Liang Wenfeng: DeepSeek Founder Who Disrupted the AI Market
The Chinese hedge fund manager who built an AI lab that matched Western models at a fraction of the cost, triggering the largest tech stock drop in history.
👁 112 reads
Liang Wenfeng is not a Silicon Valley computer scientist, but the founder of a Chinese investment fund who built an AI lab on the side. His company, DeepSeek, showed in January 2025 a reasoning model with quality comparable to the best Western ones, trained at a fraction of the costs declared by competitors. The reaction was the largest single-day tech stock drop in history.
How He Got the GPUs
The High-Flyer fund that Liang ran was buying graphics cards for financial models before US export restrictions closed Chinese access to the latest hardware. That stockpile became DeepSeek’s infrastructure. This is a case where export restrictions forced efficiency instead of blocking it: since there was less hardware, training had to be better designed.
What This Meant for the Market
Three things. Prices dropped, because it turned out quality doesn’t require the most expensive hardware in the largest quantities. Open weights, because DeepSeek released the model publicly, making it a reference point for everyone. Narrative shift, because the advantage of Western labs stopped looking permanent. Current model positions, including Chinese ones like Qwen, are shown in my ranking.
What to Remember When Using Chinese Models
A model run locally doesn’t send data anywhere and in that respect is safer than any cloud service. Using Chinese providers’ apps and APIs is a different matter: different data access laws apply there, and responses on politically sensitive topics are sometimes filtered. For working with company data, the proper path is running the model yourself, which I describe in articles about LM Studio and Ollama.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Common Questions
Is DeepSeek free?
Model weights are publicly available, and the app and API have both free and paid plans. Running the model locally is free.
Are Chinese models worse?
In typical text tasks, leading Chinese models are comparable to Western ones. Differences show up in topic filtering and in quality of specific languages. That’s an assessment from use.
More on this from QUANTSLANT
guides·Free 80s Photo Edits on Your Phone: What Actually WorksWe checked which AI tools let you create 80s-style photos on your phone without a subscription, what the real limits are, and how to fix results.