Liang Wenfeng Companies: DeepSeek, High-Flyer & Fire-Flyer Explained

Quick answer

Liang Wenfeng's business empire is two companies: DeepSeek, the AI lab he founded in July 2023 (he owns an estimated ~84%), and High-Flyer, the quantitative hedge fund he co-founded in 2015 (estimated ≥76% owned). The relationship is the story: High-Flyer's trading profits funded DeepSeek without a dollar of venture capital, and High-Flyer's GPU clusters became DeepSeek's training backbone. Together they underpin his $36 billion Bloomberg-estimated fortune — see the net worth breakdown for the numbers.

DeepSeek: the AI lab worth ~$50 billion

DeepSeek (Hangzhou DeepSeek Artificial Intelligence) was founded in July 2023 with a mission most startups wouldn't dare print: artificial general intelligence. Liang told Chinese tech publication 36Kr that year that "basic science research has a very low return on investment. But we're doing this because we believe in the mission." The lab's method — open-source releases, engineering efficiency over brute-force scale — produced DeepSeek-V3 (December 2024), the DeepSeek-R1 reasoning model (January 20, 2025) that rattled global markets, and DeepSeek-V4 (April 2026).

The business model is almost anti-business: give the models away, charge little, reinvest everything. It works because the owner doesn't need dividends — he owns ~84%. The June 2026 financing round at roughly $50 billion post-money was the first external validation of that value, and the first outside capital DeepSeek ever took. For how the lab was built, read how he made his money.

Strategically, DeepSeek functions as both a company and a public good. Its open weights are used by researchers and startups worldwide, which builds enormous soft power — and a talent pipeline that flows back to Hangzhou. Critics note the model has a cost: training at reported sub-$6 million budgets is only possible because the GPUs were already paid for, and US export controls mean the next 10,000-chip cluster can't simply be reordered. The company's answer, so far, has been to make each chip count for more — the efficiency research is the business strategy.

High-Flyer: the quant fund that paid for everything

High-Flyer (Hangzhou Huanfang Technology Co., Ltd.) is the older, quieter half of the empire. Liang co-founded it in June 2015 at age 30 to apply machine learning to Chinese equity trading. It worked spectacularly: by 2021, assets exceeded 100 billion yuan (~$14 billion), ranking it among China's top quant funds and making Liang a billionaire on fund economics alone — management and performance fees plus partners' own capital.

Regulatory tightening from 2019–2023 squeezed the industry, and High-Flyer wound down its flagship product in February 2024. Assets today sit around $10 billion. But by then the fund had already served its historic purpose: financing an AI lab that venture capital never touched. Forbes estimates Liang owns at least 76% of it — a second concentrated bet, same as the first.

Fire-Flyer: the supercomputers in between

Neither company fully explains the fortune without Fire-Flyer, the compute infrastructure High-Flyer built and DeepSeek inherited. Fire-Flyer I came online in 2020 with 1,100 Nvidia A100 GPUs; Fire-Flyer II followed in 2021 with roughly 10,000 A100s — assembled before US export controls cut off China's access to cutting-edge chips. A partner later told the Financial Times that Liang's plan for a 10,000-chip cluster sounded delusional at the time: "We thought this was only possible from giants like ByteDance and Alibaba."

That stockpile — reportedly worth hundreds of millions of dollars — is the hidden asset behind DeepSeek's famous efficiency claims. When the company says V3 cost under $6 million to train, the figure excludes the GPUs it already owned. Competitors must buy compute at market prices; DeepSeek trains on sunk cost. Whether Fire-Flyer sits legally inside High-Flyer or straddles both entities is not publicly disclosed — but economically, it belongs to Liang either way.

Ownership structure: how concentrated it is

Liang's empire is remarkable for what it lacks: co-founders with big slices, venture investors, public shareholders. The table below summarizes what is verifiable versus estimated — and the estimates (Forbes) are the best available because Chinese private-company filings are opaque.

Compare the concentration with Dario Amodei's Anthropic, where seven co-founders split ~1.6% each after venture rounds. Liang's structure is closer to a 19th-century industrialist than a 21st-century startup founder — one man, two companies, no dilution. It is the single biggest reason he, not a better-funded rival, became the richest AI founder on Bloomberg's index. Our head-to-head comparison breaks down the math.

R2, V4, and what's next for the portfolio

The company roster is small, but the product pipeline is not. DeepSeek-V4 shipped in April 2026 as a model family — a 1.6-trillion-parameter V4-Pro and a lighter V4-Flash, both with 1-million-token context windows and switchable thinking modes — keeping Liang's open-weights promise even as the models scale. The product everyone is actually waiting for, though, is DeepSeek-R2, the next dedicated reasoning model. As of mid-2026 it has not launched: Reuters reported Liang personally held it back because he wasn't satisfied with its performance.

That delay is revealing about how he runs his companies. A venture-backed CEO facing a $50 billion valuation would feel pressure to ship on schedule; Liang, owning ~84% and answering to no board of outside investors, can afford to wait until the model meets his bar. It is the same quality-over-hype instinct that defined the Fire-Flyer build years. Whether R2 lands in late 2026 or slips further, the decision rights sit with one man — which is precisely why the ownership table above matters more than any product roadmap.

Beyond models, the open question is whether DeepSeek stays a two-company empire or grows a third leg: cloud services, hardware, or an enterprise arm to monetize the open-source base. Nothing has been announced. Given Liang's history — a quant fund, then a compute cluster, then an AI lab, each funding the next — the pattern suggests whatever comes third will be paid for by the first two, not by outsiders.

Hangzhou: the base of operations

Both companies sit in Hangzhou, the Zhejiang-province tech hub also home to Alibaba — and to Zhejiang University, Liang's alma mater. The geography is not accidental: his entire adult life, from enrolling at 17 to building two companies, has played out within the same city. Hangzhou's government has actively courted AI investment, and DeepSeek's presence has turned the city into a magnet for young researchers who once defaulted to Beijing or Shenzhen. He operates from the Hangzhou headquarters, rarely travels for publicity, and gives almost no interviews. For a man whose companies are worth tens of billions, the physical footprint is a few office buildings and some server halls — the lightest empire in the billionaire class. How he draws income from it all is covered in his salary breakdown.

Frequently asked questions