DeepSeek founder Liang Wenfeng spent an investor call this month describing a company with no written vision, no KPIs, and a founder who says he mostly seeks consensus rather than issue orders. Geopolitechs compiled 64 quotes from that call on July 23. The framing is deliberately disarming. It is also, on close reading, a strategic document.
Treat everything here as positioning. Liang was talking to people deciding whether to fund him, not testifying under oath, and a founder’s self-description on a capital-raising call is marketing dressed as philosophy.
Start with open-weights. Liang said the company’s strongest model will keep being released openly, arguing that a commercial lab has no incentive to chase model efficiency because cheaper inference erodes its own revenue. As Liang put it, DeepSeek releases its models openly because “I don’t see what’s good about closed source.” That line reads as altruism. It functions as a supply-side attack. Every lab racing to build a defensible enterprise moat around a closed frontier model, OpenAI and Anthropic included, loses pricing power the moment a comparable open weight set is free to run. DeepSeek does not need to out-fund those labs. It only needs to make their closed models look overpriced, and giving away the weights does exactly that at zero marginal cost to a company already committed to thin margins.
The commercialization stance follows the same logic. Liang described API pricing built to recoup hardware costs in roughly ten months, explicitly rejecting profit maximization, and said demand at that price is inelastic enough that a 50 percent increase would barely move usage. He called both consumer and enterprise products incidental to the company’s push toward artificial general intelligence, not the business itself. A company that treats revenue as incidental can undercut every competitor that treats revenue as the point. That is restraint as a pricing weapon, not a values statement.
The AGI roadmap deserves more scrutiny than it is getting. Liang laid out a public sequence: chain-of-thought reasoning, then agents, then continual learning, then what he called a gradual singularity, then embodied intelligence. He put a number on the compute gap behind that roadmap, roughly 20,000 GPUs by his own H100-equivalent count against a leading edge that needs an estimated 200,000 cards to train the largest current models, and said DeepSeek is running 12 to 18 months behind the United States on that basis alone. Publishing a specific technical roadmap while conceding a compute deficit that large is unusual for a lab operating under active supply constraints. It reads as a hedge against this week’s discussion in Washington of sanctions targeting distillation, the practice of training smaller models on a rival’s outputs: if DeepSeek can point investors and regulators to a self-funded, low-compute path to the same destination, restricting its access to frontier chips or model outputs matters less.
None of this shows up as verified fact anywhere in the source. Liang did not disclose audited revenue. His projection that annualized API income lands somewhere in the low hundreds of millions this year is a founder’s estimate, not a filed number. The compute figures, the timeline, and the talent claims all came from the founder describing his own company to people he is asking for money.
Operators evaluating DeepSeek’s open weights for production use should separate the roadmap from the release schedule. A stated four-stage path to AGI is a pitch, not a delivery date, and the near-term decision that matters is whether the next open weight drop keeps pace with the API DeepSeek runs itself, which Liang says are identical models.
Quotes compiled by Geopolitechs from DeepSeek’s investor call, July 2026.
