A leaked recording of DeepSeek founder Liang Wenfeng's four-hour investor briefing has already done something no chip embargo managed: freeze the company's own fundraizing.
DeepSeek told prospective backers in a planned second funding round, targeting a valuation near 74 billion dollars, that the deal is suspended for now, Bloomberg reported. The pause stems in part from Liang's frustration over online reports about comments he made during the first financing round, which closed in June and raised about 7 billion dollars from Tencent, CATL, JD.com, NetEase and IDG Capital alongside a state AI industry fund. Liang personally put in roughly 20 billion yuan of his own money. The South China Morning Post reported the transcript offered an unprecedented look into a founder who has given roughly two interviews in his life. Neither Bloomberg nor Reuters has verified the transcript's authenticity, though the funding freeze it triggered is confirmed.
Liang lays out a pricing rule that recovers hardware costs in about ten months, a margin he calls sixfold profit, arguing that charging more only invites a rival willing to take less to undercut him. He frames DeepSeek's refusal to chase super-app dominance, the huge user base it built over one Spring Festival and deliberately declined to monetize, as proof no company can hoard a technology this size. He calls team stability the company's one core interest, run through a no-KPI, vision-driven structure with no formal chain of command. That is the oldest trick in monopoly capitalism: undercut margins until rivals who need those margins to survive get starved out, then own the field once they are gone.
Liang describes AI progress as a staircase, chain-of-thought last year, agents this year, then continuous learning, a self-iterating singularity, and finally embodied robots doing labor and elder care. DeepSeek is deliberately skipping video generation and world models, leaving those markets to competitors. He predicts Anthropic's current coding-agent edge will fade, with OpenAI, Anthropic and Google continuing to trade the lead among themselves.
Underneath the philosophy sits a resource wall built by Washington's own export controls. DeepSeek trains on roughly 20,000 Nvidia-equivalent GPUs while US frontier labs run models with 800 billion active parameters, a scale Liang says DeepSeek cannot yet afford to train or serve. Huawei supplies just 16,000 of its own chips, equivalent to about 4,000 Nvidia-class cards, and Liang said that ceiling comes from Huawei's manufacturing output. He argues Nvidia's CUDA software moat is already dissolving as Chinese engineers use AI to rebuild that stack around domestic chips.
Sam Altman and Dario Amodei are burning capital chasing trillion-dollar valuations off closed models. A Chinese lab undercutting
https://www.fredgao.com/p/deepseeks-liang-wenfeng-breaks-his
tangentwithtanner
izzy806
amandla
blakekathryn
origamiaround
prguitarman