The DeepSeek Thesis

Liang Wenfeng’s pitch is stranger than a normal startup story. DeepSeek’s CEO, who is now richer than Dario Amodei or Sam Altman, is using a hedge fund fortune to back a small research group in Hangzhou that gives its work away for free. The leaked four-hour meeting with investors, which spread in late July, makes one thing clear: Liang is not playing the usual Chinese tech game of chasing whatever sells fastest.

He thinks the business still works anyway. Liang told investors that business customers are DeepSeek’s biggest source of revenue, and that some corporate users stuck around after the R1 wave in 2025. Because DeepSeek has released its models openly under the MIT license since R1, those customers aren’t paying for on-premise deployments. They are paying for API access. DeepSeek’s own February 2025 analysis said R1’s API was bringing in US$562,027 a day at a 545% profit margin. Liang apparently told investors that enterprise revenue this year could reach the hundreds of millions of US dollars.

But money is not the point he keeps returning to. Liang says the “main quest” is AGI, and he treats learning as the central problem, not agents or some other fashionable subfield. World models, in his view, are irrelevant to higher intelligence. He thinks embodied AI will come later, after systems learn to self-iterate. The hard part now is building models that can keep acquiring knowledge, and he is honest that DeepSeek does not yet know exactly how to do that.

That belief shapes everything else. DeepSeek seems to assume future AI will be open, with China acting as a global token factory that drives down the price of intelligence. Liang talks about China’s hardware constraints as a historical problem that will fade, expects Nvidia’s CUDA moat to weaken, and is cautiously optimistic about training on Huawei chips. Huawei has reportedly allocated 16,000 Ascend 950 GPUs to DeepSeek, though Liang said larger internet companies need them more. DeepSeek’s own reason for buying domestic chips, he said, is to help build a domestic hardware ecosystem.

Inside the company, he is selling a very different management style from the usual China internet grind. Liang says mandatory work should take up less than half of an employee’s time, with the rest left for self-directed research if compute is available. He argues that research needs slack, that good ideas come from idle curiosity, and that DeepSeek succeeds by saying no to projects that don’t fit the core mission. In his telling, the company’s edge is focus, not hero worship, and certainly not 996.

The comparison that hangs over all of this is Demis Hassabis. Liang sees mission and commercialization as compatible, while Hassabis ended up unhappy inside Google’s machinery. DeepSeek’s founder is trying to avoid that trap by keeping the company aligned around a single thesis. The leak may have rattled his funding plans, but it also exposed the bet at the center of everything: that the shortest path to a business is still a long march toward machine intelligence.