Liang Wenfeng breaks his silence: DeepSeek's open-source doctrine and the road to AGI

Key takeaways from a rare nearly four-hour investor presentation by DeepSeek founder Liang WenfengDealroom has a profile for this one. Try Dealroom →, held on 20 May 2026 and made public in July 2026 via a translated transcript:

• AGI, not profit, is the goal: Liang frames DeepSeek's founding intention as building something useful for humanity rather than maximising commercial return or chasing an IPO; he argues AGI will carry enormous commercial value regardless, so business wins are treated as side products of the technology, not the objective.
• Restraint as strategy: he repeatedly returns to "restraint" — deliberately limiting the share of value the company captures — as the thing that makes success more likely, arguing AI is too large (potentially ~10% of GDP) for any single player to monopolise without being "abandoned by history".
• Open source is core, not tactical: unlike a traditional software firm, Liang contends AI's market is big enough that open-sourcing does not destroy commercial value; open source is the foundation of the ecosystem and the mechanism that lets partners build on DeepSeek's models.
• Reasonable-profit pricing: API prices are set to recover hardware cost in roughly ten months rather than to maximise profit; he notes demand is largely inelastic at these levels, and that rivals cannot match the underlying cost because they lack DeepSeek's optimisation — estimating Alibaba's and Tencent's costs run several times higher.
• US lead is cyclical and compute-driven: he argues the edge of OpenAI, Anthropic and Google stems mainly from greater computing power, not talent, and will not last; Anthropic's early coding-agent advantage, in his view, will fade.
• China's structural advantages: Liang sees built-in Chinese advantages in cost and user experience, expects the CUDA ecosystem wall to be eroded by new technology, and views building an ecosystem for Chinese chips as a historic opportunity that would close the compute gap once fabrication limits are overcome.
• A future without giants: he predicts the base-model field will consolidate to three or four fully-competing players, that anyone chasing very high profits will be pushed out, and that DeepSeek will focus narrowly on the road to AGI — language models, chain-of-thought, agents and continual learning — while deliberately ceding areas like video generation and world models to the wider ecosystem.
• Continual learning as the missing piece: Liang argues current approaches (including chain-of-thought, which he says has already surpassed top humans on olympiad-level maths) are hitting a ceiling, and that continual learning is needed to progress toward AGI.

Read more: Inside China (Fred Gao) · via X