Liang Wenfeng: The quant behind Nvidia’s $600B shock

On January 27, 2025, the most valuable company on earth lost close to $600 billion. Nvidia had not been hit by a rival chipmaker or a bad quarter. It had been hit by a free app from a Chinese company most people had never heard of. There was no launch campaign, no billion-dollar funding round, and no famous founder. The team had never shipped a consumer product, its leader had given roughly two substantial interviews in his life, and its breakthrough model was built using less-powerful Nvidia chips because U.S. export controls restricted China’s access to advanced processors. Everyone asked how. The better question is why nobody in Silicon Valley had done it first. The answer is the most useful thing any executive can take away from Liang Wenfeng’s story, and it has very little to do with genius. The man who thinks like a trader Liang is usually described as the quiet genius who made AI cheap. He is something more specific: a quantitative trader. That is not a biographical footnote; it is the whole explanation. Quant funds compete on a cruel premise. Everyone sees the same market data at the same moment, so nobody wins by having better information. You win by extracting more signal from identical inputs than the person across from you. Liang spent a decade doing that, co-founding High-Flyer in 2015 and building it into one of China’s largest quant funds. That career instills a single habit: when you cannot get more resources, extract more from the ones you have. How a hedge fund became an AI lab High-Flyer spent RMB 1.2 billion ($178 million) on two AI supercomputing clusters in 2020 and 2021. The second, Fire-Flyer II, comprised roughly 10,000 Nvidia A100 chips used to train AI models. At the time, this looked like an extravagance. Within a year, it looked like foresight, as Washington began restricting sales of those chips to China in October 2022, leaving the mid-sized Hangzhou fund with hardware that the country’s technology giants could no longer legally buy. Then, the ground moved under the day job. Through 2024, Chinese regulators bore down on the quant industry, restricting short selling, suspending accounts, and tightening rules on program trading. Quant investing stopped being the reliable machine it had been. What remained was a stockpile of scarce chips and a team whose defining skill was doing more with less. DeepSeek was founded in 2023 as a separate company, funded by High-Flyer rather than outside investors. Liang’s reason for refusing venture capital was that it comes with exit timetables and pressure to commercialize, which is incompatible with research that may take years. His stated goal was not a product but a hypothesis: that human thought may essentially be a linguistic process, making language models a potential route to understanding the mind. It sounded like an indulgence. It was also a strategy, because it meant nobody could tell him to ship. What scarcity forced DeepSeek to discover The number that detonated in January 2025 was $5.6 million, the reported cost of the final training run behind DeepSeek’s V3 model. The figure needs an honest caveat: it covers one training run, not years of research, salaries, failed experiments, or the chips. What gave the claim weight was DeepSeek’s decision to publish the underlying engineering for anyone to inspect. The reason those techniques existed at all is the heart of the story. DeepSeek trained largely on H800 chips, the deliberately throttled version Nvidia was permitted to sell to China, whose particular weakness is how fast chips can talk to one another. Blocked from better hardware, the DeepSeek team rebuilt the parts of the process that hardware normally brute-forces, cutting the model’s memory demands and waking only the sliver of the network each question actually needs. Better-funded labs had not done this work because they had no reason to. After all, buying more chips was faster than rethinking the method. That is the uncomfortable lesson under the headlines: the restriction was the design brief, and the restricted team found what the unrestricted teams had skipped. Scarcity is an unpleasant gift, but it is a gift. Two theories about where value lives DeepSeek gave the models away, publishing them openly for anyone to download, modify, and run. This is where the story stops being about one founder and becomes a window into two different national instincts. America’s leading labs have moved steadily towards keeping their best models closed. The logic is sound: the model is the product, access is metered, safety is easier to control, and the margin per query belongs to you. Value is captured by owning the best thing. China’s leading labs have moved the other way. DeepSeek and Alibaba’s Qwen publish their weights. That is the logic of a player behind on hardware. If you cannot win on compute, win on adoption. Open models spread your architecture into other people’s products, set the default thousands of developers build on, attract researchers who want their work seen, and let an ecosystem improve your system at no cost to you. It is the difference between selling the best car and making sure the world’s roads are built to your dimensions. Both are viable, which is what most commentary misses. But they collide commercially: open models put a ceiling on the price of closed ones. Once a capable model is free to download, every enterprise negotiating an AI contract has a credible alternative, and that disciplines pricing across the market. America is optimizing to own the best model; China is optimizing to make good models ordinary. Those are different bets about where the money ends up, and the second is quietly reshaping what any company should expect to pay. When the same constraint bit back Having proved Chinese AI could advance without American chips, DeepSeek was encouraged to go further and train its next flagship on domestic Huawei Ascend hardware. It did not work. Through 2025, the effort encountered unstable performance, slow communication between chips, and immature software. The follow-up model slipped by months, and DeepSeek returned to Nvidia hardware for training, using Huawei chips to run finished models. Liang has been blunt that the chip embargo, not money or talent, is his real limitation. So the same scarcity that produced the breakthrough later cost him most of a year, exactly when his lead was most valuable. Constraint is a superb teacher and a terrible landlord. Any company celebrating its own scrappiness should sit with that. He recovered. In April 2026, DeepSeek previewed its V4 family, including a 1.6-trillion-parameter flagship, still under an open license. And in June, the founder who rejected venture capital on principle raised roughly $7.4 billion at a valuation above $50 billion, making DeepSeek China’s most valuable AI startup. The move is less contradictory than it looks. Liang refused money that came with control over his timetable, then took it when compute became the binding constraint and he could set the terms. Lessons for executives Treat your constraint as the brief, not the excuse. DeepSeek’s efficiency work exists because its engineers could not buy their way past a bottleneck. In most companies, the best engineering sits exactly where somebody was forced to think instead of spend. Own the scarce input before anyone calls it scarce. The A100 stockpile was acquired before Washington restricted exports to China, turning an earlier investment in AI infrastructure into a strategic advantage. Optionality is cheap in advance and unpurchasable afterwards. Let the funding structure match the time horizon. Liang refused outside capital when the work was slow and unglamorous, then accepted it when the constraint became the cost of buying compute at scale. Who funds you determines what you are allowed to attempt. Give away what can be copied; keep what cannot. Weights can be downloaded in an afternoon. The team, the research culture, and the habit of questioning expensive assumptions cannot. The bottom line Return to the question at the start. Silicon Valley did not miss DeepSeek’s efficiency breakthroughs because its engineers were worse. It missed them because it could afford to. Abundance removes the reason to question your most expensive assumption, and a competitor who cannot afford that assumption will eventually test it for you, publish the answer, and hand it to your customers for free. Liang never set out to make AI cheap. He spent a decade learning to squeeze more out of a limited position, first in markets and then in machines. The chips he bought for a trading desk ended up resetting the price of intelligence. The question worth carrying out of his story is not, “How did he do it?” It’s, “Which of your expensive assumptions has never been tested?” Because somewhere, a smaller competitor is being forced to test it right now.

Liang Wenfeng: The quant behind Nvidia’s $600B shock

How 2026 welcomed ‘Too Cool’ back

Upgrade to Pro

Sectors

Insights

Company

Liang Wenfeng: The quant behind Nvidia’s $600B shock

How 2026 welcomed ‘Too Cool’ back