DeepSeek’s Four‑Hour Investor Briefing: Pursuing AGI and the Path to Embodied Intelligence

DeepSeek recently completed a 500 billion RMB financing round, after which founder Liang Wenfeng spent 3 hours 44 minutes in a closed‑door discussion with investors, covering strategy, business model, technology roadmap, resource bottlenecks, and organizational culture.

Liang emphasised “restraint” as the core strategy: the company will not chase profit maximisation, C‑end traffic, or B‑end hype, and will deliberately avoid video‑generation and world‑model projects, treating them as distractions from the AGI goal.

On commercialisation, Liang clarified that open‑source is not charity. DeepSeek aims for a six‑fold profit margin, pricing its V3.2 API at 0.2 CNY per million input tokens and 3 CNY per million output tokens, a level that recovers costs in ten months and leaves little room for third‑party profit.

Both C‑end and B‑end products are viewed as “by‑products” that support the AGI mission; B‑end API revenue could reach hundreds of millions of dollars this year, with a $1 billion ARR unlocking positive cash flow for R&D.

Liang presented a five‑step AGI ladder: Chain‑of‑Thought (CoT) → Agent → Continuous Learning → Self‑Iterating singularity → Embodied intelligence. He noted that CoT is already mature, agents are emerging, and continuous learning is the next mountain to climb; without it, future models would be mere incremental improvements.

The founder argued that scaling limits are primarily compute‑bound, not algorithmic, and that video generation or world‑model research does not belong on the core path.

Team stability is the only non‑negotiable factor. Liang quoted, “If everyone stays, I can achieve AGI.” The recent financing is intended to secure talent via equity, as the company has no KPIs, no formal vision, and relies on a vision‑driven structure where employees spend up to 50 % of their time on self‑directed research and avoid overtime.

Additional signals: DeepSeek will not vertically integrate chip production, preferring reasonable chip pricing; the Chinese large‑model market is expected to consolidate to two major and two minor players; cost, speed, and user experience will dominate competition, with OpenAI and Google alternating leadership and Anthropic’s lead seen as temporary.

In conclusion, DeepSeek’s “restraint” reshapes AI business logic by targeting modest profit margins that create a defensible moat, contrasting sharply with the traditional internet model of burning cash for scale.