--- title: "Liang Wenfeng (梁文锋) and the cheap-AI doctrine that rattled Silicon Valley" date: 2026-09-29 category: Builders site: NeuroAI canonical: https://neuroai.site/a/na-people-liang-wenfeng language: en --- # Liang Wenfeng (梁文锋) and the cheap-AI doctrine that rattled Silicon Valley > Liang Wenfeng (梁文锋), the former quantitative-fund manager who founded DeepSeek in 2023, turned a hedge-fund compute cluster into one of the world's most disruptive AI labs. His public record — sparse interviews, a quant background, and a fixation on cost-efficient architecture — explains more about DeepSeek's rise than the myth-making around it. The man who forced Silicon Valley to rethink its AI budget was, until his late thirties, a hedge-fund quant who rarely spoke in public. Liang Wenfeng (梁文锋) founded DeepSeek in 2023, and within two years his lab had shipped models that matched US frontiers at a fraction of the cost. The story is less a genius myth than a case study in how a different funding base produces a different kind of AI company. ## From Zhanjiang to Zhejiang University Liang was born in **1985 in Zhanjiang, Guangdong**. He studied at **Zhejiang University**, earning a Bachelor of Engineering in Electronic Information Engineering in 2007 and a Master's in Information and Communication Engineering in 2010 — his thesis work touched target-tracking algorithms for low-cost camera systems. As a master's student in 2008 he and classmates began scraping market data and applying machine learning to trading predictions, an early sign of where he was headed. ## The High-Flyer years In **2015–2016** Liang co-founded **High-Flyer (幻方量化)**, a quantitative hedge fund built on mathematical modelling and algorithms. The firm scaled fast: assets under management reached roughly **RMB 10 billion by 2019** and, by Liang's own and reported accounts, exceeded **RMB 100 billion by 2021**. Crucially for what came next, High-Flyer poured trading profits into AI infrastructure. Before US export controls tightened, the fund assembled a cluster of **more than 10,000 Nvidia A100 GPUs** — a private supercomputer that later became DeepSeek's training base. That detail reframes the usual narrative. DeepSeek was not starved of compute; it inherited a quant fund's already-paid-for GPU stockpile and a culture of squeezing maximum value from hardware. ## Founding DeepSeek Liang launched **DeepSeek in May 2023** as an AI research arm of High-Flyer, explicitly framed as long-horizon, curiosity-driven research rather than a conventional product venture. He gave only a handful of interviews — to the tech media 36Kr (specifically its 暗涌 / Anyong brand) in 2023 and 2024 — and the quotes that circulate are consistent: basic research has low ROI, China undervalues top talent, and innovation needs "as little intervention and management as possible." In January 2025 he was invited to a **government advisory session chaired by Premier Li Qiang** alongside experts from other fields — a factual marker of how centrally AI policy now sits in Beijing, not a political statement about the man. ## The technical approach: do more with less DeepSeek's public models show a coherent architectural philosophy: - **Multi-head Latent Attention (MLA)** compresses the key-value cache, slashing the memory and compute needed to serve long contexts. - **DeepSeekMoE** uses fine-grained mixture-of-experts routing so only a small fraction of parameters activates per token. - **Extreme training-cost discipline.** DeepSeek-V2 (May 2024) was reported trained for about **US$6 million**; V3, released months later, for a similar order of magnitude over roughly two months — figures the lab cited to argue it could match far richer labs. The result was a **price war** across Chinese model providers in 2024, as DeepSeek's efficient APIs undercut incumbents. When **DeepSeek-R1** launched in January 2025 as a reasoning model rivaling OpenAI's o1, the shock was global: Nvidia lost about **US$600 billion** in market value in a single day, and DeepSeek's app briefly passed ChatGPT as the top free app on Apple's US store. ## Why the doctrine matters Liang's bet is that architectural ingenuity, not brute hardware spend, is the durable edge — a direct challenge to the assumption that frontier AI requires unlimited GPUs. He has argued that China "cannot remain a follower forever" and must shift from consuming others' innovations to contributing its own. Whether or not one agrees, the stance produced a lab that publishes weights openly and prices access cheaply, widening who can build on frontier-class models. ## What readers can do now - **If you build applications**, benchmark DeepSeek's open-weight models (V3, R1) against closed APIs on your cost-per-token; the efficiency gap is the whole point of Liang's approach. - **If you study AI strategy**, read the 36Kr interviews directly rather than secondhand summaries — the "long-termism" and talent philosophy are the clearest primary source on his thinking. - **If you invest**, separate the man from the model: DeepSeek remains privately funded via High-Flyer, with no IPO signal, so exposure runs through the ecosystem, not a listing. ## Honest limitations Core biographical facts (born 1985 Zhanjiang, Guangdong; Zhejiang University BEng 2007 / MEng 2010; co-founded High-Flyer 2015–2016; AUM ~RMB 10B by 2019 and >RMB 100B by 2021 per reported accounts; >10,000 Nvidia A100 GPUs; founded DeepSeek May 2023; V2 May 2024 ~US$6M training; R1 Jan 2025; Nvidia ~US$600B single-day drop; app topped US App Store; invited to Premier Li Qiang advisory session Jan 2025; TIME 100 in AI 2025) come from CGTN, Channel News Asia, LiveMint, IFANN wiki and Reuters-sourced market reporting. The US$6M training-cost figures are DeepSeek's own cited claims, not independently audited. Liang's quoted philosophy is drawn from 36Kr (Anyong) interviews as reported by CGTN/CNA; no fabricated quotations are used. This article is current to 29 September 2026 and is not investment advice. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-people-liang-wenfeng Free to quote with attribution and a link to the original.