A leaked transcript from DeepSeek founder Liang Wenfeng’s investor meeting is circulating on the Chinese internet, offering the most detailed public look yet at how one of China’s leading AI labs thinks about the path to AGI and where agents fit into it. Momentum Works published selected Q&A from the four-hour session on July 24.
The meeting reportedly took place during DeepSeek’s first external fundraise, a ¥50 billion ($7 billion) round that closed in May 2026 with full subscription. Liang personally contributed 40% of the capital, roughly ¥20 billion ($3 billion), according to Momentum Works.
Agents Are What Investors See, Not What Researchers Chase
The most striking passage for the agent ecosystem comes early. When asked about the timeline for continual learning breakthroughs, Liang drew a sharp line between investor perception and internal research priorities: “For investors, what you see most of is agents; but for those of us doing the research, what we see more of is learning, and how to solve it.”
Liang argued that today’s agents are limited precisely because they cannot learn continually. Solve that problem, he said, and general intelligence follows. Without it, building general intelligence by hand is “tiring and bitter, data-intensive, labour-intensive, poor value for effort.”
The implication for agent builders: DeepSeek views current agent architectures as constrained by static model capabilities. The company’s roadmap prioritizes making models that improve themselves over making agents that orchestrate better.
No Scaling Ceiling, No Critical Point
Liang told investors he sees no ceiling to language model scaling at either DeepSeek’s current capability level or that of US frontier labs. Asked whether AGI would arrive through a sudden jump, he rejected the idea of a critical point but described the trajectory as “non-linear,” because AI can accelerate AI research. “You can use AI to speed up your own research, so further down the line it may become non-linear.”
DeepSeek is already operationalizing this. Liang described an internal framework where each new model’s first purpose is to raise DeepSeek’s own R&D efficiency and help build the next version faster. “The first goal of the models we build is not that everyone finds them useful, but that we find them useful,” he said.
China’s Model Market Will Consolidate to Three or Four
Liang was blunt about China’s AI competitive landscape. The US has roughly three foundation model companies. China has “far too many,” and the redundancy is wasteful. He predicted convergence to three or four serious competitors, driven by margin compression: “There won’t be windfall profits. Those who control costs well earn a bit more, those who don’t earn a bit less.”
On AI talent, Liang dismissed the shortage as “just a phase,” comparing it to early internet engineering. “There has never been a long-term shortage of any one type of person,” he told investors.
Embodiment as the Endpoint
Looking past software agents, Liang identified embodied AI as DeepSeek’s likely endpoint. “An ordinary person doesn’t need a computer; in their daily life, what they need is embodied intelligence to meet real, physical needs,” he said. The ambition: once general intelligence works, apply it to iterate on physical robots autonomously.
The Capital Signal
The ¥50 billion round represents one of the largest AI fundraises globally, comparable to OpenAI’s $40 billion round that closed in March 2026. Liang’s personal 40% stake signals an unusual level of founder conviction, tying roughly $3 billion of his own capital to the company’s AGI thesis.
The transcript’s publication comes as Chinese AI labs are matching Western frontier labs in both model capability and capital intensity, while taking a fundamentally different approach to commercialization. Liang expects low margins, open-source distribution, and consolidation. For agent platform operators, the question is whether DeepSeek’s continual learning breakthrough, if it arrives, would reshape what agents can do from the model layer up.