DeepSeek Is Not an Idealistic Company, It Just Subordinates Reality to Ideals

A leaked transcript of Liang Wenfeng's investor communication meeting reveals DeepSeek's strategy. The long-term goal is AGI, broken down into steps: language model, CoT, Agent, continuous learning, self-iteration, and embodied intelligence. The company only works on the AGI main line, excluding video generation, 3D, and other areas deemed irrelevant to intelligence. Research is organized as '摸奖' (lottery) — low-threshold exploration without strict KPIs, with half of core researchers also doing data labeling. DeepSeek believes in scaling but is limited by compute resources. The company has no formal organizational structure, no mandatory overtime, and half of employees' time is free for self-directed exploration. Founder authority comes from consensus, not command. Open-source is pursued as both a vision and a business tactic: the moat is not the model itself but deployment cost and engineering efficiency. Pricing targets a 10-month payback on equipment cost with 6x profit margin, deliberately avoiding profit maximization. Team stability is the top priority; Liang believes that as long as the core team stays, AGI is achievable. DeepSeek avoids vertical integration except for self-built clusters, and does not plan to develop chips unless the benefits are substantial. The company does not compete aggressively for users, but still captures casual opportunities like C-end daily active users without significant effort.