Summary of Key Points
DeepSeek’s recent valuation has soared to 350 billion yuan, marking its first introduction of external capital from companies such as Tencent and Alibaba. The funds raised are being invested in two main areas: building a massive own infrastructure for computing power (constructing data centers and managing operations) and developing advanced application products (creating code intelligence tools). Actions like hiring IDC design engineers, maintaining grassland-based data centers, and forming a code development team modeled after Claude Code reveal DeepSeek’s ambition to go beyond being just an AI model company; it aims to establish a complete ecosystem from underlying computing power to end-user applications.
1. Financing Valuation of 350 Billion Yuan: Where Does the Money Come From, and How Will It Be Spent?
DeepSeek previously operated on its own funds generated through Huafang Quantization, with almost no external investment. However, with this recent round of financing, its valuation has jumped to 350 billion yuan (approximately $48 billion), attracting potential investors including giants like Tencent and Alibaba. The plan for the funds is as follows:
- Half will be invested in infrastructure: Hiring IDC design engineers and personnel for maintaining data centers, with the goal of building super-large-scale facilities.
- Half will be invested in product development: Establishing a code intelligence team to develop tools comparable to Claude Code, aiming to capture the developer market.
In simple terms, DeepSeek is focusing on “laying a solid foundation (computing power) before constructing the actual buildings (applications).”
2. Building Super-Large-Scale Data Centers: Why Would an AI Company Compete with Cloud Providers?
The most surprising hire is for an IDC Design and Planning Engineer, who is responsible for the entire process of data center construction, from site selection to design to implementation. This indicates that DeepSeek no longer intends to rely on rented data centers; instead, it plans to build its own facilities on a GW-level (1GW = 1000MW), which is capable of supporting thousands of GPUs simultaneously, matching the scale of a large-scale intelligent computing center.
The reasons for building its own data centers are:
- Insufficient Rental Options: Training large models with trillions of parameters requires massive computing power. Renting data centers is not only expensive but also risky, as resources may be prioritized for cloud providers.
- Lower Costs: By building its own facilities in a cool climate (e.g., Ulanqab), DeepSeek can reduce energy consumption (PUE < 1.2, which is over 20% lower than in first-tier cities), saving on electricity costs.
- Greater Control: With self-owned data centers, DeepSeek can ensure the stability of large model training and future inference services without relying on others.
Currently, Ulanqab’s data center facilities are already in the process of hiring maintenance and delivery personnel, suggesting that they will be put into use soon.
3. Advanced Applications: Forming a “Code Intelligence Team” to Compete with Claude Code
While infrastructure development focuses on the foundation, the new hire of an Agent Harness Product Manager and research engineers demonstrates DeepSeek’s focus on application development. A senior researcher, Chen Deli, is also recruiting to develop a tool similar to Anthropic’s popular AI programming tool, Claude Code.
The objectives of this move are clear:
- Differentiation in a Homogeneous Market: As large models become more similar, the company aims to differentiate by making its tools more user-friendly.
- Capturing the Developer Ecosystem: These tools will help programmers write code and debug, similar to GitHub Copilot, thereby encouraging them to use DeepSeek’s models for various applications and fostering an ecosystem.
- Commercialization: Tool-based products are easier to monetize, such as through usage fees or subscription models, providing more stable revenue sources than simply selling model APIs.
4. A Comprehensive Approach: Liang Wenfeng’s Ambition to Control the Entire Chain
Liang Wenfeng, DeepSeek’s founder and leader, has been quite low-key. However, these recent hires reveal his ambition to become a full-stack player in the AI industry, controlling the entire value chain from computing power to applications:
- Infrastructure: Solving bottlenecks in computing power to ensure sufficient resources for training large models.
- Application Products: Transforming model capabilities into user-friendly tools for commercialization.
- Long-Term Goal: Creating a closed-loop ecosystem that combines computing power, models, and applications, similar to the combination of OpenAI (models), Microsoft Azure (computing power), and Copilot (applications), but with DeepSeek taking control of all aspects.
This “heavy assets + product development” strategy is costly, but if successful, it can create a competitive barrier that others难以 overcome. After all, not everyone has the funds to build data centers or the expertise to turn models into useful tools.
Conclusion: DeepSeek’s Bold Bet
With a valuation of 350 billion yuan, DeepSeek is taking a risky but essential step to take control of its own destiny. By building its own computing infrastructure and developing valuable applications, it aims to become a dominant player in the AI industry. If this strategy succeeds, it could emerge as a “full-stack player” in China’s AI landscape. If not, it may exhaust its resources. However, Liang Wenfeng is betting on the future, where competition will ultimately boil down to computing power and ecosystem dominance.
DeepSeek’s approach reflects a clear understanding that success in AI depends on a comprehensive strategy that combines both heavy infrastructure investment and innovative product development.