DeepSeek's founder Liang Wenfeng has articulated a radically different philosophy for building artificial general intelligence (AGI), one that prioritizes technological breakthroughs over commercial returns and rejects the pursuit of quick wins in adjacent markets. In a 3-hour and 44-minute closed-door investor meeting held in Hangzhou on May 20, 2026, Wenfeng outlined a thinking model centered on strategic restraint, unwavering focus on AGI development, and a willingness to abandon profitable opportunities to maximize the probability of achieving his ultimate goal.

What Makes Liang Wenfeng's Strategy So Different From Other AI Leaders?

Wenfeng's philosophy stands in stark contrast to the mainstream logic of China's technology industry. At the investor meeting, he made several statements that challenged conventional wisdom about how AI companies should operate. He declared that "DeepSeek has only one main line, which is AGI, and everything else is a strategy," and emphasized that "the more restrained you are, the more likely you are to make this happen." He also stated explicitly: "I don't want to be the next ByteDance or Tencent, and we don't take commercialization as our goal".

This meeting occurred just days after DeepSeek completed the largest single round of financing in China's AI industry history, raising approximately 51 billion yuan with a post-investment valuation of about 400 billion yuan. Investors including Tencent, CATL, NetEase, JD, and the National Artificial Intelligence Fund participated in the funding round. Remarkably, at this capital-rich moment, Wenfeng repeatedly emphasized restraint to investors rather than growth or monetization.

How Does Wenfeng Define "Restraint" as a Business Strategy?

Wenfeng's concept of restraint is not conservatism but rather a deeper form of strategic thinking. He redefines the relationship between restraint and achievement, arguing that giving up immediate opportunities increases the probability of achieving the ultimate goal. This restraint manifests in concrete choices: he explicitly stated that DeepSeek would not pursue 3D generation, video generation, or world models, despite having the technical capability to dominate these high-demand markets.

He also set boundaries around commercialization itself. When discussing API pricing, Wenfeng explained that DeepSeek does not seek exorbitant profits but only reasonable ones. The pricing standard is calculated based on recovering equipment costs in ten months plus a profit margin of approximately six times the base cost. He argued that further price reductions would not generate meaningful demand elasticity and therefore make no business sense.

This philosophy has remained consistent over three years. In a 2024 interview, Wenfeng stated that "grabbing users is not our main purpose" and that "we just do things at our own pace." By 2026, this position had become clearer and more assertive, reflecting his conviction that at moments of technological generational change, allocating resources to the "right thing" rather than the "profitable thing" yields higher long-term returns.

Wenfeng

What Is DeepSeek's Specific Roadmap to Achieving AGI?

If restraint is the surface of Wenfeng's strategy, persistence in AGI is its core. What distinguishes his approach is that he has broken down the abstract concept of AGI into an extremely specific and clear technical route. At the 2026 investor meeting, he outlined a precise intelligence development ladder with defined stages and identified bottlenecks.

• Language Models: The foundation of current AI systems, capable of understanding and generating human language at scale.
• Chain of Thought (CoT): Enabling AI to reason through problems step-by-step, improving accuracy on complex tasks.
• Agents: AI systems that can take autonomous actions and interact with their environment to accomplish goals.
• Continuous Learning: The next critical bottleneck, allowing AI to learn and improve from new data without retraining from scratch, a problem not yet solved by the global academic community.
• Self-Iteration Singularity: A theoretical stage where AI systems can improve themselves autonomously.
• Embodied Intelligence: AI integrated with physical systems, capable of interacting with the real world.

Wenfeng emphasized that the next key bottleneck is "continuous learning," noting that "AI now does not lack taste and intuition, but the ability to continuously learn." He also clarified that multimodality and search are only "components," not the main line of intelligence itself. This roadmap demonstrates that his pursuit of AGI is not vague belief but rather a combat map with clear milestones and priorities.

Wenfeng

How Has DeepSeek's Technical Progress Validated This Strategy?

DeepSeek's technological achievements in 2026 provide concrete evidence that Wenfeng's focused approach is yielding results. On April 24, 2026, DeepSeek released V4, which continues the Mixture of Experts (MoE) architecture with a core feature of million-token context. The model uses a hybrid mechanism of Compressed Sparse Attention (CSA) and Heavy Compressed Attention (HCA) to significantly reduce attention calculation complexity from O(n²), making the model more efficient.

The performance metrics are substantial. V4 achieved a comprehensive score of 98.5 on SuperCLUE, surpassing GPT-5.4. On Codeforces, a competitive programming platform, V4 scored 3206, reaching the top 0.01% of human programmers. By May 2026, V4-Flash had reached monthly token usage of 18.4 trillion, surpassing GPT-4o to become the world's most widely used large model.

These benchmarks translate to practical capabilities: the model can process roughly 100,000 words at once, responds nearly instantly, and demonstrates exceptional reasoning ability on complex programming and mathematical problems. The widespread adoption of V4-Flash indicates that Wenfeng's strategy of focusing on core model capabilities rather than pursuing adjacent markets has resonated with users seeking reliable, high-performance AI tools.

Steps to Understanding DeepSeek's Competitive Positioning

• Evaluate Core Capabilities: Compare DeepSeek's performance on standardized benchmarks like SuperCLUE and Codeforces against competitors to understand where the model excels and where it faces limitations.
• Assess Resource Efficiency: Examine how DeepSeek achieves high performance with architectural innovations like Compressed Sparse Attention, which reduces computational overhead compared to traditional transformer models.
• Monitor Adoption Metrics: Track token usage and user adoption rates as indicators of real-world utility and market acceptance, rather than relying solely on marketing claims or theoretical benchmarks.
• Analyze Strategic Focus: Observe which product categories DeepSeek deliberately avoids, such as video generation and world models, to understand how the company's resource allocation reflects its long-term priorities.

Wenfeng's three-year consistency in his core philosophy, even as the landscape of China's AI industry has transformed dramatically, represents the most distinctive feature of his thinking model. His willingness to articulate this philosophy publicly to investors, emphasizing restraint over growth, suggests a founder confident that technological superiority and focused execution will ultimately determine competitive outcomes in the race toward AGI.