In an industry defined by founders who chase billion-dollar valuations, curate loud public profiles, and promise world-changing breakthroughs at every press conference, Liang Wenfeng is refreshingly invisible.
The CEO of Hangzhou-based DeepSeek has rarely been seen in public — only a couple of photographs circulate online, his face unfamiliar even to most tech journalists. Yet behind that low profile sits one of China's most intriguing and formidable AI contenders, backed by a strategy so radically different it's reshaping assumptions about what Chinese AI can achieve on a fraction of the budget.
The Leak That Changed Everything
Last week, a leaked transcript of a nearly four-hour closed-door investor meeting held in May went viral on mainland Chinese social media. What followed was arguably the clearest glimpse yet into the mind driving DeepSeek -- a philosophy rooted in corporate restraint, low-margin pricing, and open-source commitment.
Liang Wenfeng didn't talk about moonshots or timelines. He didn't brag about parameter counts or market share. Instead, he explained a company vision that deliberately rejects what almost every other AI lab is doing.
"Restraint Is a Strategy"
This is the phrase that encapsulates DeepSeek's entire approach. In the leaked transcript, Liang articulated:
"Restraint is a strategy. It lies in the fact that sometimes you can give things up in exchange for more of something else. Not open-sourcing works the same way--you could see it as pressure on us, or you could see it as us giving up margin."
In an era where OpenAI, Anthropic, and Google lock their most powerful models behind paid APIs where Microsoft and Meta chase short-term commercialization, DeepSeek has done the opposite: they commit to keeping their most advanced models open-source.
This isn't altruism. It's calculated. The company believes open weights create developer gravity that turns a model into an ecosystem -- a moat that's harder to build with closed proprietary systems.
The Ten-Month Payback Rule
DeepSeek's pricing strategy would make Silicon Valley investors scream. Liang explained that DeepSeek prices based on a "ten-month payback period" for equipment costs -- approximately six times the profit. Once costs are recovered in ten months, further pricing doesn't increase revenue significantly because third parties cannot match those deployment costs, whether open or closed.
When the model first launched, the price was too high -- it upset the team. Later, it was reduced to a quarter of the original price, causing stir in group chats. Why? Because the company invested so much effort developing the model that everyone should be able to afford it.
This contrasts sharply with Western counterparts charging $50-$100 per million tokens for premium models. DeepSeek prices so aggressively that competitors can't match them without breaking even.
AGI-Only Focus, No Distractions
The strangest thing about DeepSeek: they're ignoring the hottest trends in AI. While everyone else races to deploy video generation, 3D models, and general-purpose multimodal capabilities, DeepSeek deliberately excludes these from their roadmap.
"When video generation first came out it was very hot, as if it were something you had to do, as if you weren't an AI company if you didn't," Liang said in the transcript. "That struck me as strange, because if you actually think it through, it has nothing to do with the intelligence roadmap."
Commercially, it's a good business. But not for DeepSeek. They won't do something just because it's profitable. Only what's on the intelligence roadmap matters.
This single-minded focus creates tension with the reality of competing against companies like ByteDance, Moonshot AI, and Z.ai -- domestic rivals now racing hard in the open-model space. Pressure to monetize eventually came, leading DeepSeek to finally accept outside capital.
Money Arrives on DeepSeek's Own Terms
DeepSeek recently raised RMB 50 billion (~$7.4 billion) in its first-ever financing round at a pre-money valuation exceeding $50 billion. The company is reportedly in talks to raise another $1.5 billion at a $71-74 billion valuation ahead of a planned 2027 IPO.
This reversal reflects years of refusing outside capital -- until the competitive pressure became unavoidable. Domestic rivals advancing in open-model space created a tipping point. Money arrived, but on DeepSeek's terms: the funding will be deployed primarily to acquire compute resources, not to fund unnecessary expansion or marketing.
The Computing Power Gap -- And How They're Closing It
Liang was blunt about constraints. Computing power is DeepSeek's single biggest limitation, and he argued the real gap with the US lies in resource access rather than technical know-how.
"I hope to be able to buy chips at a reasonable price so that I don't have to make chips myself," he said, though he confirmed plans to build large computing clusters eventually.
The numbers speak volumes. China currently has approximately 20,000 H100-equivalent GPUs, compared to the US reaching massive scale. Training Large Model requires 50,000 NVIDIA H100 cards or 200,000 domestic Huawei Ascend cards. The time lag between China and the US remains approximately 12 to 18 months.
DeepSeek's strategy? Buy as many cards as possible, ideally spending all funds within six months. "Buying cards is more profitable than putting money in a bank," Liang noted with characteristic bluntness.
The Roadmap From Agents To Embodied Intelligence
Liang described DeepSeek's technological evolution as a staircase of increasingly capable systems:
Step 1 - Chain of Thought (Last Year): Discovering that chain-of-thought reasoning can push intelligence to higher levels.
Step 2 - Agents (This Year): AI agents can handle more tasks with broader capability ranges and higher ceilings, using CoT as their foundation. Agents currently can't replace employees because they lack continuous learning -- they need two months of environmental familiarization before operating effectively.
Step 3 - Continuous Learning: The next bottleneck and critical breakthrough needed worldwide. A solution that allows models to learn continuously after training, like humans do, would be transformative.
Step 4 - Singularity / Self-Iteration: Once continuous learning is solved, the model can develop its own next version and research more advanced technologies autonomously. This self-iterating singularity is gradual, not sudden.
Step 5 - Embodied Intelligence: The final step puts AI in the physical world -- housework, elder care, concrete human needs. No human intervention required; the model emerges on its own.
Two parallel development lines feed this roadmap: formal projects requiring division of labor (taking no more than half the timeline), and free exploration without KPIs or supervision (taking at least half the timeline).
No Overtime, No Stress, Just Stable Teams
One of DeepSeek's most unusual policies: no overtime. Research requires a relaxed environment; too much pressure makes research impossible. The company is focused and restrained -- few imperfect products exist, no effort wasted on unnecessary polish.
There's one core interest: team stability. This is the only non-negotiable factor. As long as core personnel stay, AGI will definitely be achieved, everything else being merely a matter of time -- at most six months to a year later.
"Money and resources are not issues," Liang stated. "The talent advantage isn't that my people are smarter than his. It's how I organize these people, how I motivate them, and then how they collaborate."
Lessons From Jack Welch
Liang Wenfeng's most admired manager is former GE CEO Jack Welch. He believes Welch was right about one crucial insight: managing a large company relies not on rules and regulations, but on vision.
Vision isn't a slogan hanging on the wall. Vision is how you act, not how you talk. It's how you actually operate. DeepSeek's vision isn't even written down -- it's never been put into words, never written anywhere. It lives in how they do things, in their attitude toward the world.
This absence of written documentation, performance reviews, and conventional management structures is deeply unconventional by Silicon Valley standards. Yet it seems to work remarkably well for DeepSeek.
A Strategy Built on Conventional Wisdom?
What appears most remarkable about DeepSeek's approach is how it simultaneously defies conventional wisdom while embodying traditional Chinese values: humility, discipline, and long-term ambition over fast commercial cashouts. In mainland Chinese culture, the leaked transcript resonated deeply precisely because it aligned with cultural values that prioritize collective progress over individual glory, patience over quick wins, and substance over showmanship. Liang's philosophy extends beyond technology choices. When asked why DeepSeek wouldn't become the next ByteDance or Tencent, he dismissed the comparison entirely. "We don't want to become the next super app," he said. "No such thought at all. The AGI opportunity ahead is enormous." The consumer users and B2B revenue aren't the goal -- they're byproducts. Byproducts of building toward AGI. This inversion of typical startup logic -- where customer growth precedes product development -- represents a fundamentally different company DNA.The Western Counterpart Comparison
As DeepSeek implements this restraint strategy, Western labs pursue aggressive monetization pathways. OpenAI's GPT-5.6 Sol commands premium pricing with enterprise contracts locked behind restrictive terms. Anthropic's Claude Opus 5 beats Fable 5 at half the price but still operates within a commercial framework. Even Google's Gemini lineup prioritizes commercial integration above pure technical advancement. DeepSeek stands alone in treating openness as competitive advantage rather than concession. Their V3 and R1 models already matched far pricier US systems despite Washington's export curbs on advanced chips. This credibility transforms their open-source commitment from idealism into strategic positioning.Market Implications
DeepSeek's trajectory carries significant implications for the global AI landscape. If their continual learning breakthrough materializes -- even incrementally -- it could render today's static model architectures obsolete overnight. The self-iterating singularity they describe isn't theoretical speculation; it's engineering target number one. Meanwhile, their pricing structure forces competition into an uncomfortable position: either match DeepSeek's near-cost pricing (accepting thinner margins) or concede leadership in developer ecosystem formation through open weights. Both paths constrain competitor flexibility. Neither offers obvious advantages.What's Next?
DeepSeek announced recent completion of first funding round at roughly $52 billion valuation, reversing years of refusal of external capital. Pressure mounted from domestic rivals racing hard in open-model space, necessitating capital infusion while maintaining operational autonomy. Plans call for deploying funds rapidly -- ideally within six months -- primarily for acquiring computational resources. Domestic chip substitution remains secondary priority compared immediate GPU acquisition necessity given current infrastructure limitations. The coming months will reveal whether sustained restraint enables genuine breakthrough or merely delays inevitable commercialization pressures. Either way, DeepSeek has already rewritten expectations about what's achievable with constrained resources and unfocused dedication.🔥 Hot Takes
1. Restraint is the ultimate competitive advantage. Every founder chasing quarterly revenue misunderstands the game. DeepSeek's "ten-month payback" pricing strategy doesn't maximize short-term gains; it builds unbeatable developer gravity and prevents competitors from matching deployment economics. The best moat isn't what you keep -- it's what you willingly give away.
2. Your worst constraint isn't talent or ideas -- it's compute. Liang admitting China's gap lies in hardware access, not genius, quietly dismantles the narrative that Chinese AI lags due to technical inferiority. With 20,000 H100-equivalents versus US massive scale, fixing this buys a 12-18 month gap closure faster any algorithmic improvement could achieve. Buy cards first, think later.
3. No overtime is a revolutionary policy nobody copies. In an industry normalizing burnout as badge of honor, DeepSeek recognizing "too much pressure makes research impossible" represents sane management grounded in actual understanding creative work dynamics rather than buzzwords. Team stability matters more than speed because steady minds produce better breakthroughs than stressed ones.
4. Ignoring video generation is braver than chasing it. While every AI company desperate appearing relevant launches video capabilities DeepSeek dismissing them as irrelevant to intelligence roadmap demonstrates rare intellectual honesty. Willingness say no commercially lucrative opportunities proves conviction toward singular mission -- ultimately differentiating factor between fleeting fads and enduring impact.
Sources: SCMP Tech + RecodeChina.ai + AiBoss summary of Investor Seminar + TechJuice + internal transcript