Summary of Key Points
Liang Wenfeng, the founder of DeepSeek, was once a symbol in the AI community for refusing to seek funding or commercialization. He relied on his quantitative fund, Huanfang, to achieve results comparable to top-tier models by adopting an open-source and low-cost approach. However, now he has suddenly launched its initial round of financing, aiming to raise 50 billion yuan (with 20 billion yuan coming from his own investment, and participation from companies like Tencent and CATL), valuing the company at 350-400 billion yuan. The reason for this move is the intensifying competition in the industry: top companies are poaching talented individuals at high salaries, the cost of computing power has soared, and giants are launching a full-stack counterattack, forcing him to consider financial support. Despite this, he is still committed to maintaining control over the company through substantial personal investment and multiple sources of financing, while continuing to uphold his principles of open-source development, low costs, and promoting domestic alternatives.
From "Having Enough Money" to "Desperately Needing Money": The Compromises Behind Liang Wenfeng's Decision
Previously, Liang Wenfeng claimed that "money was never an issue for DeepSeek," but now he is seeking funding due to three unavoidable challenges:
1. Talent Poaching: DeepSeek’s core researchers have become targets for large companies. For example, the "AI genius girl" Luo Fuli was recruited by Lei Jun with a salary of tens of millions, and Guo Daya, who proposed the core algorithm, has joined ByteDance; Ruan Chong, in charge of multimodal technologies, has moved to YuanRong Qixing. Large companies offer salaries in the tens of millions or even hundreds of millions, while DeepSeek lacks mature equity incentives to retain its talent, making financing the only way to offer them attractive benefits.
2. The Cost of Computing Power: In the early days, DeepSeek’s success relied on a large stock of GPUs (such as 10,000 A100 cards) to achieve low costs. Now, as it transitions to Huawei’s Ascend platform for domestic alternatives, the underlying code needs to be rewritten, and the cost of a single training session exceeds 500 million US dollars. The existing stock of GPUs is insufficient.
3. Giant Counterattack: In 2025, DouBao surpassed DeepSeek in monthly active users, and by 2026, giants like DouBao, QianWen, and YuanBao will all have over 100 million monthly active users. The competition has shifted from a focus on individual technological breakthroughs to a full-stack battle involving computing power, talent, products, and ecosystems. DeepSeek’s advantages in open-source development and low costs are being diluted, necessitating financial investment to address these shortcomings.
The "Control Game" Around the 50-Billion Yuan Financing: Liang Wenfeng Does Not Want to Be a "Employee"
The key to this financing round is not just to raise money but to maintain control:
- 20 Billion Yuan in Personal Investment to Lock in Dominance: By investing 20 billion yuan, Liang Wenfeng will hold 40% of the total funds. Combined with his existing direct and indirect equity holdings, he will still have significant influence, ensuring that DeepSeek follows his vision rather than being controlled by external investors.
- Rejection of Exclusive Investments: Previously, some large companies wanted to make exclusive investments but requested a 20% stake, which Liang Wenfeng refused. This time, he has chosen multiple partners (fewer than 10), ensuring that no single entity can dominate the company’s direction.
Changes and Constants at DeepSeek: Compromises Yet Unwavering Commitment
Although Liang Wenfeng has changed, some aspects remain unchanged:
Changes:
- Shift from a Laboratory Model to Accelerated Commercialization: DeepSeek is now hiring product managers, building new teams, and selling models to businesses, enhancing its capabilities in image and audio processing to turn technology into profitable products.
- From Rejecting Capital to Embracing It: He is actively seeking funding and even aligning the company with industry standards, planning for revenue growth.
Constants:
- Open-Source and Low-Cost Philosophy: The V4 version has reduced prices to just a few cents per million tokens (hundreds of times cheaper than GPT-4), and Tencent Cloud has also lowered its prices by 97.5%, continuing DeepSeek’s role as a "Pinduoduo" in the AI industry.
- Algorithm Innovation: The V4 version uses an mHC architecture to compensate for weaknesses in computing power, betting on the success of domestic Ascend platforms and focusing on reducing costs through technology rather than simply increasing investment in hardware.
- Determination to Promote Domestic Alternatives: V4 is compatible with both NVIDIA and Huawei’s Ascend platforms, showing a clear commitment to making China a contributor to AI development rather than just a beneficiary.
Industry Trends: The Era of Full-Stack Competition in AI
DeepSeek’s financing reflects a turning point in the industry:
- Shift from Focusing on Individual Technologies to Competing on Comprehensive Capabilities: The focus has shifted from algorithms to computing power (whoever has more GPUs), talent (whoever can retain top researchers), products (whoever has more consumer users), and ecosystems (whoever has broader partnerships). Giants like Tencent, Alibaba, and ByteDance possess substantial resources in these areas, and startups like DeepSeek need funding to keep up.
- Divergent Paths: While DeepSeek continues to offer affordable solutions, other companies like DouBao are launching premium paid versions. The AI industry is beginning to differentiate, with some focusing on low-cost alternatives and others on high-end services.
Future Challenges: Can Money Solve the Problems?
Raising 50 billion yuan will alleviate current pressures, but DeepSeek still faces two major challenges:
1. Revenue Growth: With a valuation of over 50 billion US dollars, the company needs annual revenue of at least 500 million US dollars. However, DeepSeek’s revenue is currently limited to API sales, and many companies use its technology without paying for it. It must diversify its revenue through product development (e.g., agent tools, enterprise services).
2. The Success of Domestic Alternatives: Transitioning to Huawei’s Ascend platform requires rewriting code, which increases training costs. Whether DeepSeek can match the performance of NVIDIA’s systems is a critical bet. If successful, it could break the monopoly on computing power; otherwise, it may fall behind.
In summary, Liang Wenfeng’s financing is not a sign of defeat but an attempt to strengthen his position with capital, allowing him to continue pursuing his vision of making AI more accessible through open-source development and algorithm innovation while promoting domestic alternatives. Whether this approach will be successful depends on his ability to balance the needs of capital, talent, and technology.