On January 27, 2025, Nvidia lost $593 billion in market value in a single
day — the largest one-day loss for any company in Wall Street history.
The trigger was not a rival chip. It was a free chatbot app from a
startup few had heard of, founded by a man who had given two interviews
in two years. His name is Liang Wenfeng. The story starts not in a
garage, but on a Chinese trading floor.
The Quant Who Read Jim Simons Before the World Did
Liang was born in 1985 in a village in Wuchuan, Guangdong, the son of
two primary school teachers. He entered Zhejiang University at 17,
scoring 806 on the gaokao — enough for Tsinghua, which he passed up.
During the 2008 crisis he and classmates began collecting market data
and testing machine-learning trading. In 2016 he co-founded High-Flyer;
by 2019 it managed over 10 billion yuan and ran its own AI lab. He
wrote the preface to the Chinese edition of Jim Simons' biography,
drawn to the line "There must be a way to model prices."
High-Flyer peaked above $12 billion in assets. Then 2021 arrived — a
public apology to investors, assets down more than a third. While the
fund apologized, Liang was buying something his partners could not
justify: thousands of Nvidia GPUs.
10,000 Chips and a Piano
One card, then 100 in 2015, 1,000 in 2019, then 10,000. "Many people
assume there's some hidden commercial logic," Liang told 36Kr. "But
actually, it was mainly driven by curiosity."
He compared the cluster to buying a piano: first you can afford it,
second, there are people eager to play it. Those 10,000 A100s were
bought before US export controls closed the door. In May 2023,
High-Flyer spun off DeepSeek to pursue AGI. Venture capital saw no exit
and passed, so the fund paid for everything — no commercial pressure,
no KPIs.
Efficiency as a Weapon
DeepSeek-V2 (2024) cut KV cache by 93.3 percent and its pricing forced
Alibaba and Baidu to slash rates by more than 95 percent. V3 (December
2024) trained on 2,048 H800 GPUs at a claimed $5.6 million — the
paper's fine print notes the figure covers only the final training run.
R1 (January 2025) matched OpenAI o1 on reasoning benchmarks: 671
billion parameters, 37 billion activated, weights open.
The team numbered under 140 people, mostly fresh graduates. "There are
no wizards," Liang said. V2's key cost innovation came from one young
researcher's personal interest, in a culture where anyone can grab GPUs
without approval.
What He Actually Believes
"OpenAI is not a god." "All playbooks are products of the previous era."
"Innovation is, first of all, a matter of belief."
The behavior matches the words: DeepSeek releases weights under MIT
license, and Liang sees no benefit in closing them. On AGI timing he
answers: two, five, or ten years — "in any case, it will happen in our
lifetimes."
By July 2026, Bloomberg put his net worth at $36 billion — the
wealthiest founder of any AI model company. The man who avoids press
sits on the TIME100 AI list and shared a table with Xi Jinping in
February 2025.
Liang ran the same race with a fraction of the compute and none of the
marketing, and bet on the one resource Western labs priced out:
patience.
The piano he bought in 2021 turned out to have more players than anyone
expected.
Top comments (0)