[In-depth] Behind DeepSeek's Success: A Profile of China's Most Technical AI Company
"A Shot Was Fired in the AI Price War": Interview with DeepSeek CEO
A dive into how a Chinese startup challenges Silicon Valley's AI dominance through radical technical innovation and unconventional choices.
The Technical Idealist
DeepSeek's CEO Liang Wenfeng stands out in China's AI landscape. Unlike his startup counterparts focused on rapid commercialization, Liang champions a different philosophy: focus on foundational research and open-source innovation.
Key aspects that define his approach:
Prioritizes technical innovation over immediate commercialization
Committed to open-source development
Runs DeepSeek more like a research lab than a traditional company
Believes China can contribute to fundamental AI innovation
Building DeepSeek Differently
Several decisions distinguish DeepSeek from other Chinese AI companies:
The Price War Pioneer
Triggered an industry-wide price reduction in May
API costs 1/53rd of Claude 3.5 Sonnet
Maintains profitability despite low prices through architectural innovation
Research Over Revenue
Only Chinese AI startup focused purely on research
No venture funding raised to date
Avoids B2C applications to focus on foundational models
Emphasizes long-term technical advancement over short-term gains
Talent Strategy
Hires primarily fresh graduates from top Chinese universities
Focuses on curiosity and passion over experience
Allows complete freedom in resource allocation
No hierarchical structure or formal approval processes
The Innovation Philosophy
Liang's perspective on innovation challenges conventional wisdom:
On China's Role: "The real gap isn't one or two years—it's between originality and imitation. China cannot remain forever in a follower position."
On Open Source: "In the face of disruptive technology, closed-source moats are temporary. Even OpenAI's closed approach can't prevent being overtaken."
On Competition: "More investment doesn't necessarily produce more innovation. Otherwise, big tech would monopolize all innovation."
Future Vision
Liang sees three critical paths to AGI:
Mathematics and coding
Multimodal capabilities
Natural language processing
He believes AGI will arrive within our lifetime, possibly in 2-10 years.
Management Style
DeepSeek's organizational structure reflects Liang's philosophy:
Bottom-up innovation encouraged
No preset task assignments
Flexible resource allocation
Open office design to encourage spontaneous collaboration
Looking Forward
Despite industry pressure to commercialize, Liang maintains his course: "All formulas are products of the previous generation. Using internet business logic to discuss future AI profit models is like discussing Coca-Cola's business model when Tencent was starting."
The story of DeepSeek and Liang Wenfeng represents a unique experiment in Chinese tech: can a purely research-focused, open-source company compete with global AI leaders? Their recent technical breakthroughs suggest it's possible, while their unconventional approach offers lessons for the broader tech industry about balancing innovation and commercialization.
[full interview script]
Q: After DeepSeek V2's release triggered a fierce price war, some call you an industry disruptor.
A: We didn't intend to be disruptive - it happened accidentally.
Q: Did this outcome surprise you?
A: Very much. We didn't expect such price sensitivity. We simply priced based on costs plus modest profit, following our principle of avoiding both losses and excessive profits.
Q: Zhipu AI followed in 5 days, then ByteDance, Alibaba, Baidu, and Tencent.
A: Zhipu only lowered their entry model. ByteDance truly led, matching our flagship pricing, triggering others. Since their costs exceed ours, we hadn't expected loss-leading pricing - it became typical internet-era subsidization.
Q: Externally, price cuts seem aimed at user acquisition.
A: That wasn't our primary goal. We reduced prices because our next-gen architecture lowered costs, and we believe AI should be accessible to everyone.
Q: Before this, most Chinese companies copied Llama's structure. Why innovate architecture?
A: Following Llama works for quick applications. But our goal is AGI, requiring new architectures for stronger capabilities with limited resources. Beyond structure, we research data construction and human-like behavior. Llama's structure lags about two generations in efficiency.
Q: Why focus solely on research when others pursue both models and applications?
A: Participating in global innovation matters most now. Chinese companies traditionally commercialized others' innovations, but this isn't inevitable. We aim to advance technology, not make quick profits.
Q: What drives these generational gaps?
A: Training efficiency differs significantly. China's best models require twice the compute for structure and dynamics, plus double the training data. Combined, that's 4x more compute. We're working to close these gaps.
Q: Why did DeepSeek V2 surprise Silicon Valley?
A: It's ordinary among U.S. innovations. The surprise came from seeing a Chinese company join as an innovator, not follower.
Q: Isn't focusing on research luxurious in China's context?
A: Innovation isn't cheap, but China now has the economic scale and corporate profits to support it. We don't lack capital - we lack confidence and expertise in organizing talent for innovation.
Q: Why do Chinese companies prioritize rapid commercialization?
A: The past thirty years emphasized profit over innovation. Innovation needs curiosity and creativity beyond commercial drivers. We're still bound by past habits, but this is temporary.
Q: How do you build competitive advantages while open-sourcing?
A: Closed-source advantages are temporary. Even OpenAI can't prevent being overtaken. Our value lies in team growth and innovative culture. Open-sourcing creates cultural attraction and honor.
Q: What's your view on market-focused perspectives like Zhu Xiaohu's?
A: His approach is internally consistent but suits quick-profit companies. America's most profitable companies are technology-focused with patient development.
Q: In large language models, pure technical leadership rarely creates absolute advantages. What bigger bet are you making?
A: We see that Chinese AI can't remain followers forever. While we say China is 1-2 years behind the US, the real gap is between originality and imitation. Without change, China remains a follower. NVIDIA's leadership isn't just one company's effort - it's the result of the entire Western tech community. They see next-gen trends and have roadmaps. Chinese AI needs similar ecosystems. Many Chinese chips struggle due to lack of supporting tech communities and relying on second-hand information. China needs frontline innovators
Q: DeepSeek has early OpenAI's idealism and open-source approach. Will you go closed-source like OpenAI and Mistral?
A: No. Building a strong technical ecosystem matters more.
Q: Any funding plans? Reports suggest Huanfang (high-flyer) plans to spin off DeepSeek. Silicon Valley AI startups eventually partner with big tech.
A: No short-term funding plans. Our challenge isn't money but high-end chip export bans.
Q: Many say AGI differs from quantitative trading - it needs alliances and publicity for more investment.
A: More investment doesn't guarantee more innovation. Otherwise, big tech would monopolize all innovation.
Q: Is your lack of applications due to missing operational expertise?
A: We see this as an era of technical innovation, not application explosion. Long-term, we want to create an ecosystem where industry uses our technology, we focus on foundation models and innovation, and others build B2B/B2C businesses. If we achieve complete industrial chain, we won't need our own applications. We can build them if needed, but research remains priority.
Q: Why choose DeepSeek's API over big tech?
A: Future likely brings specialized division of labor. Foundation models need continuous innovation - big tech has limitations here.
Q: Can technology really create gaps when there are no absolute technical secrets?
A: No secrets, but rebuilding takes time and resources. NVIDIA's GPUs have no theoretical secrets but are hard to catch up due to team-building and next-gen development time.
Q: After your price cut, ByteDance followed - suggesting they feel threatened. Your thoughts on startup vs. big tech competition?
A: We don't focus much on this. Cloud services aren't our main goal - AGI is. No new competitive solutions yet, but big tech lacks clear advantages. They have users but legacy businesses make them vulnerable.
Q: Your prediction for the other 6 AI model startups?
A: 2-3 might survive. Currently all burning money. Clear positioning and efficient operations will determine survival. Others will transform. Value won't disappear but will take new forms.
Q: In Huanfang era, you were known for independence, ignoring comparisons. What's your competition philosophy?
A: I focus on whether something improves social efficiency and finding our strength in the industry chain. If the end result increases social efficiency, it's valid. Everything else is temporary - overthinking creates confusion.
Q: Jack Clark of Anthropic thinks DeepSeek hired "mysterious talents." Who created DeepSeek V2?
A: No mysterious talents - just top university fresh graduates, late-stage PhD students, and young professionals a few years out of school.
Q: Many AI companies aggressively recruit overseas, believing top 50 AI talents aren't in Chinese companies. Where are your people from?
A: V2 team is entirely local. While top 50 talents might not be in China yet, we believe we can cultivate them.
Q: How did the MLA innovation happen?
A: After analyzing Attention architecture patterns, a young researcher designed an alternative. It took months and a dedicated team to implement.
Q: Your organizational structure seems innovation-focused. Has AGI's uncertainty required more management?
A: We remain bottom-up with natural division of labor. People bring their own ideas - no pushing needed. When ideas show promise, we allocate resources accordingly.
Q: How flexible is DeepSeek's resource allocation?
A: No limits on GPU or personnel use. Anyone can access training clusters without approval. No hierarchy means flexible collaboration based on mutual interest.
Q: Your management style depends on passion-driven people. How do you recruit?
A: We prioritize passion and curiosity. Many have unique backgrounds. Research desire outweighs financial motivation.
Q: How do corporate AI labs differ from startups in innovation?
A: Google, OpenAI, and Chinese tech AI labs all have value. OpenAI's success partly comes from historical chance.
Q: Is innovation largely chance? Your office design seems to encourage serendipity.
A: Innovation first requires conviction. Silicon Valley innovates because they dare to try. When ChatGPT emerged, China lacked confidence in frontier innovation. Innovation needs self-belief, often found in young people.
Q: How do you attract talent without fundraising publicity?
A: We tackle the hardest problems. Top talent wants world's most challenging issues. Elite talent is undervalued in China due to limited hardcore innovation opportunities.
Q: With GPT-5's delay, some question Scaling Laws. Your thoughts?
A: We're optimistic - industry progress meets expectations. OpenAI isn't invincible, can't always lead.
Q: Your AGI timeline and roadmap?
A: 2-10 years, within our lifetime. No internal consensus on roadmap. We bet on three directions: math/code, multimodal, and natural language. Math/code provides verified systems like Go. Multimodal real-world interaction may be essential. We remain open to all possibilities.
Q: What's the endgame for large language models?
A: Specialized companies will provide foundation models and services, with extensive professional specialization serving diverse social needs
Q: China's AI startup landscape changed this year, with Wang Huiwen's exit and companies differentiating. Thoughts?
A: Wang Huiwen took all losses himself, letting others exit cleanly. He made a choice that was worst for him but best for everyone. That shows his integrity, which I admire.
Q: Where's your main focus now?
A: Researching next-generation models. Many unsolved problems remain.
Q: Other AI startups balance research and products, since technology alone doesn't guarantee permanent leadership. Is DeepSeek's research focus due to insufficient model capabilities?
A: All formulas are products of their era. Using internet business logic for AI is like applying General Electric's model to early Tencent. It's outdated thinking.
Q: Is your optimism from Huanfang's successful tech-driven growth?
A: Huanfang strengthened our confidence in technology-driven innovation, but it wasn't easy. People see post-2015 success, but we worked for 16 years.
Q: Will economic downturn and cold capital markets suppress original innovation?
A: Not necessarily. China's industrial restructuring needs core technical innovation. As quick profits become harder, more will pursue real innovation.
Q: So you're optimistic about this?
A: I grew up in a small Guangdong city in the 1980s. My father was a teacher. In the 90s, many parents thought education worthless due to easy money opportunities. Now views have changed as opportunities shrink - even taxi driving jobs are scarce. One generation changed everything. Hard-core innovation will increase. Current skepticism exists because society needs proof. When innovative pioneers succeed, collective mindset will shift. We just need time and evidence.
The full interview mentions a lot of leading AI companies in China. Please let me know if you're interested in any one of them. There will be more in-depth post in the future about them.