Liang Wenfeng, the DeepSeek founder who rattled Wall Street and then went silent

Liang Wenfeng

Born	January 1, 1985 Zhanjiang
Occupation	Entrepreneur
Awards	Time 100 u00b7 Nature's 10

The question that Silicon Valley never quite answered, in the months after DeepSeek arrived, was not technical. It was philosophical. If a team of fewer than two hundred engineers, working in Hangzhou on hardware that America’s export controls were supposed to limit, could produce a model competitive with the best in the world at a fraction of the cost — what exactly was all the money for? Liang Wenfeng was the person who posed that question, but he did not stick around to hear the debate. He had already gone quiet.

Wuchuan, the coastal city in Guangdong Province where he grew up, is not a place that appears in the standard AI origin stories. His parents were primary school teachers. He taught himself advanced calculus in junior high school and finished at the top of the Zhanjiang region in the college entrance examination, earning him a place at Zhejiang University at seventeen. He studied electronic information engineering, then information and communication engineering at postgraduate level, with a thesis on object tracking algorithms for cameras. The research subjects were modest, but the sensibility — applied mathematics used to solve constrained, real-world problems — would define everything that followed.

When the 2008 financial crisis sent markets into freefall, he began experimenting with machine learning applied to trading. He moved to Chengdu, worked through a series of early ventures, and in 2016 co-founded High-Flyer Capital Management (幻方量化) in Hangzhou with two Zhejiang University classmates. The firm would become one of China’s most successful quantitative hedge funds, managing more than seventy billion renminbi and posting average returns of over fifty percent in 2025. The edge was algorithmic: deep learning models running on GPU infrastructure that Liang began accumulating at scale. By 2021, he was purchasing Nvidia A100 chips in quantities that suggested plans considerably larger than a trading operation.

That infrastructure became the foundation for DeepSeek, which Liang established as a spinoff in July 2023. The company’s stated goal was not to build a chatbot or a commercial product, but to pursue what he described as foundational AI research — the kind of work that does not have a product roadmap or a quarterly earnings target. DeepSeek employed roughly one hundred sixty people, recruited from a deliberately wide range of disciplines, and operated with a budget that would barely register in the accounts of OpenAI or Google DeepMind. When DeepSeek-V3 was released at the end of 2024, the training cost was estimated at approximately six million dollars. When DeepSeek-R1 followed in January 2025 and topped the iOS App Store in the United States, displacing ChatGPT, the figure that circulated was five point six million dollars. Models that the largest technology companies in the world had spent hundreds of millions to develop had apparently found a formidable, cheaper equal.

The disruption narrative that attached itself to DeepSeek almost immediately was not entirely wrong, but it was incomplete in ways that mattered. The story told in financial markets — that DeepSeek had proved American AI spending was wasteful, that export control strategy had failed — assumed that what Liang’s team had done was reproducible anywhere, by anyone, on a tight budget. That assumption deserves scrutiny. High-Flyer’s GPU infrastructure, built before the tightest export restrictions came into force, was not a modest resource. The engineering choices that produced efficient models at low training cost required expertise of unusual depth. What DeepSeek demonstrated was not that large AI models are cheap to build; it was that architectural innovation can compress the marginal cost of training a given capability level. The distance between those two claims is considerable, and much of the public commentary ran them together.

In February 2025, Liang attended a symposium with President Xi Jinping, alongside some of China’s most prominent technology executives. It was his most visible public moment. He has not appeared publicly since. DeepSeek released a V4 preview in April 2026 and is reported to be seeking three hundred million dollars in new investment at a ten billion dollar valuation. His ownership stake — approximately eighty-four percent through direct and indirect holdings — gives him both the control and, evidently, the latitude to be anywhere but in front of a microphone.

He keeps almost nothing on the public record about his private life. There is no verified information about a spouse or children. He maintained no public social media presence before DeepSeek became global news, and acquired none after.

DeepSeek will release more models. The trajectory in the months since the R1 moment shows no sign of slowing. What remains genuinely unclear is whether Liang Wenfeng will emerge to say anything about any of it, or whether the work will continue to speak where he has chosen not to.

Tags: artificial intelligence, business, ai, ceos, DeepSeek, technology entrepreneurs