallow the entire industry chain.
Eating too much will instead lead to an early exit.
The same attitude applies to competitors.
Liang Wenfeng said DeepSeek is willing to help Alibaba, Zhipu, and Moonshot AI reproduce open-source models better because it doesn't harm their own interests.
It sounds polite, but the underlying algorithm is consistent: If the market is large enough, others doing well is not a threat; others doing poorly and ruining the trust of the entire ecosystem is the real loss.
Restraint explains what DeepSeek does not do.
As for what it will do, the answer is a technical ladder that has long been thought through.
A Ladder That Can Be Climbed Without Overtime
Liang Wenfeng describes the evolution of AI technology like climbing a ladder layer by layer, where every step is built on the previous one and no step is wasted.
Last year's step was the Chain of Thought (CoT), letting the model learn to think for itself, pushing the upper limit of intelligence a step further, already surpassing top humans in coding and Math Olympiad problems.
This year's step is Agents, pushing the boundaries of capability further through multi-step task execution.
But the Agent path will also reach its end.
After reaching the end, the model still cannot replace a real employee.
Liang Wenfeng gave a very apt example: When a new employee joins, they spend two months getting familiar with the environment. After that, if you say "call Xiao Wang over," they understand.
AI lacks these two months of accumulation; you have to explain who Xiao Wang is, where they are, and what their position is before it can execute.
This limitation of "must provide full context" is the invisible wall between current AI and true intelligence.
The next step to cross is continuous learning: Can the model, like a new employee, accumulate experience on its own through continuous contact with the environment, rather than relying on humans to feed context into its mouth every time?
Once continuous learning is solved, a stage Liang Wenfeng calls the "singularity" will appear: The model's capability will be strong enough to iterate the next version of the model itself, do its own research, and develop a more advanced next generation.
He specifically emphasized that this singularity is not a sudden mutation but a gradual, continuous process, though it's conventionally called a singularity.
Only after that comes embodied intelligence, where robots truly enter the physical world to do housework and elderly care.
As for why they follow this order, Liang Wenfeng said: This is the most labor-saving path.
By solving continuous learning first, AI can help accelerate every subsequent step of research and development, so humans don't have to carry everything themselves.
Conversely, if they did embodied intelligence first, that would be a bitter and tiring path built on human labor, which DeepSeek does not want to choose.
There is a detail here worth lingering on: DeepSeek's primary goal in training the next version of the model is not for external users to use it well, but for their own team to develop faster.
The model serves internal R&D efficiency first, and then incidentally benefits external users.
This priority order, to some extent, determines the sequence of almost all resource allocation in this company.
20,000 Cards and a Dissolving Moat
When talking about the gap between China and the US, Liang Wenfeng's judgment was very blunt: Talent is not the bottleneck; computing power is.
China does not lack AI talent; the base is large enough, and the distribution of smart people is random. There is no such thing as "all the smartest people went to the US."
The real bottleneck is resources.
The numbers are very direct: DeepSeek currently has about 20,000 H-series equivalent compute cards, most of which arrived only in the last month or two.
To train a model of the same grade as the current largest models (with about 800B activated parameters), one needs 50,000 GB300 cards, or 200,000 of Huawei's latest cards. This is just for the training itself, not including the massive experimental costs during the research phase.
DeepSeek's current resources can support experiments at the scale of dozens of billions of activated parameters, which is still an order of magnitude away from 800B.
Regarding domestic chips, Liang Wenfeng's judgment was unexpectedly optimistic.
He believes CUDA's ecological moat is being dismantled by three forces simultaneously: AI itself can help write code, significantly lowering the threshold for building an ecosystem; high-level languages like TileLang can quickly rewrite CUDA's entire operator system; gaming cards and compute cards originally shared an architecture because the AI compute market was smaller than gaming, but now the compute market is larger, and the two no longer need to be tied together—specialized chips are becoming mainstream.
He gave a specific time judgment: Within the next year, the fact that the domestic chip ecosystem is fine will be proven by facts.
The only remaining obstacle is production capacity.
The specific cost-performance comparison is also very direct: Huawei's 950 super-node can replace NVIDIA's GB200 and GB300 in terms of performance and price, at the cost of needing four cards to match one, while lagging two years behind. It doesn't matter if it's 50% to 200% more expensive, as long as it can be bought.
DeepSeek currently gets about 16,000 cards of capacity from Huawei. This volume is only equivalent to 4,000 NVIDIA cards, which isn't even enough to train the next generation of models, but it's enough to help Huawei get the ecosystem running first.
Liang Wenfeng calls this "helping Huawei do this well first." He doesn't expect to rely on these cards to train larger models; he is paving the way for the future.
Even more interesting is the catch-up narrative he provides: In the past, Chinese AI used one-twentieth of others' computing power to produce comparable things while lagging one to two years behind.
In the future, this "lag time" needs to be compressed from one or two years to three to six months, while the computing power gap remains just as large.
Not by relying on more resources, but by relying on smarter methods.
This "using clever strength to bridge the gap" approach is also reflected in how the company manages itself.
Half of Core Researchers are Labeling Data
Someone asked Liang Wenfeng if the hallucination problem of large models would affect user experience. His answer was very frank: Hallucination is not an unsolvable problem. It can be improved with better post-training; it's just that no one has put in a lot of effort to do it yet.
Inside DeepSeek, this is classified as a product issue; it will be solved, but it's not the focus.
A detail more worth remembering than the hallucination problem is: Half of DeepSeek's core researchers are labeling data.
This is not because labeling data is cheap in China. On the contrary, Liang Wenfeng said the cost of high-quality data labeling in China is no different from that in the US, especially for high-end data where there is no cost advantage at all.
So the current strategy is a two-pronged approach: label the low-cost ones first and set aside the high-cost ones for now.
He judges this to be a matter of time, not capital.
OpenAI and Anthropic started earlier and have been running for a long time. China only started investing seriously about six months ago, but it is expected that most of the gap can be closed within a year.
A Company Without Organization: What Keeps People Together?
Liang Wenfeng said DeepSeek has "no organization" and is driven by vision.
The specific operation: Company management is divided into two lines. One is top-down for collective projects, such as releasing V4 which requires full-staff collaboration and division of labor; this part is called "formal." The other is bottom-up, where everyone researches whatever they want; no one manages them, and there are no KPIs.
The standard Liang Wenfeng sets for himself is: Formal work should not take up more than half of an employee's time.
There are two considerations behind this.
First, research requires a relaxed environment. Pushing too hard makes it impossible to explore new things. Since the goal is to stimulate spontaneous interest, space must be left for "wild thinking."
Second, DeepSeek is focused enough that there are few things to do. The company's products are generally imperfect, and they haven't put effort into fixing them, which is itself an active choice.
This organizational style has no template to copy.
Liang Wenfeng said DeepSeek has not imitated any company, including Bell Labs which is often used as a comparison, because Bell Labs did not need to be responsible for its own profits and losses, whereas DeepSeek is ultimately a company that must consider how to survive. The government will not give them a single cent.
Every choice is a result calculated based on the actual situation at hand, not a copied answer.
Returning to That Cheer
Looking back now at the cheer on the day of the DDCP price cut: That was not an isolated story; it was a microcosm of this company's entire survival algorithm.
Restraint in not pursuing the maximum share is because, in a large enough market, those who are greedy will be squeezed out first.
Picking up sesame seeds without lingering is because staring at immediate gains will prevent you from seeing the larger watermelons behind.
Not making chips or vertical applications is because eating too much leads to an early exit.
Choosing the most labor-saving order for the technical route is because the saved energy can be used where the real bottlenecks are.
Having half the researchers label data is because the gap that computing power cannot fill can only be filled with a more solid foundation.
These decisions seem unrelated on the surface, but they are actually different ways of writing the same sentence: In a large enough track, not precisely calculating how much you can get right now is the prerequisite for getting more.
This is a judgment method that can be directly used by anyone in a competitive industry.
The next time you face a choice between "how much can I get now" and "how much can I get in the future," don't be in a hurry to calculate the immediate account.
Ask a more fundamental question: Will this matter, because you are taking too much now, prematurely attract a competitor who is willing to take less?
If the answer is yes, then no matter how tempting the short-term gain is, it's not worth trading the probability of survival for it.
