A Four-Hour Meeting, Leaked and Dissected
In late July, minutes from a four-hour meeting between DeepSeek CEO Liang Wenfeng and a room full of investors began circulating quietly. By the time ChinaTalk's Irene Zhang poured over them, the document had become a rare window into the mind of one of the most consequential figures in the US-China AI race. Liang, a finance veteran who founded a hedge fund before building the Hangzhou-based lab, is now richer than Dario Amodei or Sam Altman. His wealth, however, does not come from DeepSeek. It comes from his hedge fund. The meeting, which ran for four hours, covered everything from business strategy to management philosophy to the nature of intelligence itself.
The leak did not sit well with Liang. He was reportedly furious. He reportedly paused a new funding round and pushed back IPO plans after the minutes got out. The episode raised questions about who among the audience wanted to poke holes in Liang's vision. For China+AI watchers, who have long been puzzled by Liang's motivations, the document offered a rare chance to see how he thinks about the world.
The Business of Open Source
DeepSeek has released all models since R1 under the MIT license. That open-source approach is unusual, and it has puzzled observers. But the leaked minutes clarify the business logic. DeepSeek's biggest revenue source is business customers paying for API tokens. In February 2025, the company published an analysis showing R1's API earned US$562,027 on an average day, at a 545% profit margin. That figure, which appeared in DeepSeek's own analysis, demonstrated that open weights do not have to mean open wallets.
Liang gave a hypothetical figure for DeepSeek's enterprise-end revenue this year in the hundreds of millions of US dollars. He did not commit to a precise number, but the scale suggests the API business is real. In 2025, a wave of Chinese businesses and government entities connected DeepSeek to internal systems. That integration has turned the lab into a quiet utility for the domestic economy.
Last year, Liang considered sunsetting consumer products. He kept them, he said, due to loyal users. That decision reflects a broader philosophy: missions do not cost good companies revenue. Liang argued that point directly in the meeting. He sees no contradiction between building AGI and running a profitable business. He regards harnessing commercial incentives for pure AGI development as possible.
He also revealed something personal about that choice. At one point, he said, "Don't know what the point of users is… I'll just feed them for now." That offhand remark captures his attitude toward consumer products. He does not see them as the mission. He sees them as a way to keep the lights on while he pursues something bigger.
Learning as the Path to AGI
Liang's central thesis is a confidently singular one. He believes learning is the most important problem for AGI, not world models or embodied AI. His theory posits an inevitable causal relationship between automated learning and generalized intelligence. He sees AGI as the only problem worth solving right now.
He is careful, though, not to commit to specific subfields or theories. He is honest that he and his team don't know the path to AGI yet. He expressed cautious optimism about training on Huawei chips, but he did not present a detailed technical roadmap. Instead, he emphasized the importance of self-iteration.
Liang thinks embodied AI is inevitable but will be solved after self-iteration. He expects Nvidia's CUDA moat to erode. That expectation is central to his strategic thinking. If the software ecosystem around Nvidia weakens, then the hardware constraints imposed by US export controls become less binding over time.
The topic of model training on domestic GPUs collected the most questions during the Q&A. Investors wanted details. Liang gave them a picture of a lab working within limits, but not paralyzed by them. Huawei allocated 16,000 Ascend 950 GPUs to DeepSeek. That is fewer than the chips sold to bigger internet companies like ByteDance, Alibaba, and Tencent. Liang said DeepSeek can acquire noncompliant chips, so it doesn't need Huawei as much as bigger companies. The main rationale for buying Huawei chips, he said, is to support the domestic hardware ecosystem.
Liang sees China as playing the role of token factory at global scale. He uses phrases like "historic mission" regarding the hardware chokehold. He sees Chinese domestic integration as inevitable. He aligns with Beijing on tech sovereignty. The Communist Party has coercive potential, and Liang is careful to shape the domestic narrative around DeepSeek, avoiding heroism. He takes pains to present the company as part of a larger national effort.
The dangers of potential AGI never come up in the meeting. That omission is striking. Liang is less worried about AGI dangers than many Western leaders. He sees the technology as a problem to be solved, not a threat to be managed. That confidence may be naive, or it may be a deliberate choice to focus on what he can control.
Management Without Heroics
DeepSeek is a small but well-funded lab. Its team is described as little-known researchers. Liang emphasized that DeepSeek is made up of regular people and teamwork is key. He takes pride in the company he built, but he does not sound like he is racing against time.
DeepSeek researchers rarely work overtime. Mandatory tasks take up no more than half of employees' time; the rest is self-directed. Liang believes a person can concentrate only six to eight hours a day. That belief shapes the lab's culture. He argues a relaxed environment is necessary for research.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
This stands in sharp contrast to the AI gold rush in Silicon Valley, which is unleashing workaholic founders. It also contrasts with 996, the common work culture in China's internet industry. Liang has built something different. He focuses on saying no to projects not related to the main mission. That discipline is a bet on focus.
Liang is inspired by Bell Labs but doesn't see DeepSeek as an exact replica. He draws on the model of a research institution that produced world-changing inventions, but he adapts it to his own context. The result is a lab that moves deliberately, not frantically.
Liang's conviction is hard to doubt given his investment. He has put his own money and reputation behind this approach. Whether DeepSeek can achieve AGI might not be up to them. Hardware limits, political pressures, and the sheer difficulty of the problem all stand in the way. But Liang's theory is a confidently singular thesis, and he is willing to bet on it.
The leaked minutes also reveal how Liang sees China's political economy from Hangzhou's vantage point. He is careful not to antagonize the government. He said the government won't give them a cent should the company fail. That comment, ChinaTalk notes, shouldn't be taken literally. It reflects a desire to present DeepSeek as self-reliant, not dependent on state support.
Liang positions himself as focused on realizing his vision for technology. He does not sound like a political actor. But his choices have political consequences. By buying Huawei chips and supporting domestic integration, he aligns with Beijing's push for tech sovereignty. The hardware chokehold imposed by US export controls has forced Chinese labs to innovate within limits. Liang sees that as an opportunity, not just a constraint.
The Hassabis Comparison
The ChinaTalk article draws a direct comparison between Liang and Demis Hassabis, the former CEO of Google DeepMind. Hassabis resigned from Google on August 5th. He will work on AI-assisted drug discovery at Isomorphic Labs and research societal impacts of AGI. His departure came after years of frustration.
Hassabis was disappointed in how DeepMind became entwined in Google's corporate labyrinth. He complained of becoming just another employee from Mountain View's perspective. Sebastian Mallaby's biography of Hassabis, "The Infinity Machine," documents this arc. DeepMind became a Google subsidiary, and Hassabis resented business development work. Larry Page, the former Google CEO, presented Hassabis with a choice about DeepMind's future. That choice led to the integration Hassabis later regretted.
Hassabis remains convinced of keeping powerful technology in the hands of those who understand it most. That conviction drove him to leave Google and start a new chapter at Isomorphic Labs. Liang, by contrast, has never had to answer to a corporate parent. He runs DeepSeek as an extension of his own vision.
The comparison is instructive. Both men believe AGI is the defining challenge of our time. Both have built labs that prioritize research over commercial pressure. But their circumstances differ. Hassabis fought against corporate bureaucracy. Liang has built his own structure from scratch.
Liang's firm symbolizes the US-China AI race. He is a finance veteran who turned to AI, and his hedge fund provides immense wealth. That wealth gives him independence. He does not need to please venture capitalists or public markets. He can pursue AGI on his own terms.
The leak raises questions about who wanted to poke holes in Liang's vision. The meeting was with investors, and the minutes were detailed enough to be useful. Someone in that room decided to share them. That decision may have been an act of sabotage, or it may have been an attempt to shape the narrative around DeepSeek. Either way, the leak has given the world a rare look at how Liang thinks.
Liang does not sound like he is racing against time. He speaks in long horizons. He believes learning is the most important problem for AGI, and he is willing to wait for the results. His team is made up of regular people, he says, and teamwork is key. That humility is part of his management style. He avoids heroism, both in his own presentation and in the company's culture.
The meeting lasted four hours. In that time, Liang covered a remarkable amount of ground. He talked about revenue, about chips, about management, about AGI. He gave a hypothetical figure for enterprise-end revenue this year in the hundreds of millions of US dollars. He discussed the 16,000 Huawei Ascend 950 GPUs allocated to DeepSeek. He explained why he expects Nvidia's CUDA moat to erode.
The leaked minutes do not resolve the puzzle of Liang's motivations. They deepen it. He is a finance veteran who became an AI idealist. He is a billionaire who talks about regular people and teamwork. He is a Chinese CEO who aligns with Beijing on tech sovereignty while insisting the government won't give him a cent.
ChinaTalk's article suggests Liang's approach is a bet on focus and discipline. That seems right. He has said no to projects that do not serve the main mission. He has kept mandatory tasks to no more than half of employees' time. He has built a culture that values concentration over hustle.
Whether DeepSeek can achieve AGI might not be up to them. The hardware constraints are real. The political pressures are real. The technical challenges are immense. But Liang has a plan, and he is sticking to it. He believes learning is the path to generalized intelligence. He believes the CUDA moat will erode. He believes China will play the role of token factory at global scale.