Liang Wenfeng

"We're done following," Liang has said. "It's time to lead."¹ He had prepared for the sentence quietly: close to 10,000 Nvidia A100 chips accumulated inside a quantitative hedge fund with no stated artificial-intelligence product, in place before United States export controls took effect in September 2022.²

Liang Wenfeng is the founder and chief executive of DeepSeek, a Chinese artificial-intelligence research lab that builds open-source large language models, including the R1 reasoning model released in January 2025. DeepSeek's models are notable for matching or approaching the performance of Western frontier systems at a fraction of the reported training and inference cost. Liang is also the founder of High-Flyer Quantitative Investment Management, a Chinese hedge fund that funds DeepSeek entirely and holds about 99 percent of it; Liang personally holds roughly 1 percent. He is an unusually private figure, with very few interviews and no Western-style social-media presence.

Background

Liang was born around 1985 in Mililing, a village of roughly 700 people in Wuchuan, Zhanjiang, in Guangdong Province, an area often described as a fifth-tier part of China. His father was a primary-school teacher; some accounts describe both parents as teachers. A strong mathematics student, he was the top scorer in the Zhanjiang region's college entrance examination, the gaokao. He was admitted to Zhejiang University in 2002, where he studied electronic information and electrical engineering with a later focus on artificial intelligence, and completed a master's degree. His master's thesis addressed object tracking using low-cost pan-tilt-zoom cameras, compensating for cheap hardware through software and algorithmic means. Liang and later profilers have repeatedly cited that thesis as the conceptual ancestor of DeepSeek's core approach: when resources cannot be matched, out-think competitors algorithmically.³ He is distinctive among prominent Chinese technology founders for having no overseas education or work experience, and his teams have been staffed domestically.

Getting started

Around 2008, during the global financial crisis, Liang began applying machine learning to quantitative trading, treating the market's disorder as a training dataset. He started with roughly 80,000 yuan of capital. He has said he was once personally invited by Wang Tao, the founder of the drone company DJI, to help co-found DJI, and that he declined in order to pursue artificial intelligence applied to finance.³ He built High-Flyer gradually over roughly a decade, and by about 2015 the fund had achieved recognized success. Multiple profiles report that it delivered returns above 100 percent for seven consecutive years in its early period, a figure that is widely repeated but not independently audited. Alongside the trading infrastructure, Liang steadily accumulated Nvidia GPUs over roughly a decade, described as moving from a single unit in the early days to about 100 in 2015, roughly 1,000 in 2019, and close to 10,000 A100 chips before United States export controls took effect in September 2022. Profiles frame this as patient, long-horizon infrastructure building rather than a reactive response to the AI boom.² He founded DeepSeek in May 2023, after ChatGPT's late-2022 debut made the AI race visible, carving it out of High-Flyer with no external venture funding.

What he built

High-Flyer grew into one of China's top quantitative hedge funds, with assets under management reported at various points around 100 billion yuan, or about 14 billion dollars, and elsewhere described as roughly 8 billion dollars; the figures vary by year and source. DeepSeek released a rapid sequence of open-source models, including DeepSeek-V2, V3, and R1. DeepSeek-R1, released in January 2025, matched or approached the reasoning performance of OpenAI's o1-class systems at a dramatically lower reported training cost. Its release was followed by a large sell-off in United States AI and semiconductor stocks, an event widely reported as a reassessment of the assumption that AI progress requires the largest compute budgets. Technical innovations attributed to the team include Multi-Head Latent Attention and the DeepSeekMoE architecture, both of which sharply reduced inference costs; the latent-attention work is reported to have originated from a young researcher's personal interest rather than a top-down directive. Liang built the team largely from fresh domestic graduates, PhD candidates, and early-career researchers rather than hiring people away from OpenAI, Google, or DeepMind, which he has described as a deliberate choice. He has kept DeepSeek's models open-source and priced near cost, saying the aim was neither to sell at a loss nor to seek excessive profit.¹

How he operates

Liang runs DeepSeek with a flat, bottom-up research culture in which, by his account, there are no predefined roles, researchers have largely unrestricted access to GPUs, and collaboration is driven by shared interest rather than assignment. He has said DeepSeek carries no commercial key performance indicators and is funded off High-Flyer's balance sheet specifically to remove short-term commercial pressure from research oriented toward artificial general intelligence. A decade of machine-learning work and a GPU fleet accumulated inside a hedge fund with no stated AI product turned out to be exactly the equipment an AGI lab needed, and he committed High-Flyer's balance sheet to it without hedging: the Wave Recognizer pattern of a long accumulation in one domain meeting a wave shaped like it. He describes DeepSeek's mission as "unraveling the mystery of AGI with curiosity."³ Liang believes that the real gap is between originality and imitation, and that more investment does not equal more innovation. He has argued that Chinese firms have long focused on monetizing applications while others drove innovation, and that China should become a contributor rather than a free-rider.¹ He cites the mathematician and investor Jim Simons as a model, and the night before R1's release posted a preface he had written for a Simons biography.

Where things stand

As of these sources in the mid-2020s, Liang remains a low-profile figure whose direct quotes trace back to a small number of Chinese-language interviews. DeepSeek is widely credited with demonstrating that architectural and algorithmic efficiency can substitute for raw compute scale, a claim that reshaped discussion of United States and China AI competition through 2025. High-Flyer continues as DeepSeek's sole backer, with no indication in these sources of outside investment.

Key facts

• Founded DeepSeek in May 2023, carving it out of his hedge fund High-Flyer with no external venture funding.
• Personally holds roughly 1 percent of DeepSeek, with about 99 percent held by the High-Flyer partnership.
• Reportedly declined an invitation from DJI founder Wang Tao to help co-found the drone company, choosing AI in finance instead.
• Accumulated GPUs over roughly a decade, reaching close to 10,000 A100 chips before United States export controls took effect in September 2022.
• DeepSeek-R1's January 2025 release was followed by a sharp sell-off in United States AI and semiconductor stocks.
• Built an entirely domestically educated and recruited research team by deliberate choice.