DeepSeek's Liang Wenfeng: China's Only Real AI Gap Is Compute

In a four-hour investor meeting on May 20, DeepSeek founder Liang Wenfeng laid out a strategic theory of AI competition that amounts to one bet: the United States is ahead only because it can buy more chips, and that advantage is narrowing faster than the talent gap is closing. The remarks, first transcribed and analyzed by blogger Fred Gao, have not been confirmed by DeepSeek or by any investor present, and Gao has flagged that he was not sure the original transcript could be freely disseminated. They are still the most concrete public read yet of how one of China's leading AI labs is thinking about the contest.

Liang argues that the only durable US lead in AI is compute, and that talent is "spread around the world by chance," with no real shortage in China. From that single constraint, the rest of the strategy falls out: if the lab cannot match frontier training budgets, it has to win on openness, on cost, and on the stability of the team doing the work.

In Liang's telling, open-source models help enterprises adopt AI faster and undercut the closed-source, high-margin strategy that US labs are running. Pursuing excessive profit on closed models is, in his framing, a strategic weakness: a more restrained rival will simply outcompete the rent-seeker. The implied target is the US frontier-lab business model, in which capital intensity is treated as a moat. Liang's read is that the moat is thinner than the model releases suggest.

Liang told investors the real key to AGI is not hiring the very best individual researchers but keeping the team stable. If that holds, DeepSeek's response to the compute constraint is structural: hold the people, ship the models, accept lower margins. If it does not hold, and DeepSeek's attrition matches US-lab norms, the philosophy collapses into rhetoric. The claim is falsifiable, and on a measurable timeline.

So is the consolidation forecast. Liang predicts that only three or four foundation-model companies will remain globally, in both China and the United States. He also argued that the current leads among OpenAI, Anthropic, and Google are cyclical, not durable, and that Anthropic's early coding-agent edge will fade. Both claims read as a strategic posture from a lab that wants to be one of the survivors and that does not want to be written off as a regional also-ran.

The Daoist framing that Gao's account emphasizes, restraint, fairness, and the metaphor of AGI as a tide no one can own, is texture, not load-bearing structure. It is also a hard sell for an English-language AI reader: US frontier-lab founders tend to talk in capability roadmaps and ARR, not in metaphors about pushback and shared gain. The substantive critique underneath the metaphor is that closed-model profit-maximization is structurally fragile, and that the first frontier lab to break ranks on pricing wins the long game. That claim is sharper than the metaphor.

The material is a blogger's account of an alleged meeting, with a permission flag on dissemination. DeepSeek and Liang have not publicly verified the remarks. The meeting date is May 20, with the year pinned to 2025 from the DeepSeek news cycle. Anything in the talk that is specific, the Anthropic fade claim, the three-to-four-player forecast, the team-stability thesis, should be tested against what DeepSeek actually does over the next year: who stays, who ships, and what the company charges.

The honest test of Liang's theory is not whether it sounds restrained. It is whether a Chinese lab that publishes open weights, holds its team, and prices on cost rather than scarcity can hold a frontier position into 2027. If yes, the closed-model playbook is in trouble. If no, the talk was posture.