The Smartest Moves Look Like Idealism: A Hard-Nosed Breakdown of DeepSeek's Liang Wenfeng's 4-Hour Meeting
Recently, a leaked 4-hour internal meeting recording of DeepSeek founder Liang Wenfeng, spanning 42 pages of transcript, has sent shivers down Silicon Valley's spine.
The internet is busy canonizing him, calling him 'Saint Liang' for open-sourcing for humanity. But after reading the transcript, I can only sneer: the smartest moves often look like idealism!
The Terrifying Calculation of 10-Month ROI: The Nongfu Spring of AI
Look at the big model labs in Silicon Valley—how many dare admit they're profitable? Saint Liang seems to say: 'Bro, I'm not targeting you, but everyone in this room is trash!'
The internet is hailing Liang as a 'cyber saint,' but he's actually an ice-cold 'quantitative actuary.'
Many don't get it: after raising tens of billions, why does he dare go all-in on buying GPUs even when prices are inflated two or even four times?
Because the odds of this bet are 100%—the purchased equipment pays back its full cost within 10 months through API calls alone. In the standard 5-year depreciation cycle for servers, that means for the remaining 50 months, the equipment is self-'printing money' at zero cost and high efficiency.
His most ruthless move is curbing the urge for huge profits by locking API prices at 6 times cost. Think 6x is high? No, it's the ultimate 'price defense'—it squeezes margins so thin that any third party trying to buy GPUs and self-host finds it 'completely unprofitable.' This meticulously calculated cost pricing directly eliminates arbitrage opportunities at the root, forcing closed-source giants into a dead end.
Many blindly praise DeepSeek's open source. Open source? Even if I hand you the source code as-is, can you run it? You don't have the GPUs, the data centers, or his insane algorithms and ops that stretch every penny. Your cost to run it yourself would be ten times higher than just calling his API!
In the end, you'll have to beg him: 'Saint Liang, help me buy GPUs; I just need to break even in a year.'
He's not doing tech open source; he's using supply chain management as a dimension-reduction strike to monopolize pricing power at the base. While everyone else loses money for attention, he's already the water seller who never loses.
The 'Restraint' of Dimension-Reduction: Even Zhang Xiaolong Would Call Him Big Brother
The internet is analyzing Liang's 'restraint.' Some compare him to WeChat's Zhang Xiaolong, but that's underselling it!
Zhang's restraint is at the product level. WeChat has massive traffic; one extra step gets you shut down, one less step gets you criticized—it's restraint or die.
But what is Saint Liang's restraint? It's ecosystem-level, god's-eye-view, high-dimensional restraint!
Why doesn't he chase huge profits? Because he sees the sea of stars that is 10% of human GDP! Would you be stupid enough to price Nongfu Spring at ten times the cost? Pointless! He's not just calculated how the product survives; he's calculated how investments, upstream and downstream ecosystems survive.
His micro-moves in the capital market turn restraint into a 'pig-slaughter game.'
Right after the last funding round, he starts the next one. Almost no gap, and the valuation jumps another 30% to 40%!
How? He puts in 20 billion of his own money as a buffer, brings in Tencent and the state-owned assets committee to stabilize the base. A bunch of institutions are lining up with checks, but Alibaba isn't even in the room—he won't let them in.
He closes the door, lets those who missed out squirm like ants on a hot pan, tortured by FOMO, then cracks the door open: 'Oh, we were about to close, but seeing how enthusiastic you all are, I'll extend for 15 minutes and add a few slots—but the price goes up 40%.'
And those institutions still thank him and hand over the money!
Bypassing Nvidia's Ecosystem? AI Writes Its Own Code!
Business savvy is just the appetizer; the foresight in tech path is what's truly chilling.
The whole world is kissing Jensen Huang's ring, thinking Nvidia's CUDA is an insurmountable moat. But Liang has long been positioning TileLang to bypass Nvidia and adapt to domestic chips.
Sounds like a fairy tale? Can AI really bypass low-level hardware code?
I'm telling you, it's not bullshit. Recently, something big happened: AMD released a new chip and planned to send a huge team for low-level optimization to adapt Claude. But Claude said: 'No need, just feed the context to the model and let Claude write it!'
In the end, Claude, with just one or two top engineers, nailed the early prototype!
It's like code that would've taken 100 engineers half their lives to write, now done in a flash by a large model!
The core ecosystem moat is being eroded from within by large models. How long do you think Nvidia's moat can hold? Once large models evolve enough, they'll flatten the low-level differences across hardware. Then, compute hegemony will be overturned!
Embodied Intelligence Is Stupid! Let the Model Evolve First!
In the 4-hour meeting, what impressed me most was Liang's cold-blooded planning of the AI path: Large Model -> Chain of Thought (CoT) -> Agent -> Self-Iteration -> Embodied Intelligence.
Many companies are hyping embodied intelligence. What does Liang think? The subtext is: doing embodied intelligence now is just stupid!
Why? In the virtual software world, there's no problem that can't be solved; but the physical world is full of physical walls! To make a robot peel an egg in complex reality, there are countless unpredictable accidents. Using current models to brute-force edge cases has an abysmal ROI. That's like 'chasing the infinite with the finite—doomed!' In short, it's stupid!
What's the smartest move? Still restraint.
Wait for what? Wait for the intelligence singularity. Wait for large models to cross the threshold, achieve continuous learning and self-iteration like humans, and realize a certain level of AGI in the software world first. It's like a mature Alipay that's learned to pay off its own Huabei!
Once large models can self-iterate, the evolved intelligence can then adapt to physical world problems, and the difficulty drops dramatically.
In contrast, many companies, without even getting the 'brain' right, spend fortunes on fancy dancing robots. Besides making money from rental shows, I don't see what robots can do now.
Isn't Tesla's end-to-end autonomous driving the same logic? You think the model in the car is as smart as the cloud? It doesn't need to be! As long as the on-device 'hands and feet' coordinate with the cloud 'brain,' it's enough to beat human drivers.
Conclusion
So, stop being fooled by those ethereal sentiments. Liang did talk about the grand vision of occupying 10% of human GDP. Why must he? Just like 20 years ago when Google DeepMind's boss said they'd achieve AGI to solve everything. You have to paint that big picture to awe the common folk and capital sharks—that's called grand narrative.
But behind the sea of stars is a cold-blooded calculation that optimizes supply chain costs to the extreme and plans tech paths to a terrifying degree.
In this world, pure idealists died out long ago. Those who change the world are often more calculating than the greediest capitalists.
The smartest moves often look like idealism!