Summary of Key Points
DeepSeek has completed its first round of financing, raising over 50 billion yuan. The founder, Liang Wenfeng, has firmly secured control of the company through a unique transaction structure that includes “no voting rights for external investors” and a five-year lock-up period, holding 100% of the voting rights. The company emphasizes being driven by vision rather than focusing on short-term listing or profitability. However, DeepSeek faces challenges such as technical shortcomings (including “illusions” and overthinking in its algorithms), strategic decisions (abandoning multi-modal capabilities), talent loss, and reliance on computing power. There are three possible outcomes for the company’s future: optimistic, pessimistic, or neutral. Its financing model and development path offer important insights for the AI industry.
I. Financing is Not About “Raising Money”; It’s About Managing Capital – Why Does DeepSeek’s “Unequal Treaty” Work?
What makes DeepSeek’s financing unique is not the amount of capital raised (50 billion yuan), but the transaction structure: Apart from the National Artificial Intelligence Industry Fund, other investors (such as Tencent and CATL) have no voting rights or board seats; they can only receive dividends and are restricted from exiting for five years. Through additional capital contributions, founder Liang Wenfeng now holds 84.29% of the beneficial shares and 100% of the voting rights.
Why is capital willing to accept this “unequal” arrangement?
- Technical scarcity: DeepSeek is the only major domestic company that adheres to a purely open-source approach for its large models, outperforming mainstream open-source models, thus giving it significant technical influence.
- Diverse investor priorities: The state wants to achieve “independence in computing power” (hence retaining voting rights); Tencent and CATL aim to gain strategic advantages; JD.com and NetEase seek “general-purpose API services” with immediate benefits; financial VCs are betting on future exits.
- Industry trends: Capital is converging towards leading companies. By 2025, only a few firms will raise over 1 billion yuan in financing, and DeepSeek is one of them, making it an attractive target for investors.
II. The Lack of KPIs Does Not Mean Lack of Efforts; It Reflects Technical Confidence – But There Are Costs
Liang Wenfeng claims the company has no KPIs or evaluations, which is not arbitrary but is supported by a rigorous technical methodology:
- GPU optimization: Directly optimizing NVIDIA’s high-end cards using the PTX instruction set to overcome memory bandwidth limitations.
- Efficient models: Using the MoE (Mixture of Experts) architecture to activate only necessary parameters and reduce computational load.
- Reducing communication delays: Utilizing pipelining parallelism to circumvent chip export restrictions.
However, this approach also has drawbacks:
- High costs: The three-year maintenance cost for MoE models is 1.3–1.5 times that of regular models, and the inference cost is three times higher.
- Limited compatibility: PTX optimizations are only effective on NVIDIA’s high-end cards and may not work with domestic alternatives.
- Inference efficiency: MoE models can overthink simple problems, leading to errors when handling straightforward tasks.
III. Hidden Concerns Behind the Glorious Financing: Technical Issues, Strategic Risks, and Talent Loss
DeepSeek’s problems are subtle:
1. Technical shortcomings: Its “illusions” (erroneous predictions) are more subtle; for example, recommending fictional food delivery services or suggesting “南极科考 to regain a former boyfriend” can be problematic in real-world applications like law and healthcare, potentially limiting paid customer acquisition.
2. Strategic choices: Abandoning multi-modal capabilities (such as video generation) may save resources in the short term but could lead to long-term talent loss and a lack of relevant data, hindering the development of general AI (AGI).
3. Talent turnover: Some departures are strategic decisions (e.g., focusing on other areas), but the core issue is that if multi-modal capabilities are not pursued, top talents may leave, impacting future growth.
IV. Three Possible Futures: Becoming an AI Infrastructure or Being “Kidnapped” by Capital?
How long DeepSeek’s model will be successful? There are three possible outcomes:
- Optimal outcome: Continuous technological breakthroughs (e.g., incremental learning to reduce computing power demands), validation of the AGI path, and becoming a leading AI infrastructure provider with substantial profits for all investors.
- Pessimistic outcome: Slow progress in AGI development, erosion of technical advantages, continuous losses leading to diluted equity, and pressure from capital that forces the founder to commercialize the company.
- Most likely outcome: Periodic technological improvements, but AGI will take 8–10 years to achieve. The company may focus on limited commercialization in specific industries (e.g., selling APIs), generating stable cash flows without pursuing high profits. Industrial capital will continue to invest, while financial investors may exit through share transfers, with the founder retaining control.
V. What Can AI Startups Learn from DeepSeek?
DeepSeek’s approach offers two types of lessons for the industry:
- Replicable elements:
- Control structure: Using a “no voting rights + lock-up period” to shield the company from short-term capital pressures and protect its technical direction.
- Open-source strategy: Allowing access to core optimizations (e.g., PTX instructions) while setting prices that undercut competitors’ profits, and leveraging developer communities for data collection.
- Management scale: No KPIs are suitable for companies with 300–500 employees; larger teams need more streamlined management.
- Non-replicable elements:
- The founder’s personal investment of 20 billion yuan (40% of the total financing) is a result of significant personal wealth, something not easily achievable by most startups.
- Special terms from the state-backed investors are a result of policy support and not available to everyone.
In summary, AI startups should learn to use institutional mechanisms to protect their vision, rather than simply imitating practices that may not be profitable or involve large amounts of financing.
Conclusion
DeepSeek is an outlier in the AI industry: it uses capital without being controlled by it, focuses on a clear vision with technical support, and has both strengths and weaknesses. Its story highlights that AI entrepreneurship is about balancing technology, control, and strategy – achieving funding while maintaining direction and addressing practical technical and commercial challenges. Whether DeepSeek’s approach will be successful depends on its ability to continuously innovate and the impact of geopolitical factors related to computing power. Regardless of the outcome, its financing model and strategic choices provide valuable lessons for the AI industry.